Skip to content

Big AI is Accelerating the Metacrisis: What Can We Do?

Conference: ACL 2026
arXiv: 2512.24863
Code: None (Policy position paper)
Area: AI Ethics / Position Paper / NLP Governance
Keywords: metacrisis, Big AI, ACL Code of Ethics, ecological crisis, linguistic diversity, ethics washing

TL;DR

In this ACL 2026 position paper, Steven Bird argues that "Big AI" (industrialized LLM engineering driven by a few giants) is simultaneously accelerating three intertwined crises—ecological crisis / meaning crisis / language crisis—and that ACL, as the largest publisher of LLM research, must shift from "individual compliance" to "collective action within a professional community," proposing 7 specific reforms for ACL (reaffirming public interest priority, resisting corporate capture, protecting critical NLP, establishing an NLP policy track, etc.).

Background & Motivation

Background: ACL is likely the world's largest publisher of peer-reviewed LLM research, and its Code of Ethics explicitly requires that the "public interest is the paramount consideration." However, the reality is that the vast majority of ACL papers contribute to Big AI (OpenAI / Google / Meta / Microsoft / various startup giants)—the review ecosystem, SOTA tracking, and conference sponsorships have all been captured by industry.

Limitations of Prior Work: The author lists multiple harms in the NLP field that are overlooked or "whitewashed": (1) Data centers lead to greenhouse gas emissions, e-waste, water consumption, and rare mineral mining; (2) LLM content erodes critical thinking, creative work, knowledge diversity, and democracy; (3) Approximately 90% of the world's languages lack standardized writing, and multilingual LLMs do not solve the political roots of language loss; (4) AI safety is inherently non-scalable, and "adding guardrails to play whack-a-mole" is a tactic for Big AI to maintain a deregulated space; (5) Academia is repeatedly corrupted by corporate philanthropy and ethics washing.

Key Challenge: Individual researchers face an ethical conflict between "public interest is paramount" vs. "industrial funder/employer requirements." Using Kohlberg's theory of moral development, the author points out that relying solely on individual compliance is Level 3 (postconventional), which only a minority of adults can reach—making it unreliable. More realistic is Level 2 (conventional): shaping behavior through the collective expectations and standards of a prestigious group like ACL.

Goal: (a) Systematically demonstrate the causal chain between Big AI and the metacrisis; (b) debunk the three myths of "AI self-governance / scalability / benefits outweigh harms"; (c) propose 7 actionable reforms for ACL.

Key Insight: The author stands on the position of "language engineers as a professional community," elevating the contradiction from an "individual moral dilemma" to the level of "professional organizational governance."

Core Idea: The acceleration of the metacrisis by Big AI cannot be solved by individual conscience alone; ACL, as the largest professional organization for language engineering, must act collectively through mechanisms such as collective standards, policy tracks, and independent spaces for critical NLP.

Method

Overall Architecture

The paper follows a policy paper structure:

  1. §2 Cascading Crises: Arguments for LLMs participating in ecological / meaning / language crises, analyzing the feedback loops between them;
  2. §3 More Business As Usual will not Work: Debunking the three myths of Big AI self-governance, scalability, and benefit-to-harm ratio;
  3. §4 What Can We Do?: 7 specific actions for ACL;
  4. §5 Individual vs Community Standards: Explaining the "community standards" route using Kohlberg's theory of moral development;
  5. §6 Conclusion: Urging the abandonment of scalability ethics and the reconstruction of the professional community.

Key Designs

  1. Systemic Causal Framework of Triple Crisis Feedback:

    • Function: Upgrades "LLM harm" from single-point arguments to a systemic demonstration where crises are not isolated but mutually reinforcing.
    • Mechanism: (a) Ecological ↔ Meaning: Ecological anxiety is exploited by LLM content for "attention capture," while doomscrolling numbs ecological anxiety; (b) Meaning ↔ Language: Mainstream language LLM content displaces local languages; language loss undermines elder roles and knowledge transmission; (c) Language ↔ Ecological: Language loss weakens indigenous ability to care for ancestral lands (hotspots of biodiversity); climate disasters and mining for data centers displace indigenous people and erase language communities. Together, these cycles form the metacrisis (Morin & Kern 1999, Lawrence et al. 2024).
    • Design Motivation: Traditional AI ethics literature treats "carbon emissions," "misinfo," and "low-resource languages" separately, diluting their impact; a unified perspective prevents readers from using "I only do A, I don't care about B/C" to evade responsibility.

