Paper deep dive
Toward a Sustainable Software Architecture Community: Evaluating ICSA's Environmental Impact
Mahyar T. Moghaddam, Mina Alipour, Torben Worm, Mikkel Baun Kjærgaard
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 94%
Last extracted: 4/10/2026, 2:32:41 AM
Summary
This paper presents a systematic carbon footprint audit of the IEEE International Conference on Software Architecture (ICSA) 2025, evaluating both the digital footprint from GenAI usage in research papers and the traditional footprint from conference logistics (travel, accommodation, catering, and venue operations). The study provides a methodology for measuring these emissions and offers recommendations for fostering a more climate-conscious software architecture research community.
Entities (4)
Relation Signals (3)
ICSA 2025 → hascarbonfootprint → Carbon Footprint
confidence 95% · We report two separate carbon inventories relevant to the software architecture research community... the conference attendance and operations footprint of ICSA 2025
Generative AI → contributesto → Carbon Footprint
confidence 90% · GenAI tools are increasingly integrated into software architecture research, yet the environmental impact of their computational usage remains largely undocumented.
EcoLogits → measures → Carbon Footprint
confidence 90% · EcoLogits, a tool that estimates the energy use and greenhouse gas emissions from LLM inferences.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Generative AI (GenAI) tools are increasingly integrated into software architecture research, yet the environmental impact of their computational usage remains largely undocumented. This study presents the first systematic audit of the carbon footprint of both the digital footprint from GenAI usage in research papers, and the traditional footprint from conference activities within the context of the IEEE International Conference on Software Architecture (ICSA). We report two separate carbon inventories relevant to the software architecture research community: i) an exploratory estimate of the footprint of GenAI inference usage associated with accepted papers within a research-artifact boundary, and ii) the conference attendance and operations footprint of ICSA 2025 (travel, accommodation, catering, venue energy, and materials) within the conference time boundary. These two inventories, with different system boundaries and completeness, support transparency and community reflection. We discuss implications for sustainable software architecture, including recommendations for transparency, greener conference planning, and improved energy efficiency in GenAI operations. Our work supports a more climate-conscious research culture within the ICSA community and beyond
Tags
Links
- Source: https://arxiv.org/abs/2604.04096v1
- Canonical: https://arxiv.org/abs/2604.04096v1
Trouble viewing inline? Open PDF directly →
Full Text
48,521 characters extracted from source content.
Expand or collapse full text
Toward a Sustainable Software Architecture Community: Evaluating ICSA’s Environmental Impact Mahyar Tourchi Moghaddam mtmo@mmmi.sdu.dk SDU Software Engineering Odense, Denmark Mina Alipour mial@mmmi.sdu.dk SDU Software Engineering Odense, Denmark Torben Worm tow@mmmi.sdu.dk SDU Software Engineering Odense, Denmark Mikkel Baun Kjærgaard mbkj@mmmi.sdu.dk SDU Software Engineering Odense, Denmark ©2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. Accepted for publication in ICSA-C, International Conference on Software Architecture Companion Proceedings. Abstract—Generative AI (GenAI) tools are increasingly in- tegrated into software architecture research, yet the environ- mental impact of their computational usage remains largely undocumented. This study presents the first systematic audit of the carbon footprint of both the digital footprint from GenAI usage in research papers, and the traditional footprint from conference activities within the context of the IEEE International Conference on Software Architecture (ICSA). We report two separate carbon inventories relevant to the software architecture research community: i) an exploratory estimate of the footprint of GenAI inference usage associated with accepted papers within a research-artifact boundary, and i) the conference attendance and operations footprint of ICSA 2025 (travel, accommodation, catering, venue energy, and materials) within the conference time boundary. These two inventories, with different system bound- aries and completeness, support transparency and community reflection. We discuss implications for sustainable software ar- chitecture, including recommendations for transparency, greener conference planning, and improved energy efficiency in GenAI operations. Our work supports a more climate-conscious research culture within the ICSA community and beyond. Index Terms—Sustainability, Software Architecture, GenAI. I. INTRODUCTION Generative AI (GenAI) has quickly spread across communi- ties, including software architecture. Large Language Models (LLMs) and other GenAI tools are increasingly used for tasks such as code generation, design assistance, and requirements analysis [1]. Recognizing this trend, ICSA 2025 encouraged submissions on Architecture & Generative AI, including top- ics such as the use of LLMs for design decisions, pattern identification, source code generation, design review, and trade-offs 1 . While GenAI claims to improve productivity and innovation, it also incurs high computational costs, especially high energy use, which leads to carbon emissions. Researchers have raised concerns about the environmental impact of AI, e.g., estimating that training a single large NLP model can generate hundreds of tons of CO2-equivalent (CO 2 eq) [2]. Studies [3] and companies’ disclosures [4] show that a typical query requires approximately 0.3-0.5 Wh of electricity, with 1 https://conf.researchr.org/track/icsa-2025/icsa-2025-papers? #Call-for-Papers longer or more complex prompts consuming several watt- hours under realistic serving conditions. In-person conferences also generate substantial carbon emissions, primarily due to travel. Prior studies show that attendee air travel accounts for the majority of conference emissions [5], a finding confirmed by life-cycle assessments across events [6]. While face-to-face participation remains important for community