Big AI and the Metacrisis: When Scaling Becomes a Liability

Scale is one of business’s favourite words.

A product that scales can serve more customers without proportionally increasing costs. A platform that scales becomes harder to displace. An infrastructure provider that scales can convert technical advantage into market power.

The awkward question is what else scales with it.

More AI usage can mean more useful outputs, lower unit costs, and wider access. It can also mean more infrastructure demand, more dependence on dominant platforms, more synthetic content competing for attention, and more institutional influence concentrated among the organisations able to build frontier systems.

Steven Bird’s policy paper, Big AI is Accelerating the Metacrisis: What Can We Do?, examines this second set of effects.¹ Its subject is not merely the environmental footprint of large language models. The paper presents Big AI as an institutional system that connects corporations, data centres, language technologies, governments, universities, conferences, labour markets, and public discourse.

Its central claim is that this system accelerates three already-interacting crises:

• an ecological crisis, involving emissions, water consumption, mineral extraction, pollution, and e-waste;
• a meaning crisis, involving attention capture, disinformation, epistemic harm, weakened critical thought, and declining social participation;
• a language crisis, involving the erosion of linguistic diversity, community knowledge, and cultural sovereignty.

The important word is interacting. Bird is less interested in adding three columns to an AI risk register than in drawing arrows between them.

Risk registers like tidy categories. Real systems prefer feedback loops.

The Paper’s Unit of Analysis Is the Feedback Loop

In ordinary AI governance, risks are usually divided into manageable boxes.

A sustainability team measures energy consumption. A model-risk team tests hallucinations and bias. A legal team reviews data rights. A public-policy team watches regulation. Each function produces its own controls, documentation, and reassuringly rectangular dashboard.

Bird’s argument is that these separations conceal the mechanism that matters most. Ecological, social, linguistic, commercial, and academic systems influence one another. A mitigation that improves one local metric may leave the wider system unchanged or even support further expansion.

The paper’s first figure captures this argument conceptually. Large language models and data centres sit near the centre of three overlapping crises. Beneath them, Big AI exchanges money, technology, talent, legitimacy, and influence with governments, universities, and professional conferences.

The figure should be read as a mechanism map, not a causal estimate. It identifies relationships the author believes deserve attention. It does not calculate their relative strength, prove that each relationship is equally important, or measure Big AI’s marginal contribution to each crisis.

A simplified version of the paper’s logic looks like this:

Corporate and institutional incentives
                ↓
      Larger models and deployments
                ↓
 Ecological, meaning, and language harms
                ↓
      Interactions among the crises
                ↓
 Greater social and institutional fragility

The model is therefore broader than “LLMs consume resources.” It asks what happens when scale becomes the organising principle of the entire system around language technology.

Three Crises, Each with a Different Mechanism

The three crises should not be treated as interchangeable. Each operates through a different mechanism, and each has a different evidence boundary.

Ecological pressure grows through infrastructure and rebound

The ecological mechanism is the most familiar.

Training and operating large models require data-centre capacity. That capacity is associated with electricity demand, greenhouse-gas emissions, water consumption, critical-mineral extraction, hardware production, and eventual e-waste.

The paper adds a less convenient observation: efficiency does not necessarily reduce total resource consumption. When producing an AI output becomes cheaper, organisations often produce more outputs, build more applications, and attract more investment. Lower resource use per task can coexist with higher aggregate resource use.

For businesses, this makes a per-query efficiency metric incomplete. A system that cuts energy use per inference by half while increasing inference volume tenfold has not solved its total-exposure problem. It has improved unit economics and expanded the environmental bill.

The paper does not provide a quantitative estimate of how much LLM growth contributes to global ecological stress. Its ecological argument is instead a warning about direction, incentives, and cumulative demand.

The meaning crisis concerns the surrounding business model

“Meaning crisis” is the paper’s broadest and least easily measured category.

It includes disinformation, attention capture, harmful classifications, erosion of critical thinking, weakening of creative work, declining knowledge diversity, and damage to democratic processes. The paper also stresses that an LLM does not possess reliable access to truth or shared social norms merely because it produces convincing language.

This category can sound vague until the unit of analysis is widened.

An individual model response might be accurate. A platform filled with automatically generated, engagement-optimised content can still degrade the information environment. A chatbot might help one user complete a task. A business model built around maximising interaction can still encourage dependency, passive consumption, or the replacement of human participation.