  2. Debunking Three Myths:

    • Function: Pre-emptively blocking common reader defenses ("Big AI will manage itself / will become greener / benefits outweigh harms").
    • Mechanism:
      • Myth 1 — Big AI will self-govern: Citing Phan, Zuboff, Shelby, and Ressa to point out that Big Tech weaponizes AI ethics as a tool for "delaying regulation"; ethics-washing at GMU/Stanford; the irony of the FAccT conference being sponsored by Google/Facebook/Microsoft; mimicking the Big Tobacco playbook (Abdalla & Abdalla 2021).
      • Myth 2 — Scalability is feasible: Data center carbon/water/mineral consumption has exceeded planetary boundaries; AI safety is inherently non-scalable, always resulting in whack-a-mole; annotation sweatshops exploit low-wage labor in "AI’s hidden outposts."
      • Myth 3 — Benefits outweigh harms: Sequence models are far removed from natural language (Bender & Koller 2020); NLP research SOTA-chasing is a shallow fashion; bias is treated as a bug rather than a feature of classification itself (Crawford 2021); resource consumption grows exponentially for linear performance gains (Schwartz et al. 2020).
    • Design Motivation: Before writing actionable suggestions, the author acknowledges that the "rationalizing narrative of the status quo is wrong," otherwise the suggestions in §4 would be directly countered by the myths in §3.

  3. 7 Actionable Reforms for ACL:

    • Function: Translating philosophical criticism into executable organizational actions.
    • Mechanism: (1) Reaffirming that the Code of Ethics' "public good paramount" applies to member behavior rather than just papers; (2) Protecting ACL from image-cleansing brought by corporate sponsorship from companies like Meta; (3) Re-asserting in the CFP that computational linguistics is the study of "natural human language," encouraging degrowth + small LMs; (4) Establishing an independent track + review process for critical NLP to avoid suppression by gatekeepers; (5) Establishing an NLP policy research track to prepare for future regulation; (6) Issuing public statements in the name of ACL; (7) Advocating for a life-sustaining research vision (Ethics of Care, data feminism, decolonizing methods, etc.).
    • Design Motivation: Vague criticism is easily dismissed as "another anxiety article"; listing 7 specific, board-actionable proposals makes them impossible for the ACL executive committee to ignore.

Loss & Training

Not applicable (policy paper, no model training).

Key Experimental Results

Main Results: Summary of Empirical Evidence for the Three Crises (Figures are real facts cited in the paper)

Crisis Dimension Key Facts / Citations
Planetary Boundary Breaches 6 out of 9 planetary boundaries have been breached (Richardson et al., 2023)
Data Center Consumption Increased greenhouse gases, e-waste, water usage, and rare earth mineral plunder (Crawford 2021; UNEP 2024)
Language Writing Rates Approximately 90% of global languages lack standardized writing (Bird, 2026, etc.)
Multilingual Usage The majority of the global population is already multilingual, using dozens of contact languages for information access and economic participation
SOTA Resource Ratio Exponential growth in resource consumption for linear performance gains (Schwartz et al., 2020)
Multilingual LLM Limits For low-resource languages, there "will never be enough data" to train robust models
Moral Development Levels Kohlberg Level 3 "is only reached by a minority of adults"—individual conscience alone is insufficient

Ablation Study: Comparison of 3 Myths vs Reality

Myth Big AI Narrative Reality Counter-argument in Paper Key Citation
Myth 1: Big AI will self-govern Ethics frameworks are sufficient for constraint Ethics washing; FAccT is sponsored by Big Tech Slee 2020; Ochigame 2022; Bietti 2021
Myth 2: Scalability is feasible Data/compute stacking is sustainably viable Planetary boundaries are broken; guardrail-on-guardrail is non-scalable Bender & Hanna 2025; Slee 2020; Crawford 2021
Myth 3: Benefits > Harms "Solving poverty, sustainable cities, education for all" Sequence models \(\neq\) natural language; SOTA-chasing is fashion; AI is useless for most of the world Bender & Koller 2020; Church & Kordoni 2022; Bender & Hanna 2025

Key Findings

  • Common Cause in Big AI: The author's most persuasive conclusion integrates ecological / meaning / language crisis feedback loops into a single diagram (Fig 1), unifying criticisms of "carbon emissions," "information pollution," and "language displacement" as a single systemic issue.
  • Kohlberg Level 2/3 Argument: Grounding ACL reform suggestions in the empirical psychology that "the majority of adults are at the conventional moral level" is more leveraged than simple slogans about "following ethics."
  • Obvious Conflict of Interest in Big Tech Funding: Corporate capture phenomena—Meta sponsorship, the prevalence of Big Tech employees in the review process, and infrastructure advantages forcing PhD students to follow industry trends—are explicitly named in an ACL paper for the first time.
  • Paradigm Call for Degrowth + Small LMs: Compared to over-funded SOTA-chasing, the author recommends the degrowth/small LM paths of Vetter (2017), Meyers (2023), Wang et al. (2025), and Church (2026), providing guidance for ACL review priorities.