building, more sustainable planning approaches are needed. As the ICSA community both adopts GenAI and contin- ues to meet physically, assessing the combined digital and traditional footprint becomes essential. Using ICSA 2025 in Odense, Denmark, as a case study, we measure: i) emissions from GenAI use in accepted papers, and i) emissions from travel, accommodation, venue operations, catering, and mate- rials, based on anonymous data 2 . To date, software architecture conferences have not been systematically assessed in this way, limiting awareness of their environmental impact. This paper aims to address the gap by conducting a carbon footprint assessment for ICSA 2025. We explore the following research questions: RQ1: To what extent did ICSA authors use GenAI, and what is the estimated carbon footprint of this usage? RQ2: What are the carbon emissions from traditional confer- ence activities at ICSA 2025? RQ3: What is the per-capita emission? By answering these questions, we quantify digital and physical emissions, identify improvement areas, and propose strategies for more sustainable conferences and research. The contributions are: i) We present a method for detecting and measuring GenAI tool usage in scientific research outputs. This combines automated detectors with manual verification and energy estimation techniques. i) We provide the first empirical estimates of the carbon footprint linked to GenAI in software architecture research. We also include a detailed account of all major conference-related emissions. i) We dis- cuss practical recommendations for the software architecture community, including encouraging transparency in reporting 2 https://bit.ly/41D5G8k arXiv:2604.04096v1 [cs.SE] 5 Apr 2026 Physical Conference Emissions • Long-haul and regional travel • Hotels & accommodation • Catering • Venue energy, ICT, materials Digital Research Emissions: • GenAI use in accepted papers Sustainable Architecture Topics • Energy-efficient software • Sustainable cloud architectures • Carbon-aware task offloading • Low-carbon inference pipelines Multiplier Effect of a Paper • Reuse across multiple projects • Replication across organizations • Deployment over several years Fig. 1. Net-zero framing for the ICSA community. computational use, adopting greener research practices, and rethinking conference organization (such as promoting green travel and catering) to reduce carbon impact. Beyond quantifying ICSA 2025’s emissions, this study frames the conference footprint as a measurable baseline for improvement. As Figure 1 shows, rather than viewing it solely as a drawback, it can motivate greater emphasis on sustainability-oriented software architecture research, such as sustainable architectures, carbon-aware task offloading, and low-carbon inference pipelines. An open question is how much sustainability-driven research would be required to offset the conference’s annual carbon footprint. Embedding this perspective into long-term research planning highlights the potential role of software architecture in supporting broader decarbonization efforts. The remainder of this paper is organized as follows. Section I presents background on GenAI and conference sustain- ability. Section I describes the methodology. Section IV reports the results. Section V discusses the findings, Section VI outlines limitations, and Section VII concludes. I. BACKGROUND A. GenAI in Software Architecture: its Good and Bad Software architects are increasingly exploring the use of GenAI to support complex design and engineering tasks [7]. GenAI models, particularly LLMs such as GPT-3/4 and BERT variants, can generate source code, simulate design decisions, and translate requirements into architectures [8]. For instance, researchers examined whether LLMs have ar- chitectural knowledge [9], e.g., the ability to answer design questions or recognize patterns, and assessed how effectively they can generate or refactor architectural artifacts. This is evident in the ICSA 2025 papers discussing the use of GPT for architectural knowledge and component generation [10]– [12]. The interest in GenAI is due to its ability to manage large amounts of information and automate creative tasks, increasing productivity and supporting decision-making. Several accepted papers at ICSA 2025 highlighted applications such as using LLMs for architecture traceability and generating architectural components with serverless platforms [13], [14]. Recent ICSA 2025 papers illustrate how GenAI can support more sustainable software architecture. For example, self- adaptive edge-AI approaches aim to improve energy efficiency through hyperparameter optimization and dynamic model adaptation [15]. Similarly, agent-based frameworks have been proposed to manage energy and resources in data centers by monitoring demand and adjusting cooling or power usage [16]. These works demonstrate how GenAI-enabled techniques can contribute to reducing software system emissions. At the same time, awareness of GenAI’s hidden envi- ronmental costs is increasing. LLM-based services require substantial computational resources, yet their energy use often remains invisible to end users [17]. As researchers increas- ingly rely on tools such as ChatGPT for writing or analysis, cumulative energy consumption may become significant. Prior discussions have warned that the broader environmental impact of AI can be underestimated amid rapid adoption [18]. While GenAI offers clear benefits, its energy implications warrant careful consideration. B. Carbon Cost of GenAI GenAI’s carbon footprint arises primarily from training and inference. Training large models is energy-intensive and can generate substantial CO 2 eq emissions [2]. For example, train- ing GPT-3 required approximately 1,287 MWh of electricity and emitted an estimated 552 tons of CO 2 eq [19]. In software architecture research, however, models are typically accessed via APIs or pre-trained checkpoints, making inference the relevant stage for footprint estimation. Although inference consumes less energy per request, its cumulative impact can be significant. Median energy use per LLM query is around 0.34 Wh, rising to several Wh for long prompts, and large-scale or repeated use can accumulate rapidly. As model sizes and adoption grow, inference-related energy consumption may rival or exceed training [20]. Accurately measuring inference emissions is challenging because proprietary cloud models obscure hardware and en- ergy details. Researchers therefore rely on proxy measure- ments or emission calculators. For example, prior work repli- cated typical GenAI tasks on local GPUs and measured energy with Carbontracker [18], finding that fine-tuning tasks were more carbon-intensive than prompt-based usage. Limited transparency and inconsistent reporting further complicate precise estimation. C. Measuring Sustainability of Events Sustainability research increasingly examines the carbon footprint of academic events. Life-cycle assessments compar- ing in-person, hybrid, and online formats consistently show that in-person conferences have the highest emissions, largely driven by long-distance air travel [21], [22]. Other sources, such as venue energy, accommodation, materials, and local transport, contribute smaller shares [23]. The Hotel Carbon Measurement Initiative (HCMI) pro- vides standardized estimates for hotel emissions, typically 5– 15 kg CO 2 eq per room-night for operational impacts [24], potentially rising to around 20 kg CO 2 eq when indirect effects are included [25]. Venue emissions depend on building characteristics and the local energy mix. Assessing ICSA 2025 therefore establishes a baseline to identify high-impact mitigation opportunities. Papers identified from IEEE Xplore (n=108) Logistics data obtained from organizers Energy and emission baselines gathered from literature and Danish public data Papers screened (n=108) Transportation, lodging, venue energy, catering, material, and digital infrastructure data screened for correctness Energy and emission data screened and linked to ICSA2025 context Data Collection and Scope Screening Papers excluded for no possible use of GenAI (n=89) Estimating energy and emissions for GenAI inference Estimating traditional conference emissions Estimating Energy Use Interpreting the carbon footprint breakdown Implications for sustainable software architecture Mitigation strategies for future conferences Reflections on GenAI’s role and responsible use Analysis and Recommendations RQ1. GenAI usage and emissions from Papers RQ2. Emissions from traditional conference activities RQ3. Overall footprint and GenAI’s share Formulating Research Questions Fig. 2. Research methodology. I. RESEARCH METHODOLOGY A. Data Collection and Scope Papers Data: We analyzed the 108 accepted ICSA 2025 papers (28 main track, 80 companion) to ensure a manageable and replicable scope. Ethics and data protection are discussed in Section I-F. Activity Data: To estimate emissions from conference logistics, we collected organizer-provided data on 229 in- person attendees, their geographical distribution and regis- tration types, the conference schedule, venue details at the University of Southern Denmark (room sizes and usage), and the catering menu. We considered the following emission sources: • GenAI in Papers: Inference-related computing for GenAI (text/image generation). Training and traditional computa- tion were excluded. • Transportation: Travel to and from Odense (primarily air, train, or car), including minor local transfers. • Accommodation: Hotel stays based on registration category, using HCMI-aligned per-night factors. • Venue Energy: Electricity for lighting, HVAC, audiovisual systems, and digital infrastructure, estimated from venue size and typical energy intensity. • Catering: Meals and refreshments calculated using life- cycle carbon intensity factors. • Materials and Waste: Printed materials, conference items, and waste, estimated using standard LCA factors. All emissions are reported in kg CO 2 eq using region- specific emission factors (e.g., Denmark’s electricity grid intensity). B. Detecting GenAI Usage Identifying how authors use GenAI in research is challeng- ing, as they often do not clearly state its use. We conducted our study using automated text analysis and a manual review as follows: 1. Keyword Search: We scanned the papers for specific terms related to GenAI, such as “ChatGPT”, “GPT-3”, “GPT-4”, “large language model”, “Transformer”, “BERT”, “Generative AI”, “LLM”, “OpenAI API”, “prompt”, “fine-tune” to high- light relevant sections. Mentions of papers and tools linked to GenAI also indicated potential usage. 2. AI-Generated Text Detection: We obtained access to the GPTZero 3 API for research use to determine whether parts of the papers might be AI-generated [26]. Because results can be unreliable, we considered a paper as possibly containing AI-written text only if the entire paper was flagged by the detectors with at least 50% confidence. 3. Manual Content Analysis: We reviewed the methodology sections to identify applications of GenAI, using a coding scheme to track possible research stages and purposes. We looked for quantitative data to estimate energy. After completing these steps, we created a list of papers that likely used GenAI. If a paper showed no signs of GenAI or was detected with less than 50% probability, we assumed zero GenAI-related emissions. If it mentioned GenAI without details, we estimated possible usage scenarios. C. Estimating Energy and Emissions for GenAI Inference In our assessment of the carbon footprint from generative AI usage at ICSA 2025, we used EcoLogits 4 [27], a tool that estimates the energy use and greenhouse gas emissions from LLM inferences. EcoLogits tracks energy consumption and environmental impacts from generative AI models accessed via API calls, providing metrics like energy consumed and carbon emissions. The impact of one request is: I request = E request F em + ∆T ∆L I e servere , where E request is the energy used by the server (including an overhead for cooling) to serve the request, F em is the emission factor of the electricity (kgCO 2 eq per kWh) at the data center, ∆T /∆L is the fraction of the server’s lifetime consumed by this inference, and I e server is the total embodied GHG impact of the server hardware. For our analysis, we applied EcoLogits to each instance of generative AI use in the ICSA 2025 submissions, e.g., using ChatGPT for tasks like text generation. Key inputs included the model used, the output token count, server hard- ware, data center location, and efficiency factors. By inputting these parameters, EcoLogits helped us calculate the energy 3 https://gptzero.me. 