The mechanism is therefore not simply “the model hallucinates.” It is the interaction between generative capacity, platform incentives, attention markets, and institutional trust.

This distinction matters commercially. Model accuracy is a product-quality metric. The meaning crisis concerns the consequences of repeatedly deploying the product inside a particular incentive system.

The language crisis is political before it is technical

The paper’s language-crisis argument challenges one of the industry’s more comfortable narratives: that expanding multilingual model coverage naturally supports endangered or minoritised languages.

Bird argues that language loss is primarily sociopolitical. Communities lose languages through persecution, displacement, cultural disruption, economic pressure, and the interruption of intergenerational transmission. Adding a language to a model does not reverse those forces.

The technical assumptions behind multilingual scaling are also poorly matched to many communities. Numerous languages lack standardised writing systems, large digitised corpora, or enough data to support robust model development. Communities may already use regional contact languages for external communication while preserving local languages for social, cultural, and ecological knowledge.

A model can therefore appear inclusive while extracting data, shifting participation toward dominant languages, or weakening the local practices that keep a language alive.

The relevant governance question is not merely, “Does the model support this language?”

It is, “Does this deployment strengthen the community that speaks it, respect its authority over data, and preserve meaningful participation outside the platform?”

The Cross-Crisis Loops Are the Actual Thesis

The paper becomes more distinctive when it connects the crises.

An environmental-footprint critique could be addressed with cleaner energy, better chips, and smaller models. A misinformation critique could be addressed with provenance systems and moderation. A language-access critique could be addressed with more multilingual data.

Bird argues that these responses remain incomplete because the crises reinforce one another.

Ecological crisis and meaning crisis

Environmental disruption can produce fear, anxiety, and helplessness. Attention-driven platforms can turn those emotions into engagement. Repeated exposure to alarming content may then encourage doomscrolling, distraction, or resignation rather than collective action.

The reverse connection also matters. An information environment dominated by attention capture, disinformation, and social fragmentation can weaken the capacity of communities to coordinate responses to ecological threats.

In this loop, the ecological crisis supplies emotionally powerful content. The meaning crisis reduces society’s ability to respond constructively to the conditions generating that content.

Meaning crisis and language crisis

Generative systems can produce enormous volumes of content in dominant languages at almost negligible marginal cost. Local-language communication, which depends on smaller communities and human participation, must compete against this industrial output.

Attention capture can also reduce participation in local social life. When fewer people engage with elders, community institutions, and local forms of knowledge, language transmission weakens.

The reverse effect follows naturally. Language loss disrupts the transfer of memory, identity, and knowledge. That disruption can weaken wellbeing and deepen the sense of disconnection that defines the meaning crisis.

Language crisis and ecological crisis

The language–ecology connection is the least obvious and perhaps the most important.

Languages frequently encode knowledge about land, species, medicine, seasons, and stewardship. When a language and its community institutions weaken, ecological knowledge may disappear with them. The capacity of Indigenous and local communities to care for ancestral lands can also be undermined by displacement and loss of authority.

Ecological damage then accelerates the language crisis in return. Mining, climate disasters, biodiversity loss, and forced displacement separate communities from the places, practices, and stories through which language remains meaningful.

The result is not three parallel problems. It is a system in which damage travels.

Why Technical Safety Cannot Repair Institutional Incentives

After describing the crisis interactions, the paper turns toward a more confrontational claim: business as usual will not correct them.

Bird argues that much AI ethics and safety work remains trapped inside the assumptions of the organisations producing the systems. Ethical problems are translated into technical variables. Fairness becomes a metric. Safety becomes another evaluation suite. Governance becomes a model card, a committee, or a vendor-authored framework.

These measures can be useful. The paper’s objection is that they leave the underlying incentives intact.

A corporation rewarded for expanding usage can improve its guardrails while continuing to increase total deployment. A platform can reduce the rate of harmful outputs while multiplying the volume of generated content. A company can publish an ethics framework while lobbying against binding restrictions. A conference can host critical research while relying on sponsorship and infrastructure from the firms being criticised.

Technical controls operate on the system’s outputs. The paper is concerned with who controls the system, who benefits from its expansion, and which institutions legitimise it.

This is also why the paper calls the scalability story a myth. Scale increases the number of interactions, use cases, languages, jurisdictions, and potential failure modes. Monitoring and mitigation must then expand around them. Guardrails accumulate around monitoring systems, which accumulate around increasingly complex deployments.