Highlights & Insights

  • The "metacrisis" framework unifies fragmented criticisms into a powerful impact: Previously, Bender (parrot), Crawford (Atlas of AI), Strubell (NLP carbon), and Birhane each attacked separate points; this paper strings them into a systemic argument, making the impact \(1+1>2\).
  • ACL Code of Ethics as a leverage point: The author seizes the existing commitment that "public good is paramount" to force ACL to follow through, which is more actionable than proposing entirely new clauses.
  • Precision of the Big Tobacco analogy: Philanthropic funding, academic capture, establishing "independent" ethics bodies, and industry lobbying against regulation—all four align perfectly with tobacco industry behavior from 1950-90, allowing readers to understand the situation instantly.
  • Meta-political rebuttal to "don't want to get political": Citing Black Project to point out that "those who claim to be apolitical are maintaining political structures they are comfortable with" provides a preventive strike against common "don't talk about politics" objections within ACL.
  • Shift from Individual → Community level: Transferring the moral burden from individual researchers to professional organizational governance changes the discussion from "guilt" to "collective executable mechanisms."

Limitations & Future Work

  • The author acknowledges: (1) "Big AI" refers to a general direction rather than naming specific companies; (2) The paper only covers ecological / meaning / language crises, omitting war, inequality, personal privacy, etc.; (3) Fig 1 omits direct relationships between government, military, and academia, potentially weakening the role of the state in AI governance; (4) The paper targets ACL, but ACL is inherently just a member organization with structural limits on collective action.
  • Personal Observations: (a) The article is purely normative and does not provide quantitative metrics for success (e.g., a "Big Tech paper percentage" threshold); future work needs a metric suite; (b) Definitions of "what counts as small LM / life-sustaining research" are missing, which could lead to inconsistent implementation; (c) The discussion on linguistic/economic differences for ACL members in developing countries is relatively shallow, requiring more detailed discussion on fairness in degrowth for researchers in low-resource nations.
  • Improvement Ideas: (a) Co-proposing petitions to the ACL board with scholars like Strubell, Bender, and Hanna; (b) Designing mandatory carbon/water reports for ACL paper ethics statements (similar to NeurIPS broader impact); (c) Collaborating with law schools and planetary boundary research centers for the NLP policy track to expand leverage.
  • vs. Bender et al. 2021 (Stochastic Parrots): While Stochastic Parrots focuses on the cognitive risks of LMs, this paper expands the scope to the planetary ecosystem; it can be seen as an "extended + governance-oriented" version five years later.
  • vs. Bender & Hanna 2025 (The AI Con): That book systematically critiques the Big AI business model; this paper is its actionable microcosm within the ACL academic community.
  • vs. Crawford 2021 (Atlas of AI): Atlas of AI is an anthropology of material AI; this paper borrows its evidence but provides a reform agenda specific to the NLP academic community.
  • vs. Abdalla & Abdalla 2021 (Grey Hoodie): That work quantified the penetration of Big Tech funding in NLP; this paper uses those conclusions as evidence for corporate capture and offers "how to resist."
  • vs. Schwartz et al. 2020 (Green AI): Green AI advocates for efficient research; this paper elevates efficiency claims to the political-economic level of "degrowth + anti-scalability."
  • Insights: (a) Any "AI for X" field can adopt the same "systemic crisis perspective + society-level governance reform" analysis; (b) This paper's template (demonstration → debunking myths → 7 reforms → moral development argument) can be applied to other professional society reform papers; (c) The Kohlberg community norm leverage approach is also applicable to other high-stakes engineering ethics (bioethics, nuclear ethics).

Rating

  • Novelty: ⭐⭐⭐⭐ The integration of the Metacrisis framework with 7 ACL reforms is explicitly proposed for the first time within the ACL society; the approach is novel, though the theoretical pillars (Bender/Crawford/Ressa) are established.
  • Experimental Thoroughness: ⭐⭐⭐ As a position paper, it conducts no experiments but includes wide-ranging citations (~80 references) with high argumentative density; it is not directly comparable to traditional experimental papers.
  • Writing Quality: ⭐⭐⭐⭐⭐ The levels of argumentation are clear, the prose is powerful, citations are precise, and self-refutation is complete (including 5 "why" questions for self-examination in Limitations); it is an exemplar of high-quality policy writing.
  • Value: ⭐⭐⭐⭐⭐ Directly targeting the ACL governance level, it may influence future conference policy and review practices; it is a milestone for self-reflection in the NLP community.