4 https://ecologits.ai/latest/ consumption and carbon emissions for each usage event. We then totaled the emissions across all events to quantify the conference-wide impact. It is important to note that this estimate is based on certain assumptions, particularly for models with limited publicly available information. The details of deployment can vary: factors such as hardware efficiency and the data center’s carbon intensity can influence the results. We used standardized inputs to minimize inaccuracies. D. Traditional Conference Emissions 1) Transportation: When we look at the environmental impact of in-person conferences, the largest contributor is travel. To estimate how much CO 2 eq each participant gen- erates from their travel, we first assessed their emission by their origin country and further categorized their journeys as local, neighbor, regional (within Europe), or long-haul (intercontinental). For local and neighborhood travelers, we assumed they primarily use trains, while those coming from long distances would typically opt for flights. When it comes to flights, we used specific emission factors: approximately 0.15 kg of CO 2 eq per passenger-kilometer for shorter flights and 0.11 kg/km for longer flights [28]. The impact of non- CO 2 eq emissions occurring at high altitudes should also be in- cluded. The exact calculation was performed using myclimate calculator that considers all possible contributing variables [29]. Train travel was assigned a much lower emissions factor of 0.024 kg/km per passenger [30] specified for Denmark, considering that most of Europe’s train network is powered by electricity. On the other hand, car travel for a few local attendees was estimated at 0.137 kg/km, which is typical for a single-occupancy petrol car [30]. The overall transportation impact was simply calculated by adding up the emissions from all individual trips made by attendees: C transport = X i∈P d i × EF mode(i) ,(1) where d i is the round-trip distance for participant i and EF mode(i) is the emission factor for their mode of travel. This accounted for a dominant share of the conference’s carbon footprint, as reported in Section IV. 2) Accommodation: For each non-local attendee, we es- timated emissions from hotel stays during the conference. We assumed an average of 4 nights per main conference attendee, 6 nights for full conference attendees, and 3 nights for workshop attendees. We adopted the HCMI factor of ap- proximately 5.87 kg CO 2 eq per room-night as a representative average for hotel stay emissions in Denmark [31]. Therefore, if N hotel is the total number of attendee-nights in hotels, the accommodation footprint is: C accommodation = N hotel × 5.87 kg CO 2 eq/night(2) This yields on the order of tens of kilograms CO 2 per attendee based on their registration category. Accommodation appears as a significant but smaller source than travel. 3) Venue Energy: The conference took place in university buildings, thereby consuming energy for electricity, heating, and ventilation. To estimate the venue’s energy consumption, we considered the venue’s size (approximately 1000 square meters, including auditoriums and classrooms) and the event duration (8 hours per day for 5 days). Using typical energy intensity values for modern academic buildings [32], we calcu- lated the total electricity required for lighting, AV equipment, Wi-Fi, and heating or cooling. Our estimates, together with the venue’s technical service, suggested that the overall energy demand (including digital infrastructure) for the conference was around 2,300 kWh. Next, we accounted for the carbon intensity of the local grid, which in Denmark is approxi- mately 0.173 kg of CO 2 eq per kWh [33]. This allowed us to estimate the venue’s carbon emissions at 397.9 kg CO 2 eq. In formula form, the venue footprint could be expressed as C venue = E venue × EF grid . Considering the registrations and their categories, a person’s share of venue-related energy use is 3.03 kWh/day, and venue-related emissions are 0.523 kg CO 2 eq/day. 4) Catering: When assessing the environmental impact of conference catering, we can break it down by considering the CO 2 eq emissions per person per meal. For instance, a full conference attendee would have approximately five lunches, one reception, and one banquet dinner over the course of the event, along with coffee breaks each day. For lunch (30% vegetarian and 70% meat-based), we calculated roughly 2.5 kg of CO 2 eq per person [34], [35], while the richer banquet dinner was estimated at 5 kg CO 2 eq, and reception 3.5 kg CO 2 eq. ICSA 2025 also served a light dinner (sandwiches) on 3 evenings, each of which accounted for 0.85 kg CO 2 eq. The lighter refreshment breaks (coffee and cake/snacks) contribute about 0.3 kg CO 2 eq each. When we add these together, one attendee’s total food-related emissions for the event are approximately 26.55 kg of CO 2 eq if they attended the full conference. For full conference participants, if we multiply this by the number of attendees (N ) in that category, we can use the following formula to estimate the catering emissions for a full conference attendee (other categories will have reduced days): C catering = N× (5× 2.5 + 1× 5 + 1× 3.5 + 10× 0.3 + 3× 0.85). This formula will be applied to all other registration cat- egories, in which the number of days and the number of attended events vary. We acknowledge that actual emissions may differ slightly due to special dietary options provided, but catering typically accounts for a smaller share of the event’s total emissions. 