Paper deep dive
Accurate in space, unreliable in time: how LLMs represent national cultural change
Yalda Daryani, Miranda Bogen, Madeleine I. G. Daepp
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 95%
Last extracted: 9/3/2026, 3:30:27 AM
Summary
This study evaluates the temporal cultural awareness of four state-of-the-art Large Language Models (LLMs) by comparing their representations of national cultural values against over two decades of World Values Survey (WVS) data for 40 countries. Using the Inglehart-Welzel cultural map, the authors assess 'snapshot alignment' (accuracy at a specific time) and 'change alignment' (ability to capture cultural trajectories, including direction, rate, and reversals). The findings indicate that while LLMs generally achieve high snapshot accuracy, they exhibit 'temporal flattening,' lagging several years behind current cultural positions, capturing only partial magnitude of change, and failing to reproduce trajectory reversals. This suggests that current evaluations based on static snapshots provide an incomplete picture of LLMs' cultural awareness, with implications for model governance and representational harm.
Entities (12)
Relation Signals (8)
Qwen3.6-Plus → evaluatedin → Accurate in space, unreliable in time: how LLMs represent national cultural change
confidence 98% · We test four models—Claude Opus 4.8, GPT-5.5, Gemini 3.6 Flash, and Qwen3.6-Plus...
Claude Opus 4.8 → evaluatedin → Accurate in space, unreliable in time: how LLMs represent national cultural change
confidence 98% · We test four models—Claude Opus 4.8, GPT-5.5, Gemini 3.6 Flash, and Qwen3.6-Plus...
GPT-5.5 → evaluatedin → Accurate in space, unreliable in time: how LLMs represent national cultural change
confidence 98% · We test four models—Claude Opus 4.8, GPT-5.5, Gemini 3.6 Flash, and Qwen3.6-Plus...
Gemini 3.6 Flash → evaluatedin → Accurate in space, unreliable in time: how LLMs represent national cultural change
confidence 98% · We test four models—Claude Opus 4.8, GPT-5.5, Gemini 3.6 Flash, and Qwen3.6-Plus...
World Values Survey → usedby → Yalda Daryani
confidence 95% · We examine this missing dimension of cultural awareness using more than two decades of the World Values Survey data.
LLMs → exhibits → temporal flattening
confidence 93% · Our findings show that while models generally place countries close to their most recent surveyed positions, these representations tend to lag several years behind that position... These findings point to temporal flattening
LLMs → haslowaccuracyfor → change alignment
confidence 92% · They also capture only part of the magnitude of the observed change, introduce movement where little occurred, and rarely reproduce reversals in countries' trajectories.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Assessments of cultural alignment have become an important part of the development and improvement of large language models (LLMs). However, the majority of the evaluations treat culture as a single snapshot, investigating only whether a model represents a society accurately at the current time. Research in cultural psychology shows that cultural values change at different rates and directions over time. Therefore, a "culturally aware" model should capture not only where a culture is today but also how it has changed over time. We examine this missing dimension of cultural awareness using more than two decades of the World Values Survey data. We compare the cultural trajectories of 40 countries with the trajectories produced by four state-of-the-art (SOTA) LLMs on the Inglehart-Welzel cultural map. Our findings show that while models generally place countries close to their most recent surveyed positions, these representations tend to lag several years behind that position. They also capture only part of the magnitude of the observed change, introduce movement where little occurred, and rarely reproduce reversals in countries' trajectories. These findings point to temporal flattening and suggest that snapshot accuracy can give an incomplete picture of cultural awareness in LLMs and have implications for model evaluation, representational harms, and the governance of culturally aware AI systems.
Tags
Links
- Source: https://arxiv.org/abs/2609.01902v1
- Canonical: https://arxiv.org/abs/2609.01902v1
Trouble viewing inline? Open PDF directly →
Full Text
111,803 characters extracted from source content.
Expand or collapse full text
Accurate in space, unreliable in time: how LLMs represent national cultural change Yalda Daryani 1 , Miranda Bogen 2 , and Madeleine I. G. Daepp 2 1 Department of Psychology, University of Southern California 2 Center for Democracy and Technology Correspondence: daryani@usc.edu Abstract Assessments of cultural alignment have become an important part of the development and improvement of large language models (LLMs). However, the majority of the evaluations treat culture as a single snapshot, investigating only whether a model represents a society accurately at the current time. Research in cultural psychology shows that cultural values change at different rates and directions over time. Therefore, a “culturally aware” model should capture not only where a culture is today but also how it has changed over time. We examine this missing dimension of cultural awareness using more than two decades of the World Values Survey data. We compare the cultural trajectories of 40 countries with the trajectories produced by four state-of-the- art (SOTA) LLMs on the Inglehart–Welzel cultural map. Our findings show that while models generally place countries close to their most recent surveyed positions, these representations tend to lag several years behind that position. They also capture only part of the magnitude of the observed change, introduce movement where little occurred, and rarely reproduce reversals in countries’ trajectories. These findings point to temporal flattening and suggest that snapshot accuracy can give an incomplete picture of cultural awareness in LLMs and have implications for model evaluation, representational harms, and the governance of culturally aware AI systems. Keywords: large language models, cultural alignment, cultural change, World Values Survey Introduction Half a century of social and cultural psychology re- search has established a foundational premise that so- cieties do not sit still (Jackson & Medvedev, 2024; Kurdi et al., 2025). Modernization has broadly pulled many societies from traditional toward secular-rational values and from survival toward self-expression values (Inglehart & Baker, 2000; Inglehart & Welzel, 2005; Welzel, 2013). These broad trends unfold differently across societies, however. Individualism, for instance, has risen across nearly every society studied, yet at sharply different speeds (Santos et al., 2017). Move- ment across values can also reverse. Several Soviet successor states shifted toward survival and traditional values after the economic collapse of the early 1990s (Inglehart & Baker, 2000), and, in a number of ad- vanced democracies, rising self-expression values pro- voked a traditionalist counter-movement among co- horts left behind by that shift (Norris & Inglehart, 2019). Because societies have moved at such differ- ent rates and in such different directions, the surveyed record of the past four decades shows values diverging globally rather than converging on a common destina- tion (Jackson & Medvedev, 2024). Therefore, a society is better described by a trajectory—a direction, a rate, and an occasional reversal—than by any single snap- shot in measured values. Cultural values are now studied not only in popu- lations but also in the LLMs that people across the world are increasingly using for both personal and pro- fessional tasks. In order to represent cultures accu- rately, LLMs must take into account the fact that cul- tures change. A given model may correctly represent the values of a society at a given time, but these val- ues will not necessarily remain the same in the future. To determine whether an earlier representation of a given society is still relevant today and how this soci- ety might evolve in the future, it is necessary to have information on how the society has evolved in the past. Thus, in addition to the question of whether LLMs know enough about the cultures of all the different societies, there is also the question of whether these models represent cultures as in motion over time. The extent to which LLMs accurately represent cultures is an important question both for everyday people and for science. LLMs are used daily by hundreds of mil- lions of people worldwide (Microsoft AI Economy Insti- 1 arXiv:2609.01902v1 [cs.CY] 1 Sep 2026 tute, 2026), mediating how those people retrieve and interpret information (Zhu et al., 2025) and potentially reshaping language and even cultural norms (Daryani et al., 2026). In social scientific research, moreover, LLMs have been advanced as low-cost “silicon sam- ples” that can stand in for human populations (Argyle et al., 2023). Whose values a model reflects, then, is not only a scientific question about the validity of us- ing LLMs to represent the human population but also a consequential governance question about representa- tion, cultural erasure, and hegemony (de Roock, 2024; Mollema, 2025; Shelby et al., 2023). Early studies found strong and consistent evidence that model outputs frequently resemble the opinions of the United States and of English-speaking and Protestant-European populations (Durmus et al., 2024; Santurkar et al., 2023; Tao et al., 2024). This bias is ob- served across different benchmarks and measures. On standard psychological measures, model responses fall outside the range of global data (Pataranutaporn et al., 2025) and track most closely with Western, Educated, Industrialized, Rich, and Democratic (WEIRD) popu- lations (Atari et al., 2023; Cao et al., 2023). Changing the language in which LLMs are prompted does not consistently reduce this cultural skew (Bulté & Rigouts Terryn, 2026; Mora et al., 2025), though more recent studies show that naming a country explicitly can shift model responses toward the values of that society, with cultural alignment being stronger for countries whose languages are better represented in a model’s training data (AlKhamissi et al., 2024; Tao et al., 2024). Although there are accumulating studies examining how models represent cultures at a fixed point in time, we know of no study that asks whether models capture how a society has changed. Cultural alignment audits compare a model against a single recent snapshot, a design that treats a nation’s culture as a fixed point rather than a trajectory (Masoud et al., 2025; Yuan et al., 2024). One study (Liu et al., 2026) does recog- nize the importance of change, scoring models against WVS waves 5 through 7 and reporting that alignment is highest at the most recent wave for the five coun- tries with which the models already align well. This is early evidence that models carry a recent picture of a society, but it leaves open whether and to what extent they capture change when prompted appropriately. Cultural alignment studies often draw on how cul- ture is conceptualized in social sciences while neglect- ing an important part of that understanding, which defines culture as a process that is transmitted and revised over time, rather than being a fixed property of a nation (Hong et al., 2000; Kashima et al., 2019; Oyserman, 2011). Treating a society as unchanging can also foster an essentialist view of social groups where characteristics observed at one point in time are taken to be stable features of the group (Haslam et al., 2000). This shortcoming could be consequential as cultural evaluations are incorporated into model doc- umentations and assessments (Solaiman et al., 2025), and cultural prompting is recommended as a strategy for improving alignment on these evaluations (Tao et al., 2024). Thus, cultural alignment evaluations re- quire asking both whether models distinguish among societies and whether they capture how those societies change over time. In this paper, we distinguish between snapshot align- ment and change alignment. Snapshot alignment asks whether an LLM locates a society correctly at a recent point in time by comparing its elicited values with that society’s most recent results in the World Values Sur- vey (WVS), the largest and longest-running empirical study of cultural values across countries. Change align- ment captures whether and to what extent an LLM can capture how a society’s values have changed over time, in terms of direction, rate of change, and re- versals, again compared to the WVS. WVS items are commonly reduced to the two dimensions of the In- glehart–Welzel cultural map (traditional vs. secular- rational values and survival vs. self-expression val- ues). The first dimension captures the differences in the importance placed on religion, obedience to au- thority, and traditional family norms, and the second dimension contrasts a focus on economic and physi- cal security with a focus on tolerance, gender equal- ity, participation, and quality of life. We test four models—Claude Opus 4.8, GPT-5.5, Gemini 3.6 Flash, and Qwen3.6-Plus—on ten core WVS indicators for 40 countries across Waves 4–7, using coordinates esti- mated directly from the survey microdata as our em- pirical benchmark. We begin by asking whether models place societies close to their most recent surveyed positions (snapshot alignment). We then look at time, asking which survey wave a model’s representation of a country most closely resembles when no year is given in the prompt and how far behind the most recent survey position it is (iden- tifying its “implicit anchor” year). We next evaluate the extent to which explicitly naming the year moves a model’s representation closer to the country’s sur- veyed position in that year. We examine whether mod- els recover the direction and rate of observed cultural change and, importantly, whether their movement cor- responds to when that change actually occurred rather than simply following the broad modernization trend shared across many societies. Finally, for societies whose surveyed trajectories reverse, we ask whether models reproduce that reversal or continue moving the country in its earlier direction. 