The paper describes this as a perpetual game of Whac-a-Mole. The industry’s preferred solution to the consequences of scaling is frequently another technical layer that must itself be scaled.

A safety stack can grow impressively tall without changing the direction of travel.

How to Read the Paper’s Evidence

This is a policy paper, not an empirical study. It contains no benchmark experiment, causal identification strategy, ablation study, or quantified estimate of the proposed interventions’ effects.

That does not make it useless. It changes what its evidence can support.

The distinction is important because some of the paper’s language is sweeping. It describes Big AI as accelerating the metacrisis and questions whether its benefits justify its harms. Those conclusions are argued through synthesis and institutional critique rather than demonstrated through a single comparable dataset.

The paper is strongest as a diagnostic framework. It is weaker as a measurement instrument.

Prestige Is Part of the Production System

The paper’s most useful institutional insight is easy to miss beneath its broader critique.

Big AI does not scale through computing resources alone. It also scales through prestige.

Frontier companies possess infrastructure, proprietary systems, data, and funding unavailable to most researchers. Access produces publishable results. Publications create academic recognition. Recognition attracts talent, investment, grants, partnerships, and further access.

Universities and professional conferences are not passive observers in this cycle. They help determine which research questions appear important, which methods appear rigorous, and which achievements receive status.

The paper targets the Association for Computational Linguistics because the ACL controls some of these mechanisms. It publishes research, convenes conferences, establishes ethical expectations, shapes calls for papers, and influences what counts as prestigious work in natural language processing.

Bird does not claim that the ACL can regulate Big AI by itself. The paper explicitly acknowledges that intergovernmental organisations will ultimately have greater regulatory power. The ACL matters because it can shape professional norms before formal regulation arrives.

This logic is developed through the paper’s second figure, adapted from Kohlberg’s levels of moral development.

Expecting every researcher to resist institutional pressure on the basis of universal ethical principles assumes unusually strong individual independence. Bird proposes a more pragmatic strategy: change the expectations of the professional community so that responsible conduct becomes conventional rather than heroic.

This is not just an academic concern.

An organisation that depends on employees voluntarily challenging profitable but harmful projects has designed a weak governance system. Ethical behaviour becomes more reliable when procurement rules, promotion criteria, review processes, budgets, and approval thresholds support it.

Culture is useful. Incentives are usually awake earlier.

Seven ACL Interventions, Grouped by the Mechanism They Target

The paper proposes seven actions for the ACL. Read individually, they can look like a conventional list of governance recommendations. Grouped by mechanism, they form a more coherent programme.

The proposals are deliberately institutional. Bird is not asking researchers to invent one more safety technique. He is asking the professional community to alter the conditions under which research is funded, reviewed, rewarded, and legitimised.

The paper also encourages smaller language models and forms of research less committed to unrestricted growth. It does not demonstrate that smaller models are automatically safer, more useful, or more ethical. Small systems can still exploit labour, misuse data, or damage communities.

Their relevance is that they make constraint imaginable. They challenge the assumption that every successful use case should eventually migrate toward the largest available system.

What a Metacrisis-Aware AI Review Looks Like

For businesses, the paper’s practical value lies in changing the scope of review.

Traditional AI governance asks whether a particular model meets a defined set of requirements. A metacrisis-aware review also asks what happens when the model, its supplier, its business model, and its deployment portfolio expand together.

Several operational changes follow from this broader view.

Review total demand, not only unit efficiency

Efficiency remains valuable, but it should not be mistaken for restraint. Organisations can set total compute, energy, water, and deployment budgets alongside per-task targets. Growth assumptions should be included when evaluating environmental claims.

Add an explicit necessity test

Not every process improved by generative AI requires generative AI. A necessity test asks whether a simpler model, conventional automation, search system, human workflow, or narrower domain tool can deliver sufficient value with lower exposure.

This is less glamorous than model selection. It is also where many avoidable costs disappear.

Treat community effects as governance inputs

For deployments involving minoritised languages or culturally specific knowledge, technical coverage is an inadequate success criterion. Organisations should assess community authority over data, consent, benefit sharing, local participation, and the possibility that the system will displace rather than support existing practices.

Examine institutional dependencies

Vendor risk is not limited to uptime, cybersecurity, and contractual terms. Dependence on one provider can shape an organisation’s research agenda, product roadmap, risk appetite, and ability to challenge the provider’s claims.