5) Materials and Waste: Emissions from conference mate- rials and waste were relatively lower. We included the footprint of printed materials (badges, posters) and any conference swag, as well as waste disposal. Based on rough estimates made together with the organizers’ technical service, printing and paper use per person emitted approximately 0.3 kg of CO 2 , attendee swag another 2.3 kg, and waste (food packaging, etc., not recycled) 0.5 kg. In aggregate, we estimate Materials & Waste contributed∼709.9 kg CO 2 eq. E. Assumptions Our objective is to provide a transparent, reproducible base- line of ICSA’s footprint and identify the dominant emission drivers at an order-of-magnitude level. As is common in con- ference footprinting [21], [36], some activity data (e.g., travel mode choice) are not observable post hoc without dedicated surveys, and therefore must be modeled using explicit assump- tions. To avoid overstating precision, we mentioned all key assumptions and considered scenario bounds for parameters that materially affect results, such as the share of rail vs air travel in neighbor and regional categories. F. Ethics, Governance, and Data Protection This study is a meta-research assessment of the environ- mental footprint of a scientific event and related research artifacts. It does not involve interventions, experiments, or interaction with human participants. To minimize privacy risks, we analyzed only: i) publicly available bibliographic/paper content, and i) aggregated, anonymized conference statistics. We did not receive, store, or process direct identifiers (such as names, addresses, emails, or any identities). For conference attendance and logistics, we followed GDPR-aligned data minimization and purpose limitation principles and restricted our inputs to the least granular data necessary for carbon accounting 5 . Where any underlying administrative data were used to produce aggregate statistics, these were handled under SDU’s institutional governance processes for research and data protection (including the use of SDU-approved storage and access control, and deletion or anonymization after project completion) 6 previously obtained for ICSA 2025. In line with IEEE publication policies for work involving human-related data 7 , we include this explicit ethics and data protection statement and clarify that no individual-level personal data were processed in our analysis and reporting. IV. RESULTS A. GenAI Usage in Papers (RQ1) Out of the 108 papers at ICSA 2025, we identified 26 papers (24%) that used Generative AI. This moderate but notable share reflects its emerging but not yet dominant role in software architecture research. Only one paper reported their use of GenAI as an assistant in content generation. Our methodology also flagged 11 additional papers with a probability of GenAI use below 50%, which were excluded from emission auditing. It is possible that some authors used GenAI indirectly, e.g., to get feedback and ideas, which our analysis cannot confirm with certainty. In addition to text generation, the identified use cases of GenAI in ICSA papers include: i) architecture design assistance, where models like GPT-4 and GPT-3 are used to generate and evaluate design-related code stubs through 5 https://eur-lex.europa.eu/eli/reg/2016/679/oj/eng 6 https://sdunet.dk/en/research/legal-services/researchportal legalgdpr 7 https://conferences.ieeeauthorcenter.ieee.org/author-ethics/ guidelines-and-policies/submission-policies/ interactive queries; i) requirements traceability and analysis, e.g., a paper leveraging a pre-trained LLM to extract archi- tecture component names from text, utilizing named-entity recognition, which required considerable computation; i) AI as subject, e.g., a paper studied a GenAI model, likely GPT- 3.5 or GPT-4, by asking it software architecture questions, focusing solely on inference-based responses without modify- ing the model; iv) code generation and testing, where, e.g., a study indicated the use of a generative code assistant for automating the creation of alternative code implementations to test energy efficiency. These papers demonstrate how LLMs enhance design, analysis, research, and code generation in software architecture, all of which are necessary to advance the field. Across the GenAI-using papers, no instance of training a large model from scratch was found. All usages were via API, fine-tuning smaller models, or using open pre-trained checkpoints. This suggests that most software architecture re- searchers focus primarily on using generative AI for inference rather than on creating or training generative AI models. As a result, the environmental impact, in terms of carbon emissions, is largely limited to the inference process rather than the entire development phase. The explicit mention of GenAI use was detected in the papers’ evaluation phase (e.g., comparing LLMs with human performance), the implementation phase (e.g., generating artifacts), and the data analysis phase (e.g., traceability), and as previously mentioned, only one paper explicitly mentioned using GenAI for literature review or writing. Overall, GenAI played a role in more than one-quarter of the research works, indicating an emerging presence but not yet ubiquity in software architecture research. In terms of emissions, the CO 2 eq of papers ranged from 176 g CO 2 eq (with 363 Wh of energy consumption) to 524 g CO 2 eq (with 1077 Wh of energy consumption). The total emissions from papers’ GenAI use were 8694 g CO 2 eq (with 17827 Wh of energy consumption), which would have been 36720 g CO 2 eq (with 75168 Wh of energy consumption) if all papers used GenAI to generate their content. The results confirm that, while increasing awareness is crucial, GenAI- generated emissions from research are far lower than those of traditional conference activities. B. Emissions from Traditional Conference Activities (RQ2) The carbon emissions from the traditional physical com- ponents of ICSA 2025 far exceeded those from GenAI. Travel was by far the dominant contributor, accounting for approximately 94% of total emissions, 172.218 tonnes of CO 2 eq just from attendee transportation. This aligns with prior studies that found that attendee air travel accounts for most of conference emissions. However, the share of emissions from travel is higher for ICSA 2025 due to the high number of intercontinental attendees, and as Denmark is relatively sustainable in other conference activities. The most emitting activity after travel was lodging, with 5.89 tonnes of CO 2 eq, approximately 3% of total emissions. Catering (meals and beverages) was on the order of 2% of emissions with 4.26 tonnes of CO 2 eq, Materials and waste emitted 710 kg CO 2 eq and Venue emitted 398 kg CO 2 eq. Fig. 3. The comparison of worldwide regions participants (full conference) emissions (kg CO 2 eq) and the share of various emission contributors for each region. The smaller chart shows the percentage of emissions by region, indicating that, for local and neighboring countries, participants’ lodging and catering dominate, whereas for others, travel accounts