2 Materials and Methods Study design We compared LLM representations of national culture against repeated survey measurements in order to as- sess the extent to which LLMs captured cultural values for a given country at a single point in time (snapshot alignment) versus over time (change alignment). We drew our empirical benchmark data from the World Values Survey (Inglehart et al., 2022), eliciting LLM representations for the same items. We then projected survey and LLM estimates onto the same coordinate system so that the two can be directly compared for every country and wave. Data Survey data and country sample We used WVS trend data for Waves 4 through 7, cov- ering fieldwork years from approximately 1999 to 2023. For estimating the trajectory of change parameters (i.e., direction, rate and reversal), we required at least three time points. Therefore, we only included coun- tries that had been surveyed at least three times during this period and were also part of Wave 7, which served as our reference point for snapshot alignment. We did not require all countries to have been surveyed in every single wave, as this would have biased our sample to- wards countries with more resources and better cover- age. Moreover, we did not exclude countries that were missing data for some WVS items in some waves. This was because the missing items were concentrated on politically and morally sensitive questions in precisely the authoritarian, religious, and post-conflict societies whose trajectories are of most interest. Therefore, in- cluding the item completeness requirement would have selected the sample on a variable correlated with the estimated outcome. We accounted for uneven item coverage by keeping the indicators consistent within each country, scoring model outputs on a country-by- country basis, and checking the results with sensitivity analyses. Applying the screen to the microdata yields 40 coun- tries and 134 country-waves, with 14 countries ob- served at four waves and 26 at three. Countries were selected based on their WVS coverage, without con- sidering the shape of their trajectories or their repre- sentation in the anglophone information environment. We classified their trajectories only after the sample had been selected, using the procedure described be- low. The full country list, wave coverage, and fieldwork years are provided in the SI. Models and elicitation We elicited cultural representations from four large lan- guage models: Claude Opus 4.8, GPT-5.5, Gemini 3.6 Flash, and Qwen3.6-Plus. We used the ten key indicators from the WVS with their response scales to query our models. Eight of the indicators were ordinal single-scale items (for example, a 1–10 justifiability scale or a 1–4 national-pride scale). We also included the child-qualities card sort, where participants were shown a set of cards with different qualities that children can learn at home, and asked to select up to five of these cards. Four of these qual- ities (independence (A029), determination and perse- verance (A039), religious faith (A040), and obedience (A042)) were used to create a map, and they were scored based on how often people mentioned them. Because respondents could select only five qualities, choosing one affected which of the other options they could choose. Therefore, we presented all the qualities together, like a card sort, instead of asking the mod- els to rate each one separately. The full set included ten qualities in Waves 4–6 and eleven in Wave 7. Fi- nally, we also included the WVS post-materialism in- dex (Y002), in which respondents had to pick the two most important national goals from a list of four op- tions. To elicit LLM-based representations on these indica- tors, we prompted a set of four models on these ten indicators. All queries used the same system prompt, held constant across models, conditions, and waves, which stated the task and instructed the model to an- swer as a probability distribution over the response op- tions (full instrument in SI). Every item in the instru- ment was prompted to the models in three different an- choring conditions. C0 named neither country nor year (“a population of adults”) and was elicited once per model as a country-agnostic baseline. C1 named the country but not the year, eliciting the model’s present- tense representation. C2 named both the country and the fieldwork year of a given wave and was elicited once for each retained country-wave. For each item, we asked models to give a probability distribution over the response options rather than re- peatedly sampling individual answers and using those samples to construct a distribution. This approach, known as a verbalized distribution, has been shown to recover human opinion distributions more accurately than repeated sampling or token log-probabilities over the response options (Meister et al., 2025). Repeated sampling is also not well suited to our purpose be- cause variation across model responses does not nec- essarily represent variation in the population. This approach also appears to be relatively insensitive to prompt wording. Using WVS items across countries, Liu et al. (2026) report correlations above 0.96 be- 3 tween distributions produced with their default and perturbed prompts. Given that a model does not state exactly the same distribution every time it is asked, we repeated each query k times and took the mean stated distribution as the estimate for each cell (model, coun- try, condition, item, and wave), with k = 10. Responses were treated as usable when they parsed to a well-formed distribution, either exactly or after renormalization within tolerance. Malformed, empty, or refused responses were discarded and re-elicited. For the ordered-pair post-materialism item, models oc- casionally returned distributions whose probabilities summed slightly above one. We renormalized these within a fixed tolerance and applied the identical nor- malization rule uniformly across all four models. Re- sponse validation and normalization rules are specified in full in SI. Cultural map coordinates We projected each country-wave and each model’s answers for that country-wave onto the two axes of the Inglehart–Welzel cultural map, traditional versus secular-rational (Dimension 1) and survival versus self- expression (Dimension 2). We followed the established Inglehart–Welzel map in selecting the ten core indica- tors, assigning their thirteen component variables to the two dimensions, and determining how each compo- nent was scored (Inglehart & Baker, 2000; Inglehart & Welzel, 2005). Each dimension’s components, includ- ing full item wording, WVS variable codes, response scales, and orientation indications, are provided in SI. We calculated each dimension using the equal-weight mean of its standardized components. Equal weighting is a rule that can be applied identically to survey re- spondents and model outputs, ensuring that both sides arrive at the map using the same calculations and that any discrepancy between them reflects the responses rather than the scoring. WVS data and LLM responses differ only in how a single item value is obtained. On the survey side, an item value is a respondent’s coded answer, with survey-defined missing codes set to missing. On the model side, each model’s mean stated distribution for an item was converted into a value comparable to the corresponding WVS respondent mean. For the eight ordinal items, the item value was the expected value of the stated distribution (the sum of each re- sponse option multiplied by its probability), which is directly comparable to the population mean of that item. For the child-qualities card sort, we derived the four child-autonomy components as mentioned propor- tions, taking the model’s stated selection probability for each of the four scored qualities. Because the deck differs across waves, we mapped each quality to its wave-specific position so that the correct qualities were scored. For the post-materialism item, we converted the model’s distribution over ordered goal-pairs into the WVS post-materialism index value: each ordered pair was assigned to the index category corresponding to how many of its two goals were post-materialist, and the item value was the probability-weighted mean on the resulting 1–3 index. From item values onward, the procedure is identi- cal for both sides. Each component variable was z- scored using per-item means and standard deviations computed once over the pooled WVS respondent sam- ple and sign-oriented so that higher values correspond to the secular-rational and self-expression poles. On the survey side, this standardization is applied at the respondent level and averaged to a country-wave na- tional mean using the survey design weight (S017). Given that standardization is linear, this is equivalent to standardizing the weighted national mean, which is the form the model side takes. Each dimension score is the equal-weight mean of its sign-oriented standardized components. To keep a country’s coordinate compara- ble across its own waves and to keep each model’s coor- dinate comparable to the empirical benchmark for that country, we applied a constant within-country indica- tor base, where a component was retained for a country only if non-missing for more than half of respondents in all of that country’s waves, and the same per-country base was applied when scoring model outputs. Nine countries had one or more components excluded, and their coordinates rest on fewer indicators and are less precisely estimated. We assessed the consequences by repeating the timing analysis on the countries with the complete battery (see Results). This yields, for each country-wave, one Dimension 1 and one Dimension 2 coordinate for the survey and one for each model in each condition, on identical axes. Trajectory classification We treated each country’s ordered wave coordinates as a path in the two-dimensional cultural space and de- rived four quantities from it that allowed us to classify each country’s trajectory shape and measure how far and in what direction models moved that country. Step length is the distance between consecutive waves; net displacement is the straight-line distance from a coun- try’s first surveyed position to its last; total path length is the sum of the step lengths, which exceeds net dis- placement whenever a path bends; and the turn angle at each interior wave is the angle between the incom- ing and outgoing steps. Straightness, net displacement divided by total path length, equals 1 for a path that never changes direction and approaches 0 for one that returns toward its origin. This is the straightness index of trajectory analysis, also known as the net-to-gross displacement ratio (Benhamou, 2004). 4 Sampling uncertainty was quantified by a nonpara- metric bootstrap. Within each country-wave, we re- sampled respondents with replacement, carrying the survey design weights, and recomputed that country’s coordinates and the geometry of its entire path on each replicate. In this method, the surveyed values remain fixed across replicates, while the sample of respondents changes. Thus, the bootstrap distribution of straight- ness describes how precisely a country’s path shape is determined given its sample sizes, and a country whose interval spans a wide range has a shape the survey data do not resolve. Classification followed a two-stage rule, with a sub- stantive verification stage applied afterward to the re- versal cell. First, countries whose total path length per decade fell below 0.13 standardized units were classi- fied as low-change/stable. The floor is expressed rela- tive to the scale of the map. The average country sits 0.40 units from the global centroid, so a country at the floor covers roughly a third of that distance over a decade. We repeated the classification at floors of 0.10 and 0.16 and report the consequences in SI. Sec- ond, the remaining countries were classified from the bootstrap distribution of straightness: reversal if the interval’s upper bound fell below 0.5, directional if its lower bound exceeded 0.7, and ambiguous otherwise. Using the interval rather than the point estimate helps avoid treating small shifts due to sampling noise as real changes in direction. Both cutoffs were fixed before the intervals were computed. Because they leave a gap be- tween 0.5 and 0.7, a country is labeled only when its en- tire bootstrap interval falls below 0.5 or entirely above 0.7. A country whose interval falls anywhere in be- tween or straddles either cutoff is left ambiguous rather than assigned to whichever class is nearer. Because these are hard thresholds on bootstrap percentiles, a country whose relevant bound sits near a cutoff could, in principle, flip class across Monte Carlo draws, so we repeated the classification under five bootstrap seeds and demoted to ambiguous any label that changed. At 2,000 replicates none did. Two near-boundary cases are recorded: Japan’s straightness upper bound sits just below the reversal cutoff under every seed (0.490 to 0.498), and Germany’s lower bound is just above the directional cutoff (0.706 to 0.712), with Germany’s di- rectional status additionally checked in a with/without sensitivity run. Stage 2 yielded four stable countries, ten directional countries, seven candidate reversals, and nineteen am- biguous countries. The bootstrap resamples respon- dents within waves and therefore cannot distinguish genuine doubling-back from a turn that is merely or- thogonal, since a right-angle turn also lowers straight- ness. We verified each candidate reversal on two fur- ther criteria. First, we required the turn at the bend to be obtuse, taking angles near or above 120° as back- tracking and angles near 90° as an orthogonal shift that only reads as low straightness. All seven candi- dates cleared this requirement, with turns from 129° to 169°. Second, we required shape concordance with an independent operationalization of the same dimen- sions: Welzel’s secular values and emancipative values indices (Welzel, 2013), standardized on the same scale, from which we recomputed each country’s path and its straightness. The question is whether a separately con- structed measure also bends. Six candidates do, with straightness on the Welzel path ranging from 0.13 to 0.70 (Indonesia 0.13, India 0.21, Zimbabwe 0.23, Jor- dan 0.35, Japan 0.40, South Korea 0.70). Vietnam’s Welzel path runs essentially straight at 0.97, suggest- ing that the apparent reversal is specific to the thirteen- item composite, but