Position: The Current AI Conference Model is Unsustainable! Diagnosing the Crisis of Centralized AI Conference

1 Introduction

• Centralized AI conferences are increasingly unsustainable because their growth strains peer review, sustainability, researcher well-being, venue capacity, and equitable participation.The paper identifies these pressures as undermining conferences’ core goals and social contract.Source: §1 · pp. 1–3
• Rapid paper growth overwhelms peer review, encourages hyper-competition and questionable optimization, and can make annual presentations outdated relative to AI research’s shrinking lifecycle.The paper notes that the research lifecycle is often under seven months, while conference cycles remain annual.Source: §1 · pp. 1–2
• Centralized conferences impose substantial environmental and logistical pressures through travel emissions, venue limits, and attendance growth.These pressures are presented alongside broader strains on knowledge dissemination and community building.Source: §1 · p. 2
• The model also intensifies anxiety, job insecurity, diminished belonging, and exclusion, weakening community building and the conference social contract.The paper links these human costs to high-stakes publication pressures and large-scale events.Source: §1 · pp. 2–3
• The Community-Federated Conference model separates peer review, dissemination, and networking into globally coordinated but locally organized layers.Its proposed structure combines unified rolling peer review, regional hubs, and digital connectivity.Source: §1 · p. 3

Scientific Mission

• AI conferences advance research by providing efficient peer-reviewed platforms for original, cutting-edge work and academic exchange.Acceptance at a top conference signals community recognition of research value.Source: §2 · p. 3

Knowledge Dissemination

• They disseminate results and recognize influential contributors through presentations, discussions, awards, and honors.Examples include an outstanding-paper award for direct preference optimization and a test-of-time award for Adam.Source: §2 · p. 3

Community Building

• Workshops, panels, and networking activities build community by encouraging discussion, collaboration, new ideas, and exploration of future directions.The paper describes workshops as informal venues for work-in-progress and future research directions.Source: §2 · p. 3

Social Contract

• The conference social contract promotes diversity, equity, and inclusion through measures such as travel grants, mentorship, and codes of conduct.These practices aim to broaden participation and perspectives in AI.Source: §2 · p. 4

3 Methodology for Assessing the Conference Crisis

• The methodology combines quantitative publication, environmental, and infrastructure analyses with qualitative or computational sentiment analysis of community discussions.The framework assesses publication trends, travel emissions, review and venue strain, and mental-health indicators.Source: §3 · p. 4

Data Sources

• The analysis draws on CSRankings data, official conference statistics, 405 Reddit discussion threads, and first-author affiliation data.These sources support publication, attendance, social-discourse, and travel-pattern analyses.Source: §3 · p. 4

Metrics and Analysis Framework

• Publication pressure is measured using effective faculties and contribution per person, with faculty affiliations prorated across areas.Contribution per person is calculated as pub_count/effective_faculties.Source: §3 · pp. 4–5
• Environmental impact is estimated by multiplying travel activity data by established emission factors to produce CO2-equivalent emissions.Activity data includes estimated round-trip flights and hotel stays or author-affiliation travel patterns.Source: §3 · pp. 4–5
• Research lifecycle lag measures the gap between a paper’s initial public availability and its formal conference presentation, while community well-being is assessed through sentiment and mental-health keyword prevalence.The lifecycle lag is treated as an indicator of inefficient knowledge dissemination.Source: §3 · p. 5

4 Challenges in Current AI Conferences

• AI publications grew almost exponentially, reaching 21,400 in 2024 and projected to exceed 65,000 by 2030, overwhelming peer review and shortening research relevance.The projection’s 95% prediction lower bound is 66,572, more than tripling the 2024 figure.Source: Figure 1 · p. 5, §4 · p. 6

“Publish or Perish” Treadmill: Hyper-Productivity

• AI faculty nearly doubled while per-person output doubled within a decade to exceed 4.5 papers annually, outpacing other fields and intensifying productivity pressure.The projected per-capita contribution could exceed 6 papers annually by 2030 if the exponential trend continues.Source: “Publish or Perish” Treadmill:… · pp. 6–7, Figure 3 · p. 7

Environmental Toll: Mounting Carbon Footprint

• NeurIPS 2024 flight emissions reached 8,254 tCO2e, exceeding Vancouver’s estimated daily emissions of roughly 6,849 tCO2e.The estimate covered 3,836 unique first authors and illustrates how in-person attendance creates financial, time, and participation burdens.Source: Environmental Toll: Mounting Carbon… · p. 8, Figure 4 · p. 8, Figure 5 · p. 9