for the largest share. C. Overall Footprint (RQ3) By aggregating all available data from the calculations above, the estimated total carbon footprint for ICSA 2025 is around 183 t CO 2 eq. With 229 in-person attendees, this corresponds to approximately 0.8 t CO 2 eq per participant, comparable to the emissions from a round-trip air travel from southern Italy to Denmark. This per-person share is quite high for a brief event, largely attributable to a significant proportion of long-distance travel. In Figure 3, the larger bars show CO 2 eq associated with participants traveling from local, neighboring, and intercontinental regions, and the inset specifies the composition of each region’s contribution. Long- haul travel considerably increases the emissions of interconti- nental attendees compared to those from Europe or Denmark. For participants located closer to Odense, travel emissions are relatively minor, and hotel stays and catering become the primary sources of carbon emissions for both local and neighboring delegates. Our results thus suggest a current asymmetry: the envi- ronmental cost of disseminating research far exceeds that of conducting it, at least in software architecture, where computational demands are moderate. This may not hold in fields with extreme AI computation (where research itself can emit tens of tons), but in our case, it is a striking insight. Going forward, it will be essential to monitor GenAI’s footprint, as small does not mean zero, yet the immediate priority for emissions reduction lies in rethinking conference logistics. V. DISCUSSION A. Interpreting the Carbon Footprint Breakdown The high share of transportation emissions highlights that any meaningful reduction in a conference’s carbon footprint must involve addressing travel. Improvements in other areas are beneficial with comparatively smaller gains. The ICSA Steering Committee can thus consider models such as regional hubs to reduce travel, which might yield far greater carbon savings than any on-site operational tweaks. To bend the emissions curve, decisions about where and how we meet are paramount. That said, smaller changes should not be ignored; measures such as greener catering or venue-level renewable energy use still align with sustainability values and can be implemented at low cost, thereby strengthening a culture of environmental mindfulness. Furthermore, an awareness of AI’s impact can promote greater responsibility as its use continues to grow. B. Implications for Sustainable Software Architecture The software architecture research community can draw several lessons from these findings. First, there is a strong case for greater transparency in reporting computational re- sources and emissions in our research. In our study, we had to infer GenAI usage post-hoc, but going forward, authors could be encouraged (or required) to disclose any extensive use of AI models or cloud computing and even estimate the associated energy/carbon footprint. This effort could align with best practices in other fields like machine learning that have introduced paper checklists, including environmental impact disclosures. Making such reporting a norm would raise awareness and encourage researchers to consider efficiency in their methodology. Second, our community might widen its notion of what constitutes good software architecture practice to include sustainability considerations. Just as architects de- sign systems with quality attributes in mind, we should treat carbon footprint as a relevant quality attribute of both our research processes and the systems we study. Embracing this perspective, ICSA and similar conferences could innovate in how they are structured, or better said, architected, exploring distributed or hybrid conference models that maintain schol- arly exchange while dramatically cutting emissions. Lastly, the findings reinforce that to improve sustainability, our primary focus should be on the big-ticket items (e.g., logistics) when planning events. At the same time, as GenAI becomes more common in our research, we should promote efficient use of computing (e.g., use smaller models or local resources when feasible) so that the digital side of research remains as sustainable as possible. C. Mitigation Strategies for Future Conferences Building on our results, we outline a few strategies that could reduce the carbon footprint of future ICSA editions: Hybrid conferences: as ICSA maintains strong community ties, a physical meet-up is necessary. However, even partial virtualization helps, e.g., alternating in-person and hybrid conference years, or hosting a single event with both physical and virtual tracks, could cut the travel footprint by a large margin. Regional Meetups and Co-located Events: instead of one global gathering, the conference could adopt a distributed model (multiple regional hubs linked by video) or coordinate timing with other conferences. Presumably, if ICSA is sched- uled right after another major conference in the same area, international attendees could get the chance to join both events in one trip. This could cut down on their flight emissions for each event. Also, offering regional workshops or local mirror sites could be a great option, making it easier for participants to attend something closer to home. Low-Carbon Travel Incentives: encourage and facilitate travel by train or other low-carbon means for those who are within a reasonable distance. Organizers can support more sustainable travel by offering incentives like registration discounts for train travelers or by facilitating local attendees to coordinate ride-sharing. Including carbon offsets in the registration fee or through sponsors can also help balance out travel emissions, while it is essential to handle offsets transparently and responsibly. Sustainable Catering and Accommodation: adopt plant- forward catering menus (vegetarian or vegan by default), since meals with less meat can substantially lower food-related emis- sions. Similarly, partner with hotels that have sustainability certifications or use renewable energy, and inform attendees about greener lodging options. These steps ensure that the necessary emissions from food and lodging are minimized. Reducing Materials and Waste: continue phasing out printed materials by using digital programs and conference apps. Avoid disposable swag