its three-wave span does not allow for an additional comparison. We reclassified Vietnam as ambiguous, giving the twenty ambiguous countries reported in (Table 1). Japan is retained despite its near-boundary bootstrap margin, its 129° turn on the survival–self-expression axis corroborated by a concor- dant Welzel path. Analysis Snapshot alignment We define snapshot alignment as the Euclidean dis- tance in the two-dimensional map plane between a model’s coordinate for a country and that coun- try’s most recent WVS coordinate, computed for the country-only condition (C1) and, as a year-anchored comparison, for C2. To detect directional bias, we de- composed the gap into its signed per-dimension com- ponents. We also tested whether models compress the map, representing countries as more similar to one an- other than they are. For each dimension, we computed the standard deviation of model coordinates across countries and divided it by the standard deviation of the corresponding empirical benchmark coordinates so that a ratio of 1 indicates a model separates countries as widely as the survey record does and a ratio below 1 indicates it places them closer together. Change Alignment Change alignment comprises the direction of a model’s movement, its rate, and its timing. For direction and rate, we computed the net displacement of each path (the start-to-end vector for the model and the empiri- cal benchmark) and took the cosine between the two vectors and the ratio of their magnitudes. Because the WVS country paths all mostly drift in the same dominant direction to the secular and self- expression poles, a model that was only able to move every society in this same direction would get high net displacement scores for tracking this very same general 5 direction, regardless of not actually tracking any of the countries in particular. Consequently, we computed alignment step by step: for each pair of consecutive waves, the cosine between the model’s step and the true step, averaged over a country’s steps. We mea- sured this against a permutation null (Ernst, 2004), where for each step of the model, we paired it with a random survey year (i.e., a permutation that pre- served the model’s general direction but otherwise was completely random with respect to the true country- specific paths and their corresponding years of change). We refer to this stepwise alignment subtracted by the corresponding permutation null mean as the ‘timing signal,’ which we aggregate and then bootstrap across countries. A timing signal close to zero means that a model, even if it captures change for a country, in fact conveys no information as to when in that country’s particular path the change occurred. Reversal reproduction For each of the six reversal countries, we calculated the turn angle at every interior wave. The angle was defined by the incoming and outgoing steps, with 0° indicating continuation in the same direction and 180° a complete reversal. We located the country’s reversal at the wave with the largest turn in the WVS trajec- tory and then measured the model’s turn at that same wave. Turn angle alone, however, does not show how much movement occurred. Two very short steps can form a sharp angle even when the model’s position changes very little. We therefore counted a reversal as repro- duced only if the model turned at least 120° and the combined length of the two steps around the turn was at least half the distance observed in the WVS trajec- tory. Two of the six countries had only three survey waves, so their trajectories contained a single interior wave at which a turn could be measured. Results The analytic sample is 40 countries and 134 country- waves (Table 1). For each item, we asked the mod- els to provide a probability distribution over the re- sponse options and averaged the distributions across ten independent elicitations. We used this approach rather than repeatedly sampling individual responses from the models following research showing that these verbalized distributions recover human opinion distri- butions more accurately than repeated sampling or to- ken log-probabilities (Liu et al., 2026; Meister et al., 2025). They can also be obtained from closed APIs, unlike token log-probabilities, which allowed us to use the same elicitation procedure for all four models. We prompted each model under three conditions that differed in whether the country and year were specified. C0 named neither the country nor the year, C1 named the country only, and C2 named both the country and the fieldwork year of a given wave. We use C1 to examine how a model represents a country’s values when no time period is specified. This condition is used for the snapshot and implicit-anchor analyses. For the snapshot analysis, we also report C2 to exam- ine whether adding the year changes how closely the model matches the country’s most recent surveyed po- sition. The change analyses rely on C2 because com- paring change over time requires the model and WVS data to refer to the same country and year. Finally, C0 provides a baseline for how models respond when no information about the country or year is given. Before comparing model and country positions, we established the scale of the cultural map. Country co- ordinates fall roughly within±0.70 standardized units on each axis, and the average country sits 0.40 units from the global centroid. These values provide a refer- ence for the model-country distances reported below. Because the analyses that follow examine how coun- tries move through the cultural map, we should first clarify what we mean by movement. A country’s po- sition along the two Inglehart–Welzel dimensions in a given survey wave can be compared to its position at the previous wave, enabling us to measure both the magnitude and direction of the change. We can then compare this observed movement with the movement produced by an LLM prompted to produce a response distribution for the same country and years. To evalu- ate how countries moved over time and how LLM rep- resentations compare to real-world WVS results, we calculated both the distance between coordinates on the Inglehart-Welzel map and the angle of change from one survey wave year to the next. An angle of 45° indicates some directional agreement, 90° means the two movements are perpendicular, and 180° means the model moves the country in the opposite direction. A schematic of both measures is provided in SI. We measured snapshot alignment as the distance between a model’s coordinate for a country and that country’s most recent surveyed position. We also iden- tified the implicit anchor, or the survey wave whose coordinate was closest to the placement a model gave when asked about a country without specifying a year (C1). We measured change alignment in three ways. First, for direction and timing, we calculated the co- sine between a model’s step from one wave to the next and the country’s true step over the same period and compared this with a permutation null that shuffled the model’s steps across periods. Second, we measured rate as the ratio of the distance a model moved a coun- try to the distance that country actually moved. Third, for those countries with major reversals, we measured 6 the model’s turn angle at the same wave where the country’s sharpest documented turn occurred. Snapshot alignment Given a query that named only the country (C1), all four models place countries close to their true recent position. Mean distance to a country’s most recent WVS coordinate ranged from 0.153 to 0.238, against an average country-to-centroid distance of 0.40, indi- cating broadly accurate representation of each coun- try’s current values. These representations are, more- over, a significant improvement in accuracy compared to the naive baseline in which no country is named (C0). When no country was named, models sat 0.49 to 0.62 unit from the truth, and naming the country re- duced that distance by 0.32 to 0.38 units (Claude 0.316 [0.256, 0.378], GPT 0.357 [0.284, 0.432], Qwen 0.378 [0.296, 0.463], Gemini 0.379 [0.307, 0.457]). Naming the country also relocated the answer by 0.39 to 0.55 units across the map (Claude 0.390, Qwen 0.453, GPT 0.486, Gemini 0.545), which exceeds the average coun- try’s 0.40-unit distance from the global centroid. Accuracy varied more between countries than across models. Mean placement error ranged from 0.09 for Iran to 0.40 for Thailand, with four of the five coun- tries that models placed least accurately located in East and Southeast Asia (Thailand, Taiwan, South Ko- rea, and Japan). Comparing models, Gemini placed countries closest to their most recent surveyed posi- tion (0.153, 95% CI [0.124, 0.182]), but its advantage over Claude was not statistically significant (difference 0.024 [−0.007, 0.055]). Claude (0.177 [0.146, 0.208]) and GPT (0.193 [0.160, 0.224]) were statistically indis- tinguishable from each other, while Gemini was signif- icantly closer to the surveyed position than GPT (dif- ference 0.040 [0.017, 0.065]). Qwen was the least ac- curate (0.238 [0.202, 0.278]), performing significantly worse than each of the other three under the paired bootstrap. Notably, the inclusion of the year in the model query had little effect on snapshot accuracy. Naming the country alone (C1) put models about as close to the recent snapshot as naming the country and year to- gether (C2) (0.153 versus 0.151 for Gemini, 0.177 ver- sus 0.182 for Claude, 0.193 versus 0.184 for GPT, and 0.238 versus 0.231 for Qwen). Model responses included distortions on both dimen- sions of the cultural map. On the self-expression axis, three of the four models placed countries significantly higher than the WVS records: Qwen by +0.118 ([0.067, 0.169]), GPT by +0.085 ([0.048, 0.124]), and Claude by +0.043 ([0.010, 0.076]). Gemini showed no signifi- cant displacement on this axis (0.000 [−0.029, 0.031]). The pattern differed on the tradition–secular axis. Claude (−0.049 [−0.098,−0.001]), Gemini (−0.047 [−0.093,−0.004]), and GPT (−0.046 [−0.092,−0.002]) placed countries significantly toward the traditional pole, while Qwen placed them toward the secular pole (+0.067 [0.018, 0.114]). Qwen differed significantly from each of the other three models on this axis (dif- ferences of 0.113 to 0.116, with all intervals exclud- ing zero), while Claude, Gemini, and GPT did not differ significantly from one another. Thus, the self- expression bias appears in the same direction across models when it is present, while the tradition–secular bias differs by model. Finally, we asked whether models compress the cultural map by placing countries closer relative to their placement according to the WVS data. Com- paring the spread of each model’s coordinates with the empirical spread, we found evidence of com- pression in three of the eight model–axis combina- tions. Claude compressed the tradition–secular axis (ratio 0.65 [0.59, 0.73]) and, more marginally, the survival–self-expression axis (0.88 [0.80, 0.99]), while Qwen compressed the survival–self-expression axis (0.77 [0.67, 0.93]). No other interval differed signifi- cantly from 1. Overall, the models generally preserved the differences between countries observed in the WVS rather than pulling them toward a common center, al- though there was some evidence of compression for Claude and Qwen. Implicit anchor We defined a model’s “implicit anchor” for a country as the survey wave closest, in projected space, to the answer the model gives when asked about a country with no year given. We restricted this analysis to the 36 countries with measured movement. Averaged over those countries, every model’s answer fit the most recent wave best, with fit degrading steadily as the waves got older (Fig. 1A). However, assigning each country to its own closest wave, rather than averag- ing distances across countries, gave a more nuanced picture. The most recent wave was the best-fitting wave for only 36% of countries for Claude and 42% for GPT, compared with roughly 57% for Gemini and Qwen (Fig. 1B). To further examine how far back each model’s rep- resentation falls when it is closer to an earlier wave (versus the most recent wave), we measured the num- ber of years between each country’s most recent survey and the survey wave closest to the model’s representa- tion. Averaged across countries, Claude’s anchor was 7.0 years behind the most recent survey (95% CI [4.9, 9.3]), followed by GPT at 6.0 years [4.0, 8.1], Qwen at 4.7 years [2.6, 7.0], and Gemini at 4.5 years [2.6, 6.6]. Because WVS waves are unevenly spaced, we also mea- sured the lag in number of survey waves rather than years. Using this metric, Claude anchored an average 7 Table 1 Final Trajectory Classification of the 40 Countries Trajectory classn Countries Reversal6 Indonesia, India, Jordan, Japan, South Korea, Zimbabwe Directional10 Germany, Thailand, Taiwan, Canada, Kyrgyzstan, Romania, Turkey, Serbia, Brazil, Netherlands Low- change/Stable 4 Russia, Colombia, Philippines, Cyprus Ambiguous20 Egypt, Iraq, Morocco, Malaysia, Hong Kong, Chile, Pakistan, Iran, China, Aus- tralia, Ukraine, United States, Argentina, Nigeria, Mexico, Vietnam, New Zealand, Peru, Uruguay, Singapore of 0.97 waves behind each country’s most recent sur- vey [0.69, 1.25], GPT 0.89 [0.61, 1.19], Qwen 0.69 [0.39, 1.03], and Gemini 0.67 [0.39, 0.94]. Timing and rate Models generally shifted countries in the same broad di- rection as the WVS data, with median direction agree- ment ranging from 0.71 to 0.89 across models. The an- gle between the model’s movement and the country’s actual movement was less than 90° for 83% to 92% of countries. Using a stricter criterion of 45° reduced this proportion to 50–69%. No model reflected the full amount of societies’ changes. Figure 2 compares the observed and model- estimated net displacement for all 40 countries, sep- arately for Gemini (A), Claude (B), GPT (C), and Qwen (D). For each country, we compared how far the model moved that country’s position with how far the country’s position changed in the WVS data. We ex- pressed this as a ratio, where 1.0 means that the model captured the full magnitude of the observed change. Among the ten countries with a clear directional path, every model captured only part of the observed change, with median ratios of 0.45 for Claude, 0.56 for Gemini, 0.58 for GPT, and 0.68 for Qwen. This under-travel re- mained significant across all 36 movers for Claude (0.61 [0.43, 