Human Cost: A Community Under Strain

• Among 405 conference-related Reddit threads, 71.3% were negative and 34.6% of negative threads contained mental-health keywords, including anxiety, burnout, and stress.The paper links this strain to pressure, public criticism, reduced collaboration, and weakened knowledge sharing.Source: Human Cost: A Community Under Strain · pp. 9–10, Figure 6 · p. 10
• NeurIPS submission rejections grew near-exponentially while acceptances grew linearly, and the research cycle can make work outdated before presentation.The resulting resubmissions further strain reviewers, while growing attendance has prompted hybrid formats and parallel conferences.Source: Human Cost: A Community Under Strain · p. 10, Figure 7 · p. 11, Figure 8 · p. 11

Physical Breaking Point: Venue Overload

• The Vancouver Convention Centre’s approximately 18,000-person capacity prompted a lottery for non-author registrations, limiting access for students, early-career researchers, and unaffiliated attendees.The paper presents these constraints as structural limits of centralized venues for inclusive participation.Source: Physical Breaking Point: Venue Overload · p. 11

Why palliative measures fail

• Incremental fixes leave overwhelming submission volumes, environmental impact, and equity barriers largely unresolved.Source: §5 · p. 11

Submission limits

• Submission caps shift rather than reduce publish-or-perish pressure and can disproportionately penalize junior researchers and productive laboratories.Source: §5 · pp. 11–12

Multi-site formats

• Multi-site conferences reduce some long-haul travel but preserve centralized review burdens, annual high-stakes evaluation, reviewer burnout, and author effort.Source: §5 · p. 12
• Multi-site formats may create a two-tiered system if one venue is perceived as more prestigious, contradicting equitable participation.Source: §5 · p. 12

Toward systemic reform

• A lasting solution requires replacing centralized structure, synchronized deadlines, and monolithic formats with decentralization, flexibility, and community agency.Source: §5 · p. 12

Model rationale

• The CFC model responds to environmental, financial, psychological, and logistical costs by replacing centralized conferences with a distributed architecture.Source: §6 · pp. 12–13

Core architecture

• CFC decouples peer review and publication, knowledge dissemination, and community building into distinct but interconnected layers.Source: §6 · p. 12

Global peer review and publication

• Rolling global peer review separates submissions from physical meetings, easing reviewer burden and enabling more thoughtful feedback.Source: §6 · p. 12

Regional dissemination and networking

• Regional hubs of 500 to 1,500 participants reduce mega-venue dependence and long-distance travel while supporting inclusion and meaningful interaction.Source: §6 · p. 13

Digital synchronization and collaboration

• A digital layer links local participation to global discourse through streamed plenaries, permanent poster halls, and thematic virtual channels.Source: §6 · p. 13

Distinctive design

• Unlike traditional hybrid meetings, CFC treats digital participation as fully integrated and equally valued within a community-built system.Source: §6 · p. 13

• The Community-Federated Conference model separates peer review, presentation, and community interaction into globally coordinated but locally executed components.The proposal is intended to support sustainability, accessibility, and deeper academic engagement.Source: §7 · p. 13
• The paper argues that incremental adjustments have not resolved the centralized model’s systemic pressures, including hyper-competition, reviewer overload, and logistical constraints.Source: §7 · p. 13
• The authors invite structural reform to reimagine how AI researchers gather, share knowledge, and build scientific community.Code and accompanying materials are made available by the paper.Source: §7 · p. 13

A Detailed Emission Factors

• Table 2 presents emission factors for AI conference components using data retrieved from the UK’s DEFRA GHG Conversion Factors 2025.Source: A Detailed Emission Factors · p. 17, Table 2 · p. 17

• The section combines component-level emission factors with conference-category information to support analysis of AI conference emissions.Source: A Detailed Emission Factors · p. 17, B Detailed AI Conference Categories · p. 17, Table 2 · p. 17

B Detailed AI Conference Categories

• Table 3 maps computer science areas to their corresponding top-tier conferences.The conference list is based on the CSRankings website.Source: B Detailed AI Conference Categories · p. 17

• Table 4 presents ICLR conference attendance data from 2023 through 2025.Source: C Conference Attendance Data · p. 17
• The attendance dataset concerns ICLR rather than a general set of AI conferences.Source: C Conference Attendance Data · p. 17
• The table establishes attendance as the section’s measured conference-scale variable.Source: C Conference Attendance Data · p. 17

• Table 9 presents carbon emissions from ICLR.Source: D Carbon Emissions from ICLR · p. 17
• The surrounding materials include tables for computer science conference categories and ICLR attendance data from 2023–2025.Source: Table 3 · p. 18, Table 4 · p. 19
• Figure 9 is titled “Carbon Emissions from ICLR.”Source: Figure 9 · p. 20

• The section focuses on ICLR Reddit discussions.Source: E Reddit Discussions · p. 20, Table 5 · p. 21
• Table 5 presents examples drawn from those discussions.Source: E Reddit Discussions · p. 20, Table 5 · p. 21
• The examples reflect mental health struggles among discussion participants.Source: Table 5 · p. 21