or replace it with eco-friendly or non- physical tokens (e.g., donating to an environmental cause on behalf of attendees). While these actions have a small direct carbon impact, they are symbolically important and contribute to a culture of sustainability. Greener Research Practices: promote energy-efficient com- puting in research. For example, the conference could intro- duce a recognition (an informal Green Paper Award) for work that achieves its results with remarkably low resource usage or by using innovative efficiency techniques. Encouraging authors to use renewable-powered infrastructure or to share their compute footprint in submissions can gradually make green AI a part of the community ethos. D. Reflections on GenAI’s Role and Responsible Use GenAI’s growing prevalence in research requires forward- looking consideration. If, in the near future, a much larger frac- tion of papers leverage GenAI or conduct compute-intensive experiments, the cumulative emissions could increase. The community should be proactive in establishing norms for responsible AI use. This could include expectations to disclose usage and to favor energy-efficient modeling techniques when possible. Introducing an environmental lens to research ethics is a novel but important idea; for instance, if a certain evaluation would consume an extraordinary amount of energy for marginal scientific gain, should researchers rethink it? By making energy impact a consideration in research design, we encourage innovation in more sustainable methods. Finally, it is worth exploring ways that GenAI itself could contribute to sustainability solutions. AI tools might assist in optimizing conference planning (e.g., smarter scheduling to minimize travel) or help architects design more energy-efficient software systems. As we integrate GenAI into our workflows, we should also incorporate sustainability thinking into how we use these tools. The goal is to reap the benefits of GenAI- driven innovation while keeping its energy footprint in check. VI. LIMITATIONS While this study aims to raise awareness, several limitations apply: 1) Many estimates rely on assumptions and proxy data (e.g., modeled GenAI usage), introducing uncertainty. The total footprint may vary by approximately±20%, though this does not affect the main conclusions. 2) Emission factors differ across sources and methodologies (e.g., flights, accommodation, catering, and grid intensity). We used representative averages; actual values may vary depending on context and specific providers. 3) The analysis focuses on major emission sources and ex- cludes some indirect impacts (e.g., hardware manufactur- ing, full life-cycle effects, and shared travel for multi- purpose trips). Some data center overhead may not be fully captured. 4) Detecting GenAI usage in papers is imperfect; subtle or minor uses may have gone unnoticed. However, substantial usage is unlikely to have been systematically missed. 5) We evaluated environmental impact only and did not as- sess economic, social, or research-quality implications of proposed mitigation strategies. 6) The study considers carbon footprint (CO 2 -equivalent) only and does not include other environmental dimensions such as water use or e-waste. 7) As a mid-sized software engineering conference, ICSA 2025 may not represent larger or more compute-intensive events. 8) GenAI inference estimates rely on limited publicly avail- able data; future transparency may enable more precise calculations. Despite these limitations, the analysis provides a transparent and sufficiently robust baseline to support the study’s conclu- sions and recommendations. VII. ACKNOWLEDGMENT OF AI USE In this paper, we made limited and transparent use of AI-supported tools. Grammarly was used for proofreading, editing, shortening, rephrasing, and refinement. We used Chat- GPT with a 38-word prompt to experiment with two papers totaling 21450 words, contributing 4.83 g CO 2 eq. We utilized the AI detection tool (GPTZero) to examine the ICSA 2025 papers for any AI-generated content. All the analytical rea- soning, methodological choices, and significant contributions presented in this paper are solely those of the authors. VIII. CONCLUSION We presented the carbon footprint assessment of ICSA 2025, including the traditional conference emissions and the impact of GenAI usage in research. Our analysis found that the conference’s total emissions were approximately 183 t CO 2 eq, with roughly 94% attributable to attendee travel. Improving sustainability in the ICSA community depends on reducing travel and greening our event planning, as well as optimizing researchers’ use of AI tools. The carbon cost of all GenAI-related activities at the conference was 8694 g CO 2 eq. We thus encourage transparent and efficient use of computational resources as GenAI becomes more common, so that its footprint remains minimal. REFERENCES [1] M. Esposito, X. Li, S. Moreschini, N. Ahmad, T. Cerny, K. Vaid- hyanathan, V. Lenarduzzi, and D. Taibi, “Generative ai for software architecture. applications, trends, challenges, and future directions,” Applications, Trends, Challenges, and Future Directions, 2025. [2] E. Strubell, A. Ganesh, and A. McCallum, “Energy and policy con- siderations for deep learning in nlp,” in Proceedings of the 57th annual meeting of the assoc for computational linguistics, 2019, p. 3645–3650. [3] N. Jegham, M. Abdelatti, C. Y. Koh, L. Elmoubarki, and A. Hendawi, “How hungry is ai? benchmarking energy, water, and carbon footprint of llm inference,” arXiv preprint arXiv:2505.09598, 2025. [4] S. Altman, “The gentle singularity,” Sam Altman Blog, 2025, accessed: 2025-12-10. [Online]. Available: https://blog.samaltman.com/ the-gentle-singularity [5] V. Mannheim and J. L. Avat ́ o, “Examining the carbon footprint of conferences with an emphasis on energy consumption and catering,” Energies, vol. 18, no. 2, p. 321, 2025. [6] S. Neugebauer, M. Bolz, R. Mankaa, and M. Traverso, “How sustainable are sustainability conferences?–comprehensive life cycle assessment of an international conference series in europe,” Journal of cleaner production, vol. 242, p. 118516, 2020. [7] A. Nguyen-Duc, B. Cabrero-Daniel, A. Przybylek, C. Arora, D. Khanna, T. Herda, U. Rafiq, J. Melegati, E. Guerra, K.