0.86]) and GPT (0.69 [0.55, 0.97]) but not for Gemini (0.82 [0.64, 1.18]) or Qwen (0.90 [0.70, 1.03]). Movement also appeared where little was observed in the survey data. The four low-change countries, shown as open triangles in Fig. 2, moved a median of 0.097 units. Every model moved these countries farther, from 0.113 to 0.170 units. As Fig. 2 shows, the low-change countries fall above the equality line for all four models, indicating more model-estimated movement than observed movement. With only four countries in this group, however, we treat this pattern as descriptive rather than as a tested effect. These measures compared a country’s starting and ending positions, but nearly every country in our sam- ple moved in the same broad direction over this pe- 0.15 0.20 0.25 0.30 Most recent 1 wave earlier 2 waves earlier Survey wave, relative to each country's most recent Mean distance to true position Gemini 3.6 FlashClaude Opus 4.8GPT-5.5Qwen3.6-Plus A 20961 1312101 151272 21843 Qwen3.6-Plus GPT-5.5 Claude Opus 4.8 Gemini 3.6 Flash 0%25%50%75%100% Share of the 36 moving countries Most recent1 wave back2 waves back3 waves back B Figure 1: The implicit temporal anchor. (A) Mean distance between each model’s placement when no year is specified and the country’s WVS coordinate at each survey wave, calculated across the 36 countries with measured movement. Survey waves are indexed rela- tive to each country’s most recent survey, and lower values indicate a closer match. (B) The closest survey wave to each model’s placement for each of the 36 coun- tries. Numbers show the number of countries assigned to each wave. 8 riod, toward the secular and self-expression poles. As a result, a model could score high by reproducing this general trend without capturing any country’s par- ticular path. To distinguish between these possibili- ties, we examined whether the model’s movement also corresponded to when the observed change occurred. We did this with a permutation test that treats each model’s trajectory as a series of steps between consec- utive survey waves. We first checked how closely each model step matched the country’s real step over the same years. We then shuffled the model’s steps across periods (for example, comparing a step the model made between 2005 and 2012 with what the country actually did between 2012 and 2017) and checked the match again. Shuffling preserves the model’s overall direction of movement but breaks the correspondence between its steps and the periods in which the coun- try’s observed changes occurred. If the model captures when a country changed, its alignment with the WVS trajectory should decline after shuffling. If it captures only the country’s general direction of change, shuffling the steps across periods should make little difference. Two models did better in order than shuffled, Gem- ini (+0.123 [0.047, 0.203]) and Claude (+0.085 [0.001, 0.173], only just clearing zero), while GPT (+0.049 [−0.019, 0.114]) and Qwen (+0.003 [−0.051, 0.064]) showed no significant difference. Qwen moved coun- tries farther than any other model, with a median of 0.68 of the true distance on the directional paths, yet showed no timing signal under any specification. We assessed the robustness of these results by run- ning two sensitivity analyses. The first concerned the number of steps available per country. The shuffle can only be informative if a country has enough steps to reorder, and most countries were surveyed three times, giving just two steps that can be shuffled only by swap- ping them. We therefore restricted the test to the 14 countries surveyed four times, which give three steps each. The timing signal is larger in this subset, and three of the four models clear zero: Gemini (+0.204 [0.066, 0.338]), Claude (+0.158 [0.017, 0.300]), and GPT (+0.101 [0.012, 0.200]), with Qwen again show- ing none. The second concerned how many indicators each country’s coordinates rested on. Nine countries are missing at least one of the thirteen component vari- ables, so their positions on the map are estimated less precisely. Dropping those nine left only Gemini clear- ing zero (+0.096 [0.008, 0.189]). Therefore, Gemini is the only model whose timing signal holds in every specification. We also asked whether naming the year in the query brings the model’s response closer to the country’s sur- veyed position for that year. For each country, we com- pared the model’s answer under C2, which names the fieldwork year of the most recent survey, with its an- swer under C1, which names no year, and measured 0.0 0.2 0.4 0.6 Distance the model moved it 0.00.20.40.6 Gemini 3.6 Flash A 0.0 0.2 0.4 0.6 0.00.20.40.6 Claude Opus 4.8 B 0.0 0.2 0.4 0.6 Distance the model moved it 0.00.20.40.6 Distance the country actually moved GPT-5.5 C 0.0 0.2 0.4 0.6 0.00.20.40.6 Distance the country actually moved Qwen3.6-Plus D Figure 2: Models under-travel where change hap- pened and drift where it did not. (A–D) Each point represents one country, with its observed net dis- placement on the cultural map between its first and last survey waves plotted against the corresponding net displacement in the model’s representation. All 40 countries are shown: filled circles indicate the 36 countries with measured movement, and open trian- gles indicate the four countries classified as low-change. The dashed diagonal marks equal displacement in the model and WVS data; points below the line indicate that the model captured less movement than observed, while points above the line indicate more movement. Most countries fall below the line, while the four low- change countries fall above it. whether the year brought the answer closer to that year’s true position. The difference was close to zero for every model, from−0.005 for Claude to +0.009 for GPT. A signed-rank test was null for three models and marginal for GPT, which had a median gain of +0.016 units (p = .026), but this does not survive resampling of the countries. Even at face value, that gain is about a tenth of the 0.18 that still separates GPT’s answer from the truth. Naming the most recent fieldwork year rather than no year should not, on its own, move a model far, since both cues point to roughly the same period. What the null result showed is that the year is not doing independent work. Naming the recent year does not pull a model forward out of the lag docu- mented above, and the explicit date neither refines its answer nor moves it toward the surveyed position for that year. 9 Reversal reproduction Finally, we examined the six countries whose WVS tra- jectories were verified as reversing (Table 1; Fig. 3), using the year-anchored condition (C2). At the wave where each country’s path turns, we measured the an- gle between the incoming and outgoing steps so that a path continuing straight ahead scores 0° and one dou- bling directly back scores 180°. All six real reversals were sharp, from 129° to 169° (median 155°). Models were more successful at reproducing the change in direction than the amount of movement through a reversal. Of the 24 country–model cases, seven produced a sharp turn (≥ 120 ◦ ) at the same wave as the observed reversal. These were Claude for India, Japan, and Zimbabwe; Gemini for Jordan and Japan; and GPT for Indonesia and Japan. Requiring a reproduction to be both a sharp turn and a journey at least half as long as the real one (the summed length of the two steps at the bend at least 50% of the WVS’s) leaves four cases: Japan (Claude and GPT), Indonesia (GPT), and Jordan (Gemini). Claude turned 150° at Zimbabwe’s 2012 bend, almost exactly reversing course at the right moment but covering only 19% of the ground Zimbabwe actually covered. Across all 24 cases, the models covered about half the distance the coun- try actually traveled through the bend, close to the under-travel seen on the directional paths. In the four cases where models did accurately reflect a shift, we observed no obvious pattern. Decomposing that movement by axis suggests the shortfall is not evenly distributed, at least for one model. At the bend, Claude reproduced 0.92 of the true movement along the traditional–secular axis but only 0.32 along survival–self-expression. The other three models did not show this pattern. Gemini tilted weakly in the opposite direction (0.27 and 0.55), GPT was close to even (0.33 and 0.48), and Qwen was nearly symmetric (0.56 and 0.55). Turn angle and distance traveled came apart most clearly in Claude. Its median turn at the true bend was 128°, against a true median of 155° and 37° to 58° for the other three models, yet only one of its three sharp turns also cleared the distance criterion. Claude got the timing of the reversal right more often than the other models, but it mostly changed a country’s direction without moving it very far. Qwen sat at the other extreme, never turning sharply anywhere, with the largest angle at a true bend of 99°. Discussion Our research shows that current SOTA LLMs perform poorly at capturing how cultures change over time. Al- though the LLMs we tested placed countries close to the positions those countries occupied in their most re- cent WVS wave, they made only limited adjustments when prompted to account for year as well as place. Even snapshot accuracy, though strong, varied across countries, with four of the five largest errors occur- ring in East and Southeast Asia. Given that most of these countries have official languages with substan- tial web-based pretraining data (Nguyen et al., 2024), the pattern does not appear to follow a simple distinc- tion between high- and low-resource languages. That is, data availability alone may not explain why models represent some cultures more accurately than others. More generally, LLMs tended to understate the mag- nitude of cultural change over time, introduce move- ment even for societies that changed little, and rarely reproduce genuine reversals in countries’ trajectories. Explicitly specifying a year did little to correct these errors. Together, these results suggest that LLMs cap- ture national cultures more strongly as positions than as trajectories. Our analysis of the WVS reinforces the importance of this distinction. Among the 40 coun- tries in our sample, only four met our criterion for low change. Therefore, for most societies in our sample, the most recent observed position cannot be assumed to describe an enduring state. Our findings reveal a noteworthy difference between the current practices of studying cultural representa- tion in LLMs and how culture is understood in social and cultural psychology. The existing evaluations of LLMs are largely focused on testing the models’ abil- ity to distinguish among cultures by looking at whether models reproduce country-level values (Kharchenko et al., 2024), recognize culturally specific knowledge and norms (Myung et al., 2024), or adapt their responses when given cultural context (Adilazuarda et al., 2024; Kozlowski & Evans, 2025). These are important dimen- sions of cultural competence, but they primarily test variation between cultures at a given point in time. On the other hand, cultural psychology has long empha- sized that culture is not simply a collection of stable characteristics attached to groups but instead a sys- tem of meanings and values that are transmitted, re- constructed, and revised as people respond to changing social environments (Inglehart, 2018; Kashima, 2014; Oyserman, 2017). Therefore, knowing how societies differ from one another at a single moment captures only part of what it means to represent culture ac- curately. Our findings show limitations in the extent to which LLMs capture variation within cultures over time, offering both methods for and substantive insight on an important but understudied dimension of cul- tural awareness evaluations. An accurate representation of a society today may become outdated as that society changes. Whether an earlier snapshot remains informative depends in part on what has happened before because a society that has remained relatively stable provides a different ba- 10 IndonesiaIndiaJordanJapanSouth KoreaZimbabwe Traditional → secular-rational Survival → self-expression 456745674567456745674567 -0.5 0.0 0.5 -0.50 -0.25 0.00 0.25 WVS wave World Values SurveyGemini 3.6 FlashClaude Opus 4.8GPT-5.5Qwen3.6-Plus Figure 3: Model trajectories through cultural reversals. The six columns show the countries whose WVS trajectories reversed, with traditional–secular-rational values in the top row and survival–self-expression values in the bottom row. Dark lines trace the WVS scores across Waves 4–7, and the colored lines show the corresponding model scores from the year-anchored condition (C2). Dashed lines mark the wave of the sharpest turn in each country’s WVS trajectory. The vertical scale is the same across countries within each row. WVS waves are shown at equal intervals even though the time between surveys varies, so the lines show how much values changed from one wave to the next, not the rate of change per year. sis for judging its current or future values than one that has been moving steadily or has changed direc- tion. Thus, in addition to keeping a model up-to-date, representing cultural change requires a model to have some record of a society’s past and how it got from there to here (Yin et al., 2024). Our results suggest that current models have a much weaker representa- tion of this trajectory than of a society’s most recent position. The consequences of failures in LLM cultural aware- ness are especially important in the context of grow- ing academic and commercial interest in using LLMs as proxies for people. Recent research shows that synthetic participants can flatten differences between groups (Wang et al., 2025), misrepresent minority pop- ulations (Kozlowski & Evans, 2025), amplify effect sizes (Cui et al., 2025), and generate outputs that dif- fer in systematic and often predictable ways from those of real humans that the model is intended to simulate (Tjuatja et al., 2024). Our findings identify “tempo- ral flattening” as a further source of potential error. This problem has particularly salient implications for work that attempts to understand and model attitude change, intergenerational differences, historical events, and, more generally, the design that is attempting to quantify change. Representing a society’s values as fixed is not sim- ply a neutral simplification. It can reproduce