Independent governance becomes difficult when the subject of oversight supplies the infrastructure, expertise, and funding required to perform it.

Change what receives internal prestige

Research and product teams respond to what senior management celebrates. If promotions and budgets consistently reward model scale, user growth, benchmark gains, and deployment speed, ethical review will remain an administrative obstacle attached to a growth system.

Organisations can instead reward avoided deployment, smaller adequate solutions, independent evaluation, community benefit, and evidence that a system improves human capability rather than merely replacing activity.

What the Paper Shows, What Business Can Infer, and What Remains Uncertain

The article’s business relevance depends on keeping three layers separate.

What the paper directly argues

The paper directly argues that:

• Big AI should be understood as an institutional and political system, not merely a collection of models;
• ecological, meaning, and language crises interact and can reinforce one another;
• technical safety and voluntary corporate governance are inadequate when the underlying incentives favour continued expansion;
• professional institutions influence the direction of a field through prestige, publication, sponsorship, and norms;
• the ACL can use those mechanisms to encourage critical, policy-oriented, smaller-scale, and community-centred research.

What Cognaptus infers for business practice

For companies, the framework implies that AI risk governance should examine cumulative system exposure.

A deployment can satisfy every local checklist and still contribute to a problematic portfolio. A model can be efficient per task while total resource consumption rises. A multilingual product can expand access while weakening local authority. A responsible-AI programme can produce excellent documentation while remaining dependent on the commercial incentives it is meant to constrain.

The paper therefore supports a broader governance architecture combining:

• portfolio-level exposure reviews;
• independent supplier and sponsorship scrutiny;
• total resource and deployment limits;
• community and data-sovereignty assessments;
• protected internal spaces for critical review;
• R&D incentives that do not treat scale as an achievement by default.

These are interpretations of the paper’s mechanism, not tested recommendations with established returns.

What remains uncertain

The paper does not quantify:

• Big AI’s marginal contribution to each crisis;
• the relative importance of the proposed feedback loops;
• the threshold at which scale becomes net harmful;
• the comparative impact of large and small models across use cases;
• the financial effect of adopting the proposed governance measures;
• whether the seven ACL interventions would meaningfully change corporate behaviour.

These uncertainties limit prediction. They do not eliminate the value of the questions.

The Argument Is Broad; Its Evidence Is Uneven

The paper’s ambition is also its main limitation.

“Big AI” covers a wide range of corporations, technologies, relationships, and practices. The paper deliberately does not identify a precise set of companies. This makes the concept flexible enough to describe an institutional system, but less useful for assigning responsibility or comparing actors.

The three crises are selected because language connects them. Other relevant systems—including military applications, government–university relationships, macroeconomic concentration, and geopolitical competition—receive limited treatment or sit outside the main model.

The evidence base is similarly heterogeneous. The paper connects research from ecology, labour, platform governance, linguistics, Indigenous data sovereignty, political economy, and professional ethics. This breadth is valuable for discovering relationships. It cannot establish that every relationship operates with the same strength, in every context, or primarily because of AI.

Some claims also move faster than the evidence assembled beneath them. The argument that promised AI benefits do not justify present harms is morally and politically significant, but the paper does not conduct a systematic comparison of benefits and harms. Readers should treat it as a position supported by a broad critique, rather than a measured conclusion.

The paper’s strongest contribution survives this limitation.

It demonstrates why evaluating each harm separately can produce a false sense of control. Even when the magnitude of every arrow is uncertain, the existence of interactions changes how risk should be governed.

Scaling Becomes a Liability When It Multiplies the Wrong Things

The usual case for scale is straightforward: once a system works, expand it.

Bird’s paper asks organisations to insert a prior question: what exactly has been shown to work, for whom, under which incentives, and with which consequences excluded from the calculation?

Scaling a useful capability can create value. Scaling attention capture creates more attention capture. Scaling resource consumption creates more resource consumption. Scaling a dominant language can further weaken smaller ones. Scaling a governance system designed by the governed can produce an impressive quantity of governance-shaped paperwork.

The paper does not provide a formula for deciding when AI scale becomes excessive. It provides a more demanding way to frame the decision.

An AI system should be evaluated not only by the quality of its outputs, but by the relationships, dependencies, incentives, and externalities that expand with it.

That turns scale from an automatic objective into a claim requiring evidence.

For an industry accustomed to treating bigger as a result, that is a useful inconvenience.

Cognaptus: Automate the Present, Incubate the Future.