-K. Kemell et al., “Generative artificial intelligence for software engineering—a research agenda,” Software: Practice and Experience, vol. 55, no. 11, p. 1806– 1843, 2025. [8] A. Fan, B. Gokkaya, M. Harman, M. Lyubarskiy, S. Sengupta, S. Yoo, and J. M. Zhang, “Large language models for software engineering: Sur- vey and open problems,” in 2023 IEEE/ACM International Conference on Software Engineering: Future of Software Engineering (ICSE-FoSE). IEEE, 2023, p. 31–53. [9] R. Dhar, K. Vaidhyanathan, and V. Varma, “Can llms generate architec- tural design decisions?-an exploratory empirical study,” in 2024 IEEE 21st International Conference on Software Architecture (ICSA). IEEE, 2024, p. 79–89. [10] M. Soliman and J. Keim, “Do large language models contain software architectural knowledge?: An exploratory case study with gpt,” in 2025 IEEE 22nd International Conference on Software Architecture (ICSA). IEEE, 2025, p. 13–24. [11] M. Shamsujjoha, Q. Lu, D. Zhao, and L. Zhu, “Swiss cheese model for ai safety: A taxonomy and reference architecture for multi-layered guardrails of foundation model based agents,” in 2025 IEEE 22nd Int Conference on Software Architecture (ICSA). IEEE, 2025, p. 37–48. [12] P. Liu and L. Xiao, “Improving clinical decision support: Architecture design of a multi-agent system based on an argument quality assessment ontology,” in 2025 IEEE 22nd International Conference on Software Architecture (ICSA). IEEE, 2025, p. 313–323. [13] D. Fuchß, H. Liu, T. Hey, J. Keim, and A. Koziolek, “Enabling architec- ture traceability by llm-based architecture component name extraction,” in 2025 IEEE 22nd International Conference on Software Architecture (ICSA). IEEE, 2025, p. 1–12. [14] S. Arun, M. Tedla, and K. Vaidhyanathan, “Llms for generation of architectural components: An exploratory empirical study in the server- less world,” in 2025 IEEE 22nd International Conference on Software Architecture (ICSA). IEEE, 2025, p. 25–36. [15] A. Matathammal, K. Gupta, L. Lavanya, A. V. Halgatti, P. Gupta, and K. Vaidhyanathan, “Edgemlbalancer: A self-adaptive approach for dynamic model switching on resource-constrained edge devices,” in 2025 IEEE 22nd International Conference on Software Architecture Companion (ICSA-C). IEEE, 2025, p. 543–552. [16] V. D. Sera and S. Kalra, “Dynamic architectures leveraging ai agents and human-in-the-loop for data center management,” in 2025 IEEE 22nd International Conference on Software Architecture Companion (ICSA- C). IEEE, 2025, p. 324–333. [17] E. Barnett, “Visualizing ai energy use: Can consumption reminders promote social good mindsets?” Ph.D. dissertation, Texas Christian University Fort Worth, Texas, 2025. [18] N. Inie, J. Falk, and R. Selvan, “How co2stly is chi? the carbon footprint of generative ai in hci research and what we should do about it,” in Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, 2025, p. 1–29. [19] D. Patterson, J. Gonzalez, Q. Le, C. Liang, L.-M. Munguia, D. Rothchild, D. So, M. Texier, and J. Dean, “Carbon emissions and large neural network training,” arXiv preprint arXiv:2104.10350, 2021. [20] R. Desislavov, F. Mart ́ ınez-Plumed, and J. Hern ́ andez-Orallo, “Trends in ai inference energy consumption: Beyond the performance-vs-parameter laws of deep learning,” Sustainable Computing: Informatics and Sys- tems, vol. 38, p. 100857, 2023. [21] Y. Tao, D. Steckel, J. J. Kleme ˇ s, and F. You, “Trend towards virtual and hybrid conferences may be an effective climate change mitigation strategy,” Nature communications, vol. 12, no. 1, p. 7324, 2021. [22] A. Cavallin Toscani, A. Atasu, L. N. Van Wassenhove, and A. Vinelli, “Life cycle assessment of in-person, virtual, and hybrid academic conferences: New evidence and perspectives,” Journal of Industrial Ecology, vol. 27, no. 6, p. 1461–1475, 2023. [23] S. Van Ewijk and P. Hoekman, “Emission reduction potentials for academic conference travel,” Journal of Industrial Ecology, vol. 25, no. 3, p. 778–788, 2021. [24] Sustainable Hospitality Alliance, “Hotel carbon measurement initiative (hcmi) methodology – version 2.0,” https://sustainablehospitalityalliance. org/wp-content/uploads/2020/02/HCMI-v2.0-introduction.pdf, 2020, ac- cessed: 2025-12-10. [25] V. Filimonau, J. Dickinson, D. Robbins, and M. A. Huijbregts, “Re- viewing the carbon footprint analysis of hotels: Life cycle energy analysis (lcea) as a holistic method for carbon impact appraisal of tourist accommodation,” Journal of cleaner production, vol. 19, no. 17-18, p. 1917–1930, 2011. [26] G. A. Adam, A. Cui, E. Thomas, E. Napier, N. Shmatko, J. Schnell, J. J. Tian, A. Dronavalli, E. Tian, and D. Lee, “Gptzero: Robust detection of llm-generated texts,” 2026. [Online]. Available: https://arxiv.org/abs/2602.13042 [27] S. Rinc ́ e and A. Banse, “Ecologits: Evaluating the environmental impacts of generative ai,” Journal of Open Source Software, vol. 10, no. 111, p. 7471, 2025. [28] DepartmentforEnvironment,Food&RuralAffairs (DEFRA),“Ukgovernmentghgconversionfactorsfor companyreporting,”https://w.gov.uk/government/publications/ greenhouse-gas-reporting-conversion-factors-2023, 2023. [29] Myclimate, “Myclimate flight co2 emissions calculator,” https://co2. myclimate.org/en/flight calculators/new, 2025, accessed: 2025-12-10. [30] CarbonLabel.org, “carbonlabel.org data,” https://carbonlabel.org/data, 2025, accessed: 2025-12-10. [31] Climatiq, “Climatiq emission factor, co2 (kg) per passenger-km, short-haulflight,economyclass,”https://w.climatiq.io/data/ emission-factor/31c8e342-a53e-4ae0-8fb7-dc1e69fa6965,2025, accessed: 2025-12-10. [32] A. Merabtine, C. Maalouf, A. A. W. Hawila, N. Martaj, and G. Polidori, “Building energy audit, thermal comfort, and iaq assessment of a school building: A case study,” Building and Environment, vol. 145, p. 62–76. [33] NowTricity, “Electricity mix denmark,” https://w.nowtricity.com/ country/denmark/, 2025, accessed: 2025-12-10. [34] Danish Ministry of Food, Agriculture and Fisheries (FVM), “Climate is served: Facts and figures on denmark’s food consumption and green- house gas emissions,” https://en.fvm.dk/Media/638499759964791542/ Climate isServed.pdf, 2021, accessed: 2025-12-10. [35] DenStoreKlimadatabase,“Denstoreklimadatabase,” https://denstoreklimadatabase.dk/en, 2025, accessed: 2025-12-10. [36] U. EPA, “Greenhouse gas inventory guidance,” in Indirect Emissions from Events and Conferences. US EPA Center for Corporate Climate Leadership, 2018, p. 2018–12.