a form of psychological essentialism, in which social groups are understood as having an underlying and relatively immutable nature that explains what their members are like (Haslam et al., 2000). Seen in this light, a model that gives roughly the same values for a society whether it is asked about 1999 or 2023 risks turning a particular position in time into a seemingly endur- ing characteristic of a society. In this sense, the pat- tern resembles the representational harms described in the literature (Shelby et al., 2023), and adds to ex- isting concerns about cultural erasure and linguistic hegemony in these systems (Daryani et al., 2026). This work is subject to several limitations. The first limitation concerns the operationalization of na- 11 tional culture via two Inglehart–Welzel dimensions de- rived from a subset of WVS items. Although these di- mensions capture important variation in values, they simplify culture, which also encompasses norms, prac- tices, identities, meanings, and institutions (Dev et al., 2026; Kitayama & Salvador, 2024; Markus & Kitayama, 2010). Future studies should investigate whether the snapshot-change asymmetry generalizes to other cultural realms, such as moral, political, gender, and inter-group attitudes. Second, the WVS provides repeated cross-sectional observations at irregular inter- vals rather than continuous measurements of cultural change. In other words, between any two of the sur- veyed time points, any number of events might have occurred, including changes that were not captured by the WVS; however, our trajectories capture change only at the various surveyed time points. Finally, our elicitation procedure measures the cultural representa- tions that LLMs reliably express under survey prompt- ing. The absence of recovery of a historical trajectory does not in itself imply that no relevant temporal in- formation is stored in the model’s parameters. The weak effect of explicitly naming the survey year does not suggest that such information is not present in the model’s parameters, but rather that such information, if present, is not readily accessible by means of a simple temporal instruction. Taken together, our findings show that SOTA LLMs do not adequately account for cultural change. Across forty countries and two decades of surveyed change, all four of the LLMs we tested represented how country-level values moved only in part, understat- ing real change, introducing movement where the sur- veys record little, and rarely completing a reversal. The increasing use of such systems to represent pop- ulations and to mediate what those populations read about themselves underscores the need to develop eval- uations and alignment solutions to ensure they ade- quately track a culture’s change over time in addition to correctly locating it in space at any given time. Acknowledgments We thank Daphna Oyserman and Jackson Trager for providing feedback on the early version of the manuscript. Competing interest The authors declare that they have no competing in- terests. Funding The authors received no specific funding for this work. Author contributions Y.D. conceived the research question, and Y.D., M.D., and M.B. designed the study. Y.D. conducted the ex- periments and analyzed the results. M.B. and M.D. contributed to the design of the elicitation instru- ment. M.D. supervised the project. Y.D. wrote the manuscript; M.D. contributed to writing and editing, and M.B. reviewed and edited. All authors approved the final version. Data availability Elicitation outputs for all four models, the derived Inglehart–Welzel coordinates for the model and survey sides, and all analysis code are available at https:// github.com/yalda-darya/cultural-trajectories. World Values Survey microdata are not redistributed and are available from the World Values Survey Association at worldvaluessurvey.org under their terms of use. Ethics statement This study did not involve human participants and did not require institutional review board approval. It uses publicly available, anonymized secondary data from the World Values Survey under the World Val- ues Survey Association’s terms of use and model re- sponses elicited via provider APIs in accordance with their terms of service. References Adilazuarda, M. F., Mukherjee, S., Lavania, P., Singh, S. S., Aji, A. F., O’Neill, J., Modi, A., & Choudhury, M. (2024). Towards measuring and modeling “culture” in LLMs: A survey. Proceedings of the 2024 Conference on Empir- ical Methods in Natural Language Processing, 15763–15784. https://doi.org/10.18653/v1/ 2024.emnlp-main.882 AlKhamissi, B., ElNokrashy, M., Alkhamissi, M., & Diab, M. (2024). Investigating cultural align- ment of large language models. Proceedings of the 62nd Annual Meeting of the Associa- tion for Computational Linguistics (Volume 1: Long Papers), 12404–12422. https://doi.org/ 10.18653/v1/2024.acl-long.671 12 Argyle, L. P., Busby, E. C., Fulda, N., Gubler, J. R., Rytting, C., & Wingate, D. (2023). Out of one, many: Using language models to simulate hu- man samples. Political Analysis, 31(3), 337– 351. https://doi.org/10.1017/pan.2023.2 Atari, M., Xue, M. J., Park, P. S., Blasi, D. E., & Henrich, J. (2023). Which humans? [Preprint]. https://doi.org/10.31234/osf.io/5b26t Benhamou, S. (2004). How to reliably estimate the tor- tuosity of an animal’s path: Straightness, sin- uosity, or fractal dimension? Journal of Theo- retical Biology, 229(2), 209–220. https://doi. org/10.1016/j.jtbi.2004.03.016 Bulté, B., & Rigouts Terryn, A. (2026). LLMs and cul- tural values: The impact of prompt language and explicit cultural framing. Computational Linguistics, 52(2), 407–494. https://doi.org/ 10.1162/COLI.a.583 Cao, Y., Zhou, L., Lee, S., Cabello, L., Chen, M., & Hershcovich, D. (2023). Assessing cross- cultural alignment between ChatGPT and hu- man societies: An empirical study. Proceedings of the First Workshop on Cross-Cultural Con- siderations in NLP (C3NLP), 53–67. https:// doi.org/10.18653/v1/2023.c3nlp-1.7 Cui, Z., Li, N., & Zhou, H. (2025). A large-scale repli- cation of scenario-based experiments in psy- chology and management using large language models. Nature Computational Science, 5(8), 627–634. https://doi.org/10.1038/s43588-025- 00840-7 Daryani, Y., Sourati, Z., & Dehghani, M. (2026). The homogenizing engine: AI’s role in standard- izing culture and the path to policy. Policy Insights from the Behavioral and Brain Sci- ences, 13(1), 14–27. https://doi.org/10.1177/ 23727322251406591 de Roock, R. S. (2024). To become an object among ob- jects: Generative artificial “intelligence,” writ- ing, and linguistic white supremacy. Reading Research Quarterly, 59(4), 590–608. https:// doi.org/10.1002/rrq.569 Dev, S., Prabhakaran, V., Feman, R. C., Davani, A., Denton, R., Kalia, C., Lertvittayakumjorn, P., Maji, M., Qadri, R., Rostamzadeh, N., Shelby, R., Stella, R., Stepanyan, H., van Liemt, E., Verma, A., Wahltinez, O., Wornyo, E., Zal- divar, A., & Mojsilović, S. (2026). A unified framework to quantify cultural intelligence of AI. https://doi.org/10.48550/arXiv.2603. 01211 Durmus, E., Nguyen, K., Liao, T. I., Schiefer, N., Askell, A., Bakhtin, A., Chen, C., Hatfield- Dodds, Z., Hernandez, D., Joseph, N., Lovitt, L., McCandlish, S., Sikder, O., Tamkin, A., Thamkul, J., Kaplan, J., Clark, J., & Ganguli, D. (2024). Towards measuring the representa- tion of subjective global opinions in language models. First Conference on Language Model- ing. Ernst, M. D. (2004). Permutation methods: A ba- sis for exact inference. Statistical Science, 19(4), 676–685. https : / / doi . org / 10 . 1214 / 088342304000000396 Haslam, N., Rothschild, L., & Ernst, D. (2000). Essen- tialist beliefs about social categories. British Journal of Social Psychology, 39(1), 113–127. https://doi.org/10.1348/014466600164363 Hong, Y.-y., Morris, M. W., Chiu, C.-y., & Benet- Martínez, V. (2000). Multicultural minds: A dynamic constructivist approach to culture and cognition. American Psychologist, 55(7), 709–720. https://doi.org/10.1037/0003-066X. 55.7.709 Inglehart, R., & Baker, W. E. (2000). Modernization, cultural change, and the persistence of tradi- tional values. American Sociological Review, 65(1), 19–51. https : / / doi . org / 10 . 1177 / 000312240006500103 Inglehart, R., Haerpfer, C., Moreno, A., Welzel, C., Kizilova, K., Diez-Medrano, J., Lagos, M., Norris, P., Ponarin, E., & Puranen, B. (2022). World values survey: All rounds – country- pooled datafile [Dataset Version 3.0.0]. Inglehart, R., & Welzel, C. (2005). Modernization, cultural change, and democracy: The hu- man development sequence. Cambridge Uni- versity Press. https : / / doi . org / 10 . 1017 / CBO9780511790881 Inglehart, R. F. (2018). Cultural evolution: People’s motivations are changing, and reshaping the world. Cambridge University Press. https:// doi.org/10.1017/9781108613880 Jackson, J. C., & Medvedev, D. (2024). Worldwide divergence of values. Nature Communications, 15(1), 2650. https://doi.org/10.1038/s41467- 024-46581-5 Kashima, Y. (2014). Meaning, grounding, and the con- struction of social reality. Asian Journal of So- cial Psychology, 17(2), 81–95. https://doi.org/ 10.1111/ajsp.12051 Kashima, Y., Bain, P. G., & Perfors, A. (2019). The psychology of cultural dynamics: What is it, what do we know, and what is yet to be known? Annual Review of Psychology, 70, 499– 529. https://doi.org/10.1146/annurev-psych- 010418-103112 Kharchenko, J., Roosta, T. G., Chadha, A., & Shah, C. (2024). How well do LLMs represent values across cultures? empirical analysis of LLM re- sponses based on hofstede cultural dimensions. https://doi.org/10.48550/arXiv.2406.14805 13 Kitayama, S., & Salvador, C. E. (2024). Cultural psy- chology: Beyond east and west. Annual Review of Psychology, 75, 495–526. https://doi.org/ 10.1146/annurev-psych-021723-063333 Kozlowski, A. C., & Evans, J. (2025). Simulating sub- jects: The promise and peril of artificial in- telligence stand-ins for social agents and in- teractions. Sociological Methods & Research, 54(3), 1017–1073. https://doi.org/10.1177/ 00491241251337316 Kurdi, B., Charlesworth, T. E. S., & Mair, P. (2025). International stability and change in explicit and implicit attitudes: An investigation span- ning 33 countries, five social groups, and 11 years (2009–2019). Journal of Experimental Psychology: General, 154(6), 1643–1666. https: //doi.org/10.1037/xge0001746 Liu, Y., Kaneko, M., & Chu, C. (2026). On the align- ment of large language models with global hu- man opinion. Proceedings of the AAAI Confer- ence on Artificial Intelligence, 40(44), 37673– 37681. https://doi.org/10.1609/aaai.v40i44. 41102 Markus, H. R., & Kitayama, S. (2010). Cultures and selves: A cycle of mutual constitution. Perspec- tives on Psychological Science, 5(4), 420–430. https://doi.org/10.1177/1745691610375557 Masoud, R. I., Liu, Z., Ferianc, M., Treleaven, P., & Rodrigues, M. (2025). Cultural alignment in large language models: An explanatory anal- ysis based on hofstede’s cultural dimensions. Proceedings of the 31st International Con- ference on Computational Linguistics, 8474– 8503. Meister, N., Guestrin, C., & Hashimoto, T. (2025). Benchmarking distributional alignment of large language models. Proceedings of the 2025 Conference of the Nations of the Amer- icas Chapter of the Association for Computa- tional Linguistics: Human Language Technolo- gies (Volume 1: Long Papers), 24–49. Microsoft AI Economy Institute. (2026, May). Global ai diffusion: Q1 2026 trends and insights (tech. rep.). Microsoft. https://w.microsoft.com/ en-us/research/wp-content/uploads/2026/05/ Microsoft-AI-Diffusion-Report-2026-Q1.pdf Mollema, W. J. T. (2025). A taxonomy of epistemic injustice in the context of AI and the case for generative hermeneutical erasure. AI and Ethics, 5(5), 5535–5555. https://doi.org/10. 1007/s43681-025-00801-w Mora, D., Aryabumi, V., Ko, W.-Y., Hooker, S., Kreutzer, J., & Fadaee, M. (2025). The art of asking: Multilingual prompt optimization for synthetic data. https : / / doi . org / 10 . 48550 / arXiv.2510.19806 Myung, J., Lee, N., Zhou, Y., Jin, J., Putri, R. A., Antypas, D., Borkakoty, H., Kim, E., Perez-Almendros, C., Ayele, A. A., Gutiérrez- Basulto, V., Ibáñez-García, Y., Lee, H., Muhammad, S. H., Park, K., Rzayev, A. S., White, N., Yimam, S. M., Pilehvar, M. T., . . . Oh, A. (2024). BLEnD: A benchmark for LLMs on everyday knowledge in diverse cul- tures and languages. Advances in Neural Infor- mation Processing Systems, 37, 78104–78146. Nguyen, T., Nguyen, C. V., Lai, V. D., Man, H., Ngo, N. T., Dernoncourt, F., Rossi, R. A., & Nguyen, T. H. (2024). CulturaX: A cleaned, enormous, and multilingual dataset for large language models in 167 languages. Proceed- ings of the 2024 Joint International Confer- ence on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), 4226–4237. Norris, P., & Inglehart, R. (2019). Cultural back- lash: Trump, brexit, and authoritarian pop- ulism. Cambridge University Press. https:// doi.org/10.1017/9781108595841 Oyserman, D. (2011). Culture as situated cognition: Cultural mindsets, cultural fluency, and mean- ing making. European Review of Social Psy- chology, 22(1), 164–214. https://doi.org/10. 1080/10463283.2011.627187 Oyserman, D. (2017). Culture three ways: Culture and subcultures within countries. Annual Review of Psychology, 68, 435–463. https://doi.org/ 10.1146/annurev-psych-122414-033617 Pataranutaporn, P., Powdthavee, N., Archi- waranguprok, C., & Maes, P. (2025). Simulating human well-being with large language models: Systematic validation and misestimation across 64,000 individuals from 64 countries. Proceedings of the National Academy of Sciences, 122(48), e2519394122. https://doi.org/10.1073/pnas.2519394122 Santos, H. C., Varnum, M. E. W., & Grossmann, I. (2017). Global increases in individualism. Psy- chological Science, 28(9), 1228–1239. https:// doi.org/10.1177/0956797617700622 Santurkar, S., Durmus, E., Ladhak, F., Lee, C., Liang, P., & Hashimoto, T. (2023). Whose opinions do language models reflect? Proceedings of the 40th International Conference on Machine Learning, 202, 29971–30004. Shelby, R., Rismani, S., Henne, K., Moon, A., Ros- tamzadeh, N., Nicholas, P., Yilla-Akbari, N., Gallegos, J., Smart, A., Garcia, E., & Virk, G. (2023). Sociotechnical harms of algorith- mic systems: Scoping a taxonomy for harm re- duction. Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, 723– 14 741. https : / / doi . org / 10 . 1145 / 3600211 . 3604673 Solaiman, I., Talat, Z., Agnew, W., Ahmad, L., Baker, D. K., Blodgett, S. L., Chen, C., Daumé, I., Hal, Dodge, J., Duan, I., Evans, E., Friedrich, F., Ghosh, A., Gohar, U., Hooker, S., Jernite, Y., Kalluri, P. R., Leidinger, A., Lusoli, A., . . . Vassilev, A. (2025). Evaluating the social im- pact of generative AI systems. In P. Hacker, A. Engel, S. Hammer, & B. Mittelstadt (Eds.), The oxford handbook of the foundations and regulation of generative AI (p. 461–515). Ox- ford University Press. https : / / doi . org / 10 . 1093/oxfordhb/9780198940272.013.0025 Tao, Y., Viberg, O., Baker, R. S., & Kizilcec, R. F. (2024). Cultural bias and cultural alignment of large language models. PNAS Nexus, 3(9), pgae346. https://doi.org/10.1093/pnasnexus/ pgae346 Tjuatja, L., Chen, V., Wu, T., Talwalkar, A., & Neu- big, G. (2024). Do LLMs exhibit human-like response biases? a case study in survey design. Transactions of the Association for Computa- tional Linguistics, 12, 1011–1026. https://doi. org/10.1162/tacl_a_00685 Wang, A., Morgenstern, J., & Dickerson, J. P. (2025). Large language models that replace human participants can harmfully misportray and flatten identity groups. Nature Machine Intel- ligence, 7(3), 400–411. https://doi.org/10. 1038/s42256-025-00986-z Welzel, C. (2013). Freedom rising: Human empower- ment and the quest for emancipation. Cam- bridge University Press. Yin, X., Jiang, J., Yang, L., & Wan, X. (2024). History matters: Temporal knowledge editing in large language model. Proceedings of the AAAI Conference on Artificial Intelligence, 38(17), 19413–19421. https://doi.org/10.1609/aaai. v38i17.29912 Yuan, X., Hu, J., & Zhang, Q. (2024). A compara- tive analysis of cultural alignment in large lan- guage models in bilingual contexts [Preprint]. https://doi.org/10.31219/osf.io/6hpcf Zhu, Y., Yuan, H., Wang, S., Liu, J., Liu, W., Deng, C., Chen, H., Liu, Z., Dou, Z., & Wen, J.-R. (2025). Large language models for information retrieval: A survey. ACM Transactions on In- formation Systems, 44(1), 1–54. https://doi. org/10.1145/3748304 15 Supplementary MaterialsDaryani, Bogen, & Daepp Supplementary Materials Accurate in space, unreliable in time: how LLMs represent national cultural change S1. Sample Table S1 Country sample, wave coverage, and fieldwork years. CountryISO-3 WavesFieldwork years (modal S020) n per waveTotaln Components retained Trajectory class ArgentinaARG W4, W5, W6, W7 (4) 1999, 2006, 2013, 2017 1,280, 1,002, 1,030, 1,003 4,31513Ambiguous AustraliaAUS W5, W6, W7 (3) 2005, 2012, 2018 1,421, 1,477, 1,8134,71113Ambiguous BrazilBRA W5, W6, W7 (3) 2006, 2014, 2018 1,500, 1,486, 1,7624,74813Directional CanadaCAN W4, W5, W7 (3) 2000, 2006, 2020 1,931, 2,164, 4,0188,11313Directional ChileCHL W4, W5, W6, W7 (4) 2000, 2006, 2012, 2018 1,200, 1,000, 1,000, 1,000 4,20013Ambiguous ChinaCHN W4, W5, W6, W7 (4) 2001, 2007, 2013, 2018 1,000, 1,991, 2,300, 3,036 8,32710Ambiguous ColombiaCOL W5, W6, W7 (3) 2005, 2012, 2018 3,025, 1,512, 1,5206,05713Stable CyprusCYP W5, W6, W7 (3) 2006, 2011, 2019 1,050, 1,000, 1,0003,05013Stable EgyptEGY W4, W5, W6, W7 (4) 2001, 2008, 2013, 2018 3,000, 3,051, 1,523, 1,200 8,77410Ambiguous GermanyDEU W5, W6, W7 (3) 2006, 2013, 2018 2,064, 2,046, 1,5285,63813Directional Hong KongHKG W5, W6, W7 (3) 2005, 2014, 2018 1,252, 1,000, 2,0754,32711Ambiguous IndiaIND W4, W5, W6, W7 (4) 2001, 2006, 2012, 2023 2,002, 2,001, 4,078, 1,692 9,77313Reversal IndonesiaIDN W4, W5, W7 (3) 2001, 2006, 2018 1,000, 2,015, 3,2006,21513Reversal IranIRN W4, W5, W7 (3) 2000, 2007, 2020 2,532, 2,667, 1,4996,69812Ambiguous IraqIRQ W4, W5, W6, W7 (4) 2004, 2006, 2013, 2018 2,325, 2,701, 1,200, 1,200 7,4268Ambiguous JapanJPN W4, W5, W6, W7 (4) 2000, 2005, 2010, 2019 1,362, 1,096, 2,443, 1,353 6,25413Reversal JordanJOR W4, W5, W6, W7 (4) 2001, 2007, 2014, 2018 1,223, 1,200, 1,200, 1,203 4,82613Reversal KyrgyzstanKGZ W4, W6, W7 (3) 2003, 2011, 2020 1,043, 1,500, 1,2003,74313Directional MalaysiaMYS W5, W6, W7 (3) 2006, 2012, 2018 1,201, 1,300, 1,3133,81413Ambiguous MexicoMEX W4, W5, W6, W7 (4) 2000, 2005, 2012, 2018 1,535, 1,560, 2,000, 1,741 6,83613Ambiguous MoroccoMAR W4, W5, W6, W7 (4) 2001, 2007, 2011, 2021 1,251, 1,200, 1,200, 1,200 4,85112Ambiguous NetherlandsNLD W5, W6, W7 (3) 2006, 2012, 2022 1,050, 1,902, 2,1455,09713Directional New ZealandNZL W5, W6, W7 (3) 2004, 2011, 2020 954, 841, 1,0572,85213Ambiguous NigeriaNGA W4, W6, W7 (3) 2000, 2012, 2018 2,022, 1,759, 1,2375,01813Ambiguous PakistanPAK W4, W6, W7 (3) 2001, 2012, 2018 2,000, 1,200, 1,9955,19513Ambiguous PeruPER W4, W5, W6, W7 (4) 2001, 2006, 2012, 2018 1,501, 1,500, 1,210, 1,400 5,61110Ambiguous PhilippinesPHL W4, W6, W7 (3) 2001, 2012, 2019 1,200, 1,200, 1,2003,60013Stable RomaniaROU W5, W6, W7 (3) 2005, 2012, 2018 1,776, 1,503, 1,2574,53613Directional RussiaRUS W5, W6, W7 (3) 2006, 2011, 2017 2,033, 2,500, 1,8106,34313Stable S1 Supplementary MaterialsDaryani, Bogen, & Daepp CountryISO-3 WavesFieldwork years (modal S020) n per waveTotaln Components retained Trajectory class SerbiaSRB W4, W5, W7 (3) 2001, 2006, 2017 1,200, 1,220, 1,0463,46613Directional SingaporeSGP W4, W6, W7 (3) 2002, 2012, 2020 1,512, 1,972, 2,0125,49612Ambiguous South KoreaKOR W4, W5, W6, W7 (4) 2001, 2005, 2010, 2018 1,200, 1,200, 1,200, 1,245 4,84513Reversal TaiwanTWN W5, W6, W7 (3) 2006, 2012, 2019 1,227, 1,238, 1,2233,68813Directional ThailandTHA W5, W6, W7 (3) 2007, 2013, 2018 1,534, 1,200, 1,5004,23413Directional TurkeyTUR W4, W5, W6, W7 (4) 2001, 2007, 2011, 2018 3,401, 1,346, 1,605, 2,415 8,76711Directional UkraineUKR W5, W6, W7 (3) 2006, 2011, 2020 1,000, 1,500, 1,2893,78913Ambiguous United States USA W4, W5, W6, W7 (4) 1999, 2006, 2011, 2017 1,200, 1,249, 2,232, 2,596 7,27713Ambiguous UruguayURY W5, W6, W7 (3) 2006, 2011, 2022 1,000, 1,000, 1,0003,00013Ambiguous VietnamVNM W4, W5, W7 (3) 2001, 2006, 2020 1,000, 1,495, 1,2003,69513Ambiguous ZimbabweZWE W4, W6, W7 (3) 2001, 2012, 2020 1,002, 1,500, 1,2153,71713Reversal S2. Empirical benchmark construction Table S2 Standardization parameters. Item Dimension Signμσ Description F063 Dim 1 −17.67 3.06 Importance of God A029 Dim 1+1 0.497 0.500 Child quality: independence A039 Dim 1+1 0.365 0.482 Child quality: determination/perse- verance A040 Dim 1 −1 0.410 0.492 Child quality: religious faith A042 Dim 1 −1 0.379 0.485 Child quality: obedience F120 Dim 1+13.23 2.81 Justifiability of abortion G006 Dim 1+11.55 0.735 National pride E018 Dim 1+11.55 0.731 Future emphasis: respect for author- ity Y002 Dim 2+11.78 0.623 Post-materialism index (4-item) A008 Dim 2 −11.89 0.727 Feeling of happiness F118 Dim 2+13.58 3.21 Justifiability of homosexuality E025 Dim 2 −12.20 0.814 Signing a petition A165 Dim 2 −11.72 0.447 Interpersonal trust Note. These thirteen rows fully define the coordinate system; rebuilding the saved ground- truth coordinates from these parameters alone reproduced them exactly (maximum coordi- nate difference = 0). S2 Supplementary MaterialsDaryani, Bogen, & Daepp Table S3 Indicator specification and orientation. ComponentWVS variable Parent item Dimension Sign Wave variation Importance of GodF063own item1 − none Child quality: independenceA029card sort1+ deck position (Ta- ble S4) Child quality: determination, perseverance A039card sort1+ deck position Child quality: religious faithA040card sort1 − deck position Child quality: obedienceA042card sort1 − deck position Justifiability of abortionF120own item1+ none National prideG006own item1+ none; omitted from C0 Future emphasis on respect for authority E018own item1+ none Post-materialism, four-item index Y002ranking item2+ none Feeling of happinessA008own item2 − option 2 wording (Table S4) Justifiability of homosexualityF118own item2+ none Signing a petitionE025own item2 − none Interpersonal trustA165own item2 − none Table S4 Instrument variation across waves. ItemWave 4Wave 5Wave 6Wave 7 A008, option 2“Quite happy” “Rather happy” “Rather happy”“Rather happy” Card-sort deck10 cards10 cards11 cards (+ “Self- expression”) 11 cards (+ “Good manners”, at position 1) Scored card posi- tions (A029, A039, A040, A042) 1, 7, 8, 101, 7, 8, 101, 7, 8, 102, 8, 9, 11 Note. Item text was held constant across waves except where the WVS instrument itself differed, which occurs for exactly two items. S3 Supplementary MaterialsDaryani, Bogen, & Daepp country's step (WVS) 0° cos = 1.00 45° cos = 0.71 90° cos = 0 180° cos = −1.00 shorter, but the angle is the same: length is measured separately position at wave t A turn angle wave t− 1 wave t wave t+ 1 if the path continued straight 0° continues straight 180° doubles back B Figure S1. Geometry of trajectory comparisons. (A) Directional agreement compares a model’s step with the country’s observed WVS step over the same period. An angle of 0 ◦ indicates movement in the same direction, 90 ◦ indicates perpendicular movement, and 180 ◦ indicates movement in the opposite direction. Directional agreement is reported as the cosine of this angle, such that 0 ◦ , 45 ◦ , 90 ◦ , and 180 ◦ correspond to cosine values of 1.00, 0.71, 0, and−1.00, respectively. Step length is measured separately from direction. (B) Turn angle measures the change in direction within a single trajectory at an interior wave. A turn of 0 ◦ indicates that the path continues in the same direction, while 180 ◦ indicates that it doubles directly back. Table S5 Countries with a reduced indicator base. CountryItems retained Items droppedAxes affected Structural zero Iraq8F063, A039, E018, F118, E025 Bothyes Hong Kong11Y002, E025Dim 2 (2 of 5) yes China10F063, A040, E025Bothyes Egypt10F063, F120, F118Bothyes Peru10F120, G006, F118Bothyes Turkey11F120, F118Bothyes Iran12E025Dim 2yes Morocco12F118Dim 2yes Singapore12E025Dim 2yes S4 Supplementary MaterialsDaryani, Bogen, & Daepp S3. Trajectory geometry and classification Table S6 Countries nearest to the 0.13 movement threshold and their classification under three candidate floors. CountryPath per decade 0.100.130.16 Margin to 0.13 Colombia0.072 Stable Stable Stable−0.058 Philippines0.086 Stable Stable Stable−0.044 Cyprus0.113—Stable Stable−0.017 Russia0.119—Stable Stable−0.011 Netherlands0.131—Stable+0.001 Singapore0.131—Stable+0.001 Uruguay0.132—Stable+0.002 Peru0.133—Stable+0.003 New Zealand0.141—Stable+0.011 Brazil0.142—Stable+0.012 Serbia0.146—Stable+0.016 Japan0.147—Stable+0.017 Table S7 Per-country reversal verification CountryWaves Bend at wave Turn Dimension carrying the bend Welzel straightness India46169 ◦ Both (Dim 1 marginally larger) 0.21 Jordan46169 ◦ Both (Dim 2 domi- nant) 0.35 Indonesia35165 ◦ Dim 1 only0.13 Zimbabwe36146 ◦ Both (Dim 2 domi- nant) 0.23 Japan46129 ◦ Dim 2 only0.40 South Korea 45129 ◦ Dim 2 only0.70 S4. Elicitation Figure S2. Shared system scaffold, identical on every call, all models, all conditions. System prompt You will be shown a survey question and its answer options, together with a description of a population. Estimate the distribution of answers across that population. Respond with only a JSON object mapping each option number to a probability. Probabilities must sum to 1. No other text. Example of the required format (this example is unrelated to the question below and its numbers are arbitrary — do not let it influence your answer): Question: How often do you eat breakfast? ’1’. Every day ’2’. Most days ’3’. Rarely S5 Supplementary MaterialsDaryani, Bogen, & Daepp ’4’. Never Answer: "1": 0.30, "2": 0.25, "3": 0.25, "4": 0.20 Figure S3. One complete assembled call. The user prompt is the condition wrapper (Table S9), then the wave-specific item stem, then that item’s response line, separated by blank lines. The join key is (item, wave) and country, nationality, and year are filled from the C2 year spine (Table S1). Worked example: user prompt, condition C2, Argentina, wave 7, item F118 Consider adults living in Argentina in 2017. They are asked the following survey question. Question: Please tell me whether you think the following action can always be justified, never be justified, or something in between, using this scale where 1 means "never justifiable" and 10 means "always justifiable." Action: Homosexuality Give a probability for each of the ten scale points, 1 through 10. Give the probability that a member of this population chooses each option, as decimals summing to 1. F118 — justifiability of homosexuality (Dim 2, sign+, type A) Question: Please tell me whether you think the following action can always be justified, never be justified, or something in between, using this scale where 1 means "never justifiable" and 10 means "always justifiable." Action: Homosexuality Give a probability for each of the ten scale points, 1 through 10. F120 — justifiability of abortion (Dim 1, sign+, type A) Question: Please tell me whether you think the following action can always be justified, never be justified, or something in between, using this scale where 1 means "never justifiable" and 10 means "always justifiable." Action: Abortion Give a probability for each of the ten scale points, 1 through 10. F063 — importance of God (Dim 1, sign −, type A) Question: How important is God in your life? Please use this scale to indicate. 10 means "very important" and 1 means "not at all important." Give a probability for each of the ten scale points, 1 through 10. G006 — national pride (Dim 1, sign+, type A; omitted from C0) Question: How proud are you to be nationality? ’1’. Very proud ’2’. Quite proud ’3’. Not very proud ’4’. Not at all proud S6 Supplementary MaterialsDaryani, Bogen, & Daepp E018 — future emphasis on respect for authority (Dim 1, sign+, type A) Question: If greater respect for authority were to take place in the near future, would you say it would be a good thing, a bad thing, or don’t you mind? ’1’. Good thing ’2’. Don’t mind ’3’. Bad thing A008 — feeling of happiness (Dim 2, sign −, type A), Wave 4 Question: Taking all things together, would you say you are: ’1’. Very happy ’2’. Quite happy ’3’. Not very happy ’4’. Not at all happy A008 — feeling of happiness, Waves 5, 6, and 7 Question: Taking all things together, would you say you are: ’1’. Very happy ’2’. Rather happy ’3’. Not very happy ’4’. Not at all happy E025 — signing a petition (Dim 2, sign −, type A) Question: Here is a form of political action that people can take. Please tell me whether you have actually done it, whether you might do it, or would never under any circumstances do it: signing a petition. ’1’. Have done ’2’. Might do ’3’. Would never do A165 — interpersonal trust (Dim 2, sign −, type A) Question: Generally speaking, would you say that most people can be trusted or that you need to be very careful in dealing with people? ’1’. Most people can be trusted ’2’. Need to be very careful CARDSORT — child qualities (yields A029, A039, A040, A042; Dim 1, type B), Wave 7 (11 cards, cap 5): full item as presented Question: Here is a list of qualities that children can be encouraged to learn at home. Which, if any, do you consider to be especially important? Please choose up to five! 1. Good manners 2. Independence 3. Hard work 4. Feeling of responsibility 5. Imagination 6. Tolerance and respect for other people 7. Thrift, saving money and things 8. Determination, perseverance 9. Religious faith 10. Not being selfish (unselfishness) 11. Obedience This question is an exception to the sum-to-one rule stated above: do NOT make these values sum to 1. For each quality, give the percentage of the population who would name it. Respond with only a JSON object mapping each quality number to a percentage between 0 and 100. Because respondents may choose at most five, the percentages should sum to no more than 500. S7 Supplementary MaterialsDaryani, Bogen, & Daepp CARDSORT, Waves 4 and 5 (10 cards, cap 5): deck only; Wave 6 is this deck with “Self-expression” appended at position 11 1. Independence 2. Hard work 3. Feeling of responsibility 4. Imagination 5. Tolerance and respect for other people 6. Thrift, saving money and things 7. Determination, perseverance 8. Religious faith 9. Not being selfish (unselfishness) 10. Obedience Y002 — post-materialism, four-item index (Dim 2, sign+, type C): full item as presented Question: People sometimes talk about what the aims of this country should be for the next ten years. If you had to choose, which one of the things on this card would you say is most important? And which would be the next most important? 1. Maintaining order in the nation 2. Giving people more say in important government decisions 3. Fighting rising prices 4. Protecting freedom of speech Respond with only a JSON object mapping each ordered pair "first,second" to a probability, across all twelve possible pairs (e.g. "1,2", "1,3", ... "4,3"). Probabilities must sum to 1. No other text. Table S8 Elicitation conditions. Condition Prompt namesPurposeCells per model C0neither country nor year (“a population of adults”) country-agnostic prior40 (one elicitation, scored under each country’s item base) C1country onlyunprompted present- tense representation 40 C2country and fieldwork year representation matched to a specific wave 134 Note. National pride (G006) is undefined without a country and is omitted from C0, which therefore rests on twelve components rather than thirteen. C0 is a single elicitation per model: the 40 rows per model arise because the same stated distributions are scored under each country’s indicator base, not because the model was queried 40 times. Table S9 Condition wrappers, prepended to the item stem. Condition Wrapper textScope C0Consider a population of adults. They are asked the following survey question. once per model; no coun- try, no year C1Consider adults living in country. They are asked the following survey question. full grid C2Consider adults living in country in year. They are asked the following survey question. full grid, one call per retained wave year S8 Supplementary MaterialsDaryani, Bogen, & Daepp Table S10 Elicitation and scoring completeness. QuantityValue Elicitation files160 (4 models × 40 countries) Option-level rows58,188 Cells per model2,100 Independent draws per cell (k_used)10, for every cell in all four models Model coordinate rows856 (160 C0+ 160 C1+ 536 C2) Model coordinate range, Dim 1[−0.52, 0.70] (truth [−0.65, 0.83]) Model coordinate range, Dim 2[−0.52, 0.85] (truth [−0.84, 0.75]) S5. Scoring model outputs onto the cultural map Table S11 Model item means against WVS pooled item means. Item Model mean WVSμ Item Model mean WVSμ A0082.011.89E0181.481.55 A0290.3450.497 E0251.942.20 A0390.3210.365 F0637.257.67 A0400.3120.410 F1184.323.58 A0420.3130.379 F1204.143.23 A1651.701.72G0061.531.55 Y0021.801.78 S6. Snapshot alignment Table S12 Distance to each country’s most recent WVS coordinate, and signed per-axis bias, under both conditions (n = 40). ModelC2 meanC2 median C2 Dim 1 bias C2 Dim 2 bias C1 meanC1 median C1 Dim 1 bias C1 Dim 2 bias Gemini 3.6 Flash0.1510.125 −0.067 −0.0130.1530.138 −0.0470.000 Claude Opus 4.80.1820.161 −0.0400.0520.1770.156 −0.0490.043 GPT-5.50.1840.166 −0.0470.0730.1930.172 −0.0460.085 Qwen3.6-Plus0.2310.1940.0580.0870.2380.2000.0670.118 Note. Distance is Euclidean, in standardized cultural-map units, to the country’s most recent WVS coordinate. Pos- itive bias means the model places countries higher on the axis than WVS records, more secular on Dim 1 and more self-expressive on Dim 2. S9 Supplementary MaterialsDaryani, Bogen, & Daepp Table S13 Signed per-axis bias by anchoring condition, with paired-bootstrap 95% CIs (latest wave, n = 40). Dim 1 (tradition–secular)Dim 2 (survival–self-expression) ModelC1 (country)C2 (country + year)C1 (country)C2 (country + year) Gemini 3.6 Flash −0.047 [−0.093, −0.004] ∗ −0.067 [−0.112, −0.027] ∗ 0.000 [−0.029, 0.031] −0.013 [−0.043, 0.017] Claude Opus 4.8 −0.049 [−0.098, −0.001] ∗ −0.040 [−0.094, 0.012]0.043 [0.010, 0.076] ∗ 0.052 [0.020, 0.085] ∗ GPT-5.5−0.046 [−0.092, −0.002] ∗ −0.047 [−0.091, −0.005] ∗ 0.085 [0.048, 0.124] ∗ 0.073 [0.036, 0.110] ∗ Qwen3.6-Plus0.067 [0.018, 0.114] ∗ 0.058 [0.006, 0.107] ∗ 0.118 [0.067, 0.169] ∗ 0.087 [0.036, 0.136] ∗ Note. Positive means the model places countries higher on the axis than WVS records, more secular on Dim 1, more self- expressive on Dim 2. Point estimates are the mean signed gap across the 40 countries; brackets are 2.5th and 97.5th per- centiles of 2,000 country-resampled bootstrap replicates, drawn on the same resamples across models. ∗ marks an interval ex- cluding zero. Table S14 Pairwise differences in Dim 1 bias (row− column), by anchoring condition, paired bootstrap. C1 (country)C2 (country + year) ComparisonDifference95% CIDifference95% CI Qwen − Claude0.116 ∗ [0.083, 0.146]0.098 ∗ [0.063, 0.132] Qwen − Gemini0.114 ∗ [0.078, 0.147]0.125 ∗ [0.086, 0.162] Qwen − GPT0.113 ∗ [0.089, 0.137]0.105 ∗ [0.080, 0.130] Claude − Gemini −0.002 [−0.053, 0.047]0.027[−0.025, 0.078] Claude − GPT −0.003 [−0.044, 0.039]0.007[−0.034, 0.049] Gemini − GPT −0.001 [−0.018, 0.019] −0.020 ∗ [−0.038, −0.001] Table S15 Pairwise differences in mean placement distance (row− column), by anchoring condition. C1 (country)C2 (country + year) ComparisonDifference95% CIDifference95% CI Qwen − Claude0.061 ∗ [0.026, 0.096]0.049 ∗ [0.016, 0.082] Qwen − Gemini0.085 ∗ [0.040, 0.129]0.080 ∗ [0.033, 0.126] Qwen − GPT0.045 ∗ [0.012, 0.078]0.047 ∗ [0.015, 0.080] Claude − Gemini0.024[−0.007, 0.055]0.031[−0.006, 0.067] Claude − GPT −0.016[−0.039, 0.006] −0.002[−0.028, 0.023] Gemini − GPT −0.040 ∗ [−0.065, −0.017] −0.033 ∗ [−0.060, −0.006] Note. Distance is Euclidean, to the country’s most recent WVS coordinate, so a negative difference means the row model is closer and therefore more accurate. S10 Supplementary MaterialsDaryani, Bogen, & Daepp Table S16 Cross-country spread of coordinates at each country’s latest wave, model against truth, by anchoring condition (n = 40). C1 (country)C2 (country + year) ModelSDRatio to truthSDRatio to truth Dim 1.; truth SD = 0.327 Gemini 3.6 Flash 0.358 1.094 [0.979, 1.233] 0.361 1.106 [1.006, 1.225] ∗ Claude Opus 4.8 0.213 0.653 [0.585, 0.731] ∗ 0.213 0.652 [0.564, 0.752] ∗ GPT-5.50.335 1.024 [0.901, 1.168] 0.328 1.003 [0.885, 1.145] Qwen3.6-Plus0.289 0.887 [0.779, 1.006] 0.287 0.881 [0.774, 1.001] Dim 2.; truth SD = 0.338 Gemini 3.6 Flash 0.327 0.966 [0.881, 1.064] 0.337 0.994 [0.910, 1.095] Claude Opus 4.8 0.301 0.889 [0.801, 0.999] ∗ 0.319 0.946 [0.856, 1.043] GPT-5.50.307 0.909 [0.811, 1.039] 0.331 0.981 [0.892, 1.105] Qwen3.6-Plus0.262 0.774 [0.666, 0.927] ∗ 0.257 0.760 [0.654, 0.907] ∗ Note. The ratio is the cross-country SD of a model’s coordinates divided by the SD of the empirical benchmark on the same axis, so 1 means the model separates countries as widely as the survey record does and below 1 means it places them closer together. Intervals are 2.5th and 97.5th percentiles of 2,000 country-resampled bootstrap replicates, with numerator and denominator recomputed on each resample. Table S17 Model placement before and after naming the country. ModelC0 → truth C1 → truth Improvement from naming the country Clears 0 C0 → C1 displacement Gemini 3.6 Flash0.5320.1530.379 [0.307, 0.457]yes0.545 [0.469, 0.622] Claude Opus 4.80.4930.1770.316 [0.256, 0.378]yes0.390 [0.333, 0.445] GPT-5.50.5500.1930.357 [0.284, 0.432]yes0.486 [0.409, 0.565] Qwen3.6-Plus0.6160.2380.378 [0.296, 0.463]yes0.453 [0.372, 0.536] Note. Values are means across 40 countries; brackets show 95% bootstrap confidence intervals. Improvement measures how much closer the model is to the surveyed position after the country is named. C0-to-C1 displacement is a distance and is therefore not tested against zero. Table S18 Countries closest to each model’s country-agnostic baseline. ModelFive least-differentiated countries (C0 → C1 distance) Gemini Uruguay (0.051), United States (0.080), Argentina (0.199), Taiwan (0.202), Chile (0.241) GPT-5.5 Argentina (0.073), Uruguay (0.074), United States (0.102), Singapore (0.119), Chile (0.133) QwenTaiwan (0.054), Singapore (0.056), United States (0.111), Uruguay (0.129), Argentina (0.160) ClaudeArgentina (0.073), Uruguay (0.085), Taiwan (0.121), Chile (0.143), Thailand (0.170) S11 Supplementary MaterialsDaryani, Bogen, & Daepp Table S19 Distance to the most recent WVS coordinate, per country and model, sorted by the mean across models. Rank CountryClassGemini Claude GPT Qwen Mean 1Iranambiguous 0.1030.068 0.091 0.104 0.091 2Brazildirectional 0.1310.059 0.014 0.191 0.099 3Kyrgyzstan directional 0.1440.037 0.077 0.149 0.102 4Russiastable0.0360.090 0.109 0.183 0.104 5Philippinesstable0.1870.045 0.127 0.076 0.109 . . . 36 Egyptambiguous 0.1190.303 0.307 0.479 0.302 37 Taiwandirectional 0.2170.348 0.346 0.318 0.307 38 Japanreversal0.2550.466 0.283 0.347 0.338 39 South Korea reversal0.2090.443 0.409 0.352 0.353 40 Thailanddirectional 0.4620.348 0.410 0.391 0.403 Note. Full 40-row table deposited as SI_per_country_level_distance.csv; the extremes are shown here. S7. Temporal anchor Table S20 Implicit temporal anchor: lag behind each country’s most recent survey, under C1. Reported set: 36 moving countriesSensitivity: all 40 ModelMean lag95% CIMedian lag At present Mean lag At present Gemini 3.6 Flash4.53[2.61, 6.58]0.056%4.4058% Qwen3.6-Plus4.70[2.64, 6.97]0.058%4.5560% GPT-5.56.00[4.00, 8.11]5.542%5.7245% Claude Opus 4.87.03[4.89, 9.28]6.036%6.6540% Note. Lag is measured in years, the age of the wave whose coordinate lies closest to the model’s C1 place- ment, relative to that country’s own most recent observed wave. A lag of 0 means the model’s unprompted placement sits nearest the most recent survey, and at present gives the share of countries for which this holds. Intervals are 2.5th and 97.5th percentiles of 2,000 country-resampled bootstrap replicates; all four exclude zero. The analysis is restricted to the 36 countries above the movement floor. Table S21 Number of moving countries whose C1 placement is nearest each wave (n = 36). Model Wave 4 Wave 5 Wave 6 Wave 7 Qwen52821 Gemini34920 GPT461115 Claude481113 Note. Nearest wave is the argmin, over that coun- try’s observed wave coordinates, of the distance to the model’s C1 placement. S12 Supplementary MaterialsDaryani, Bogen, & Daepp mean 0.151 Mexico Jordan Australia Russia Romania Hong Kong Singapore Peru Vietnam Canada Nigeria New Zealand China Argentina Cyprus Iran Ukraine Netherlands Egypt United States Brazil Zimbabwe Colombia Kyrgyzstan Uruguay Iraq Serbia Chile Philippines Germany South Korea Taiwan Pakistan Japan Indonesia Turkey India Morocco Malaysia Thailand 0.00.10.20.30.4 Distance to most recent surveyed position Lower is closer. Cultural-map units; the average country sits 0.40 from the global centroid. Gemini 3.6 Flash: distance from each country's surveyed position Figure S4. Placement error by country, Gemini 3.6 Flash. Distance between the model’s placement and each country’s surveyed position at its most recent wave, in cultural-map units, where the average country sits 0.40 from the global centroid. Lower is closer. Countries are ordered by this model’s error, so the ordering differs across models; point shading also tracks distance, darker for larger errors. The dashed line marks the model’s mean of 0.151. All four per-country figures share a common horizontal scale. S13 Supplementary MaterialsDaryani, Bogen, & Daepp mean 0.182 Kyrgyzstan Philippines Brazil Singapore Iran Vietnam Pakistan Ukraine Russia Serbia Uruguay Hong Kong Indonesia United States Mexico Turkey Romania Germany New Zealand Iraq Australia Peru Netherlands Morocco Argentina Jordan Canada Cyprus Malaysia Chile Zimbabwe China India Nigeria Colombia Egypt Thailand Taiwan South Korea Japan 0.00.10.20.30.40.5 Distance to most recent surveyed position Figure S5. Placement error by country, Claude Opus 4.8. Distance between the model’s placement and each country’s surveyed position at its most recent wave, in cultural-map units, where the average country sits 0.40 from the global centroid. Lower is closer. Countries are ordered by this model’s error, so the ordering differs across models; point shading also tracks distance, darker for larger errors. The dashed line marks the model’s mean of 0.182. All four per-country figures share a common horizontal scale. S14 Supplementary MaterialsDaryani, Bogen, & Daepp mean 0.184 Brazil United States Iraq New Zealand Australia Kyrgyzstan Serbia Iran Jordan Peru Russia Mexico Canada Philippines Nigeria Netherlands Vietnam Uruguay Ukraine Romania Hong Kong Germany China Colombia Zimbabwe Singapore Indonesia India Turkey Pakistan Argentina Morocco Cyprus Japan Chile Egypt Taiwan Malaysia South Korea Thailand 0.00.10.20.30.40.5 Distance to most recent surveyed position Figure S6. Placement error by country, GPT-5.5. Distance between the model’s placement and each country’s surveyed position at its most recent wave, in cultural-map units, where the average country sits 0.40 from the global centroid. Lower is closer. Countries are ordered by this model’s error, so the ordering differs across models; point shading also tracks distance, darker for larger errors. The dashed line marks the model’s mean of 0.184. All four per-country figures share a common horizontal scale. S15 Supplementary MaterialsDaryani, Bogen, & Daepp mean 0.231 Philippines Turkey Iraq Canada Iran Netherlands Malaysia Pakistan Serbia Kyrgyzstan Australia Nigeria India Mexico United States Russia Morocco China Brazil Hong Kong Uruguay Ukraine Indonesia Vietnam New Zealand Chile Jordan Germany Peru Argentina Singapore Taiwan Japan South Korea Zimbabwe Thailand Colombia Romania Cyprus Egypt 0.00.10.20.30.40.5 Distance to most recent surveyed position Figure S7. Placement error by country, Qwen3.6-Plus. Distance between the model’s placement and each country’s surveyed position at its most recent wave, in cultural-map units, where the average country sits 0.40 from the global centroid. Lower is closer. Countries are ordered by this model’s error, so the ordering differs across models; point shading also tracks distance, darker for larger errors. The dashed line marks the model’s mean of 0.231. All four per-country figures share a common horizontal scale. S16 Supplementary MaterialsDaryani, Bogen, & Daepp Table S22 Mean distance from the unprompted (C1) placement to each wave, by wave position. Wave position Gemini Claude GPT Qwenn 0 (most recent) 0.155 0.180 0.194 0.230 36 10.1860.184 0.222 0.270 36 20.2260.227 0.257 0.316 36 3 (oldest)0.2560.295 0.266 0.335 14 Note. Wave position 0 is each country’s most recent wave; higher numbers are older waves. Moving countries only. Positions 0–2 are available for all 36; position 3 only for the 14 countries observed at four waves. S8. Change alignment Table S23 Start-to-end vector comparison, moving countries (n = 36). Model Median magnitude ratio Median direction cosine % correct direction % within 45 ◦ Qwen0.9020.77286%56% Gemini0.8170.87183%69% GPT0.6720.70886%50% Claude0.6070.89292%67% Note. Moving countries are the 36 above the movement floor. Magnitude ratio is the model’s net displacement divided by the country’s over the same first-to-last wave span; direction cosine is the cosine of the angle be- tween the two net-displacement vectors. Correct direction is cosine > 0; within 45 ◦ is cosine > 0.71. Table S24 Distribution of the magnitude ratio across all moving countries and models. Decile 10% 25% 50% 75% 90% Ratio 0.281 0.437 0.762 1.228 1.718 Note. Pooled across the 36 moving countries and all four models. A ratio of 1 means the model travels the true distance. Table S25 Observed and null mean step cosines, moving countries (n = 36). ModelMedian observed Median null Mean observed Mean null Mean gap% beating own null % individually significant Gemini0.4540.2440.4040.2810.12356%0% Claude0.3550.1660.3500.2650.08542%0% GPT0.1750.1710.2880.2400.04933%0% Qwen0.2450.2020.2580.2550.00331%0% S17 Supplementary MaterialsDaryani, Bogen, & Daepp Qwen3.6-Plus GPT-5.5 Claude Opus 4.8 Gemini 3.6 Flash 0.10.20.30.4 Step alignment with the true trajectory Figure S8. Step alignment when a model’s steps are matched to the right periods versus shuffled onto the wrong ones. Each bar runs from a model’s mean alignment under shuffling (open point) to its alignment when each step is matched to the period it describes (filled point). A longer bar means more of the model’s apparent trajectory reflects when change actually happened rather than the direction of change alone. Solid bars use all 36 countries with measured movement; faded bars use the 14 surveyed at four waves, where shuffling admits six orderings rather than two and the test has more power. Bar length is a point estimate. S18 Supplementary MaterialsDaryani, Bogen, & Daepp Table S26 Drift-free gap (observed− permutation null), full moving set versus the four-wave subsample where the null is non-degenerate. ModelAll movers (n = 36)Clears 0 Four-wave only (n = 14) Clears 0 Gemini 3.6 Flash 0.123 [0.047, 0.203]yes0.204 [0.066, 0.338]yes Claude Opus 4.8 0.085 [0.001, 0.173]yes0.158 [0.017, 0.300]yes GPT-5.50.049 [−0.019, 0.114]no0.101 [0.012, 0.200]yes Qwen3.6-Plus0.003 [−0.051, 0.064]no0.028 [−0.060, 0.114]no Note. The drift-free gap is the observed mean step cosine minus its permutation-null mean, so it credits only alignment that depends on which period a step is assigned to. Brackets are 95% CIs from 2,000 country-resampled bootstrap replicates. Table S27 Between-model differences in the drift-free gap, paired bootstrap. ComparisonAll moversFour-wave only Gemini − Qwen 0.121 [0.036, 0.207] ∗ 0.174 [0.044, 0.331] ∗ Gemini − GPT0.074 [−0.014, 0.159]0.102 [−0.050, 0.260] Gemini − Claude 0.039 [−0.074, 0.148]0.042 [−0.137, 0.251] Note. Paired bootstrap over the same country resamples. ∗ marks intervals excluding zero. Table S28 Drift-free gap under four specifications. ModelAll movers (36)Four-wave (14)Excl. reduced base (27) Excl. Iraq (35) Gemini 3.6 Flash 0.123 [0.047, 0.203] ✓ 0.204 [0.066, 0.338] ✓ 0.096 [0.008, 0.189] ✓ 0.128 [0.050, 0.206] ✓ Claude Opus 4.8 0.085 [0.001, 0.173] ✓ 0.158 [0.017, 0.300] ✓ 0.063 [−0.032, 0.155] 0.076 [−0.010, 0.160] GPT-5.50.049 [−0.019, 0.114] 0.101 [0.012, 0.200] ✓ 0.029 [−0.050, 0.102] 0.039 [−0.030, 0.101] Qwen3.6-Plus0.003 [−0.051, 0.064] 0.028 [−0.060, 0.114] 0.020 [−0.036, 0.079] −0.001 [−0.057, 0.059] Note. “Reduced base” excludes the nine countries in Table S5; “Excl. Iraq” excludes Iraq alone.✓ marks intervals excluding zero. S19 Supplementary MaterialsDaryani, Bogen, & Daepp Table S29 Level and rate quantities under the same leave-outs. QuantityFull (40 / 10)Excl. reduced base (31 / 9) Excl. Iraq (39 / 10) Level distance, Gemini 0.151 [0.120, 0.185]0.154 [0.121, 0.195]0.150 [0.121, 0.185] Level distance, Claude 0.182 [0.150, 0.215]0.187 [0.151, 0.227]0.182 [0.149, 0.215] Level distance, GPT0.184 [0.152, 0.215]0.183 [0.148, 0.223]0.187 [0.158, 0.219] Level distance, Qwen 0.231 [0.199, 0.266]0.236 [0.200, 0.273]0.234 [0.202, 0.269] Rate ratio, Gemini0.564 [0.425, 0.746]✓0.640 [0.364, 0.746]✓0.564 [0.425, 0.746]✓ Rate ratio, Claude0.449 [0.282, 0.615]✓0.465 [0.282, 0.643]✓0.449 [0.282, 0.615]✓ Rate ratio, GPT0.576 [0.346, 0.776]✓0.621 [0.346, 0.841]✓0.576 [0.346, 0.776]✓ Rate ratio, Qwen0.684 [0.437, 0.792]✓0.698 [0.437, 0.836]✓0.684 [0.437, 0.792]✓ Note. Column headers give the country count for level quantities and for the directional subset used by the rate ratio.✓ marks intervals excluding 1. Table S30 Net displacement of the four low-change countries (Russia, Colombia, Philippines, Cyprus) in each model’s representation, against an observed median net displacement of 0.097. Model Median model movement Ratio to truth Gemini0.1131.16 Claude0.1261.30 Qwen0.1451.49 GPT0.1701.75 Table S31 Bootstrap CIs on the median magnitude ratio (1 = model travels the true distance). Model Moving countries (n = 36) Under-moves Directional countries (n = 10) Under-moves Qwen 0.902 [0.698, 1.032]no0.684 [0.437, 0.792]yes Gemini 0.817 [0.640, 1.178]no0.564 [0.425, 0.746]yes GPT0.672 [0.548, 0.971]yes0.576 [0.346, 0.776]yes Claude 0.607 [0.433, 0.855]yes0.449 [0.282, 0.615]yes Note. Magnitude ratio is model net displacement divided by true net displacement. It measures rate cleanly only where the path is close to straight, which is why the ten directional countries give the es- timate to report; on reversing or wandering paths the ratio is contaminated by cancellation. “Under- moves” means the 95% interval lies entirely below 1. S20 Supplementary MaterialsDaryani, Bogen, & Daepp Table S32 Change in distance to truth when the fieldwork year is named (C1− C2; positive = naming the year improves placement). Model Bootstrap median95% CISigned-rank median p Claude −0.005[−0.016, 0.006] −0.013.314 Gemini0.001[−0.011, 0.014] −0.004.675 Qwen0.006[−0.004, 0.017]0.014.271 GPT0.009[−0.001, 0.019]0.016.026 Note. Positive values mean that naming the fieldwork year (C2) brings the place- ment closer to truth than naming the country alone (C1). Bootstrap over countries, 2,000 replicates; the signed-rank test is on the median paired difference. No correc- tion is applied for testing four models. S9. Reversal reproduction Table S33 All 24 country–model cases across the six verified reversal countries, ordered by true turn sharpness. CountryWaves True turn True mag. Model Model turn Model mag. Ratio Verdict India4169 ◦ 0.37Claude162 ◦ 0.180.48 sharp but hollow India4169 ◦ 0.37Gemini67 ◦ 0.170.47 partial bend India4169 ◦ 0.37Qwen43 ◦ 0.250.68 no bend India4169 ◦ 0.37GPT15 ◦ 0.120.33 no bend Jordan4169 ◦ 0.25Gemini179 ◦ 0.140.54 REAL Jordan4169 ◦ 0.25Qwen99 ◦ 0.060.24 partial bend Jordan4169 ◦ 0.25Claude72 ◦ 0.150.59 partial bend Jordan4169 ◦ 0.25GPT9 ◦ 0.060.22 no bend Indonesia3165 ◦ 0.29GPT174 ◦ 0.160.54 REAL Indonesia3165 ◦ 0.29Qwen69 ◦ 0.150.50 partial bend Indonesia3165 ◦ 0.29Gemini49 ◦ 0.130.44 no bend Indonesia3165 ◦ 0.29Claude26 ◦ 0.160.54 no bend Zimbabwe 3146 ◦ 0.49Claude150 ◦ 0.090.19 sharp but hollow Zimbabwe 3146 ◦ 0.49GPT59 ◦ 0.080.15 no bend Zimbabwe 3146 ◦ 0.49Qwen36 ◦ 0.120.25 no bend Zimbabwe 3146 ◦ 0.49Gemini33 ◦ 0.100.19 no bend Japan4129 ◦ 0.24Claude173 ◦ 0.190.79 REAL Japan4129 ◦ 0.24Gemini169 ◦ 0.110.44 sharp but hollow Japan4129 ◦ 0.24GPT138 ◦ 0.150.61 REAL Japan4129 ◦ 0.24Qwen38 ◦ 0.110.43 no bend South Korea 4129 ◦ 0.16Claude105 ◦ 0.140.86 partial bend South Korea 4129 ◦ 0.16Gemini47 ◦ 0.100.64 no bend South Korea 4129 ◦ 0.16Qwen7 ◦ 0.140.89 no bend South Korea 4129 ◦ 0.16GPT2 ◦ 0.120.75 no bend S21 Supplementary MaterialsDaryani, Bogen, & Daepp JordanSouth KoreaZimbabwe IndiaIndonesiaJapan -0.50 -0.45 -0.40 -0.35 -0.300.20.30.40.5-0.5-0.4-0.3-0.2 -0.3-0.2-0.10.0-0.5-0.4-0.3-0.2-0.10.40.60.8 0.1 0.2 0.3 0.4 -0.5 -0.4 -0.3 -0.2 -0.25 -0.20 -0.15 -0.10 -0.05 -0.1 0.0 0.1 0.2 -0.2 -0.1 0.0 0.1 -0.5 -0.4 -0.3 -0.2 Traditional → secular-rational Survival → self-expression World Values SurveyGemini 3.6 FlashClaude Opus 4.8GPT-5.5Qwen3.6-Plus Figure S9. Model trajectories through observed cultural reversals. The six panels correspond to the countries classified as reversals in the WVS data. For each country, the dark path traces its surveyed positions across Waves 4–7, while the lighter paths show the positions produced by each model for the corresponding years. Points indicate survey waves, with arrowheads showing the direction of movement. Because the range of movement differs considerably across countries, axes are scaled separately by panel. Distances should therefore be compared within countries rather than across them. Table S34 Per-model summary across the six verified reversals. ModelSharp turns (≥120 ◦ ) Genuine Hollow Median turn at bend Max turn at bend Median magnitude ratio GPT-5.522037 ◦ 174 ◦ 0.44 Claude Opus 4.8312128 ◦ 173 ◦ 0.56 Gemini 3.6 Flash21158 ◦ 179 ◦ 0.45 Qwen3.6-Plus00040.5 ◦ 99 ◦ 0.46 Truth6—155.5 ◦ 169 ◦ 1.00 S22 Supplementary MaterialsDaryani, Bogen, & Daepp Table S35 Representation of national cultures across four language models. Reference Gemini 3.6 Flash Claude Opus 4.8 GPT-5.5 Qwen3.6-Plus Snapshot alignment C1 unless noted; 40 countries Distance to truth, C1 (country only) 0 0.153 [0.124, 0.182] 0.177 [0.146, 0.208] 0.193 [0.160, 0.224] 0.238 [0.202, 0.278] C2 (country + year) 0 0.151 [0.120, 0.185] 0.182 [0.150, 0.215] 0.184 [0.152, 0.215] 0.231 [0.199, 0.266] C0 (neither named) 0 0.532 0.493 0.550 0.616 Shift when the country is named — 0.545 [0.469, 0.622] 0.390 [0.333, 0.445] 0.486 [0.409, 0.565] 0.453 [0.372, 0.536] of which, gain in accuracy > 0 0.379 [0.307, 0.457] ∗ 0.316 [0.256, 0.378] ∗ 0.357 [0.284, 0.432] ∗ 0.378 [0.296, 0.463] ∗ Bias, survival–self-expression 0 0.000 [ − 0.029, 0.031] 0.043 [0.010, 0.076] ∗ 0.085 [0.048, 0.124] ∗ 0.118 [0.067, 0.169] ∗ Bias, tradition–secular 0 − 0.047 [ − 0.093, − 0.004] ∗ − 0.049 [ − 0.098, − 0.001] † − 0.046 [ − 0.092, − 0.002] ∗ 0.067 [0.018, 0.114] ∗ Spread ratio, tradition–secular 1 1.09 [0.98, 1.23] 0.65 [0.59, 0.73] ∗ 1.02 [0.90, 1.17] 0.89 [0.78, 1.01] Spread ratio, survival–self-expression 1 0.97 [0.88, 1.06] 0.89 [0.80, 0.99] † 0.91 [0.81, 1.04] 0.77 [0.67, 0.93] ∗ Implicit temporal anchor C1; 36 moving countries Lag behind the present, years 0 4.5 [2.6, 6.6] ∗ 7.0 [4.9, 9.3] ∗ 6.0 [4.0, 8.1] ∗ 4.7 [2.6, 7.0] ∗ Lag behind the present, wave-steps 0 0.67 [0.39, 0.94] ∗ 0.97 [0.69, 1.25] ∗ 0.89 [0.61, 1.19] ∗ 0.69 [0.39, 1.03] ∗ Anchored to the most recent wave 100% 56% 36% 42% 58% Change alignment C2; moving countries unless noted Drift-free timing signal, all movers (36) > 0 0.123 [0.047, 0.203] ∗ 0.085 [0.001, 0.173] † 0.049 [ − 0.019, 0.114] 0.003 [ − 0.051, 0.064] four-wave countries only (14) > 0 0.204 [0.066, 0.338] ∗ 0.158 [0.017, 0.300] ∗ 0.101 [0.012, 0.200] ∗ 0.028 [ − 0.060, 0.114] excluding reduced-base countries (27) > 0 0.096 [0.008, 0.189] ∗ 0.063 [ − 0.032, 0.155] 0.029 [ − 0.050, 0.102] 0.020 [ − 0.036, 0.079] Rate ratio, directional countries (10) a 1 0.56 [0.43, 0.75] ∗ 0.45 [0.28, 0.62] ∗ 0.58 [0.35, 0.78] ∗ 0.68 [0.44, 0.79] ∗ Movement, low-change countries (4) a,b 0.097 0.113 0.126 0.170 0.145 Effect of naming the year (C1 − C2) 0 0.001 [ − 0.011, 0.014] − 0.005 [ − 0.016, 0.006] 0.009 [ − 0.001, 0.019] 0.006 [ − 0.004, 0.017] Reversals C2; 6 verified countries Sharp turn at the true bend 6 / 6 2 / 6 3 / 6 2 / 6 0 / 6 Genuine reproduction 6 / 6 1 / 6 1 / 6 2 / 6 0 / 6 Median turn at the bend 155.5 ◦ 58 ◦ 127.5 ◦ 37 ◦ 40.5 ◦ Note. Values are means over countries; brackets are 95% confidence intervals from 2,000 bootstrap replicates, resampled identically across models so that between- model comparisons are paired. Reference is the value a model would show if it represented the quantity perfectly and ∗ marks an interval excluding it. † marks three intervals that exclude the reference by less than 0.002 and are better read as marginal than as decisive; two of the three reverse under the alternative anchoringcondition. The C0 row is descriptive and carries no interval; the accuracy gain below it is the tested quantity. The shift when the country is named is a Euclideandistance, non-negative by construction, so its interval is not tested against zero. Quantities are in standardized cultural-map units unless noted. The axes spanroughly ± 0.7 units, and the mean country sits 0.40 units from the global centroid. a Median rather than mean, because the magnitude ratio has a near-zero denominator on reversing paths and a skewed distribution. b Descriptive only; with four countries this row is not bootstrapped and carries no interval. S23