Skip to main content

The Kehoe Rule

What is the Kehoe Rule?

The Kehoe Rule describes a regulatory pattern where powerful new technologies are allowed to spread widely while the burden of proof falls on demonstrating harm, rather than proving safety before deployment. Named after Robert Kehoe, the lead industry's dominant medical authority in the mid-20th century, who insisted that leaded gasoline should remain in use until opponents could prove it caused harm.

Historical Origin

A century ago, public health experts warned that adding tetraethyl lead to gasoline could contaminate cities and cause "slow lead poisoning." Industry proceeded anyway. Robert Kehoe became the lead industry's dominant medical authority and insisted that leaded gasoline should remain in use until its opponents could prove that it caused harm.

The Pattern Across Industries

The Kehoe Rule has appeared repeatedly across industries:

  • Leaded gasoline — dispersed into air worldwide, fueling epidemics of coronary heart disease
  • Cigarettes — billions sold, spreading lung cancer globally
  • Asbestos — exposed workers and communities, now impossible to fully contain
  • PFAS — synthetic chemicals that contaminated virtually everyone's blood on Earth
  • Pesticides — spread across farms and communities with unknown long-term effects

In each case, companies created useful, profitable products. The trouble began when evidence of harm emerged: companies questioned the science, magnified uncertainty, funded more studies, hired experts and lawyers, lobbied regulators, and delayed restrictions. Too often, they concealed what their own scientists had discovered.

"The people making those decisions were not monsters. The corporation did not need monsters. It needed people to do their jobs."

Human Cost

  • 9 million premature deaths per year from pollution (Lancet Commission on Pollution and Health, 2022)
  • 3.5 million deaths in 2023 attributed to lead alone (Global Burden of Disease study)
  • Toxic chemicals and pollutants may contribute to nearly one in five deaths worldwide

Once dispersed around the world, these substances become enormously expensive — and sometimes impossible — to contain or remove.

The Kehoe Rule Applied to AI: A Critical View

What the Article Gets Wrong: "Agency"

The article's central fear is that AI systems possess agency — autonomous motivations that no human can comprehend. This is not supported by how these systems actually work.

LLMs are next-token predictors. They are facilitated by their harnesses, prompted by users, and deployed by corporations with clear commercial incentives. They have no goals of their own, no desires, no hidden intentions. The claim that AI systems can "pursue goals in ways creators never anticipated" conflates emergent behavior (which is explainable through the training process and prompt context) with agency (which implies autonomous intent).

This framing appears to be driven by effective altruism circles and Silicon Valley's more alarmist voices, which have a vested interest in portraying AI as an existential threat. This narrative serves certain funding priorities and institutional agendas far more than it serves accurate understanding of the technology.

Real Harms: Present-Day and Documented

The real harms from AI are not science fiction — they are concrete, measurable, and happening right now:now. Below is a synthesis of evidence from peer-reviewed research, government reports, and independent investigations.

1. Erosion of the Commons

  • TrainingMass-scale scraping of public content: The Common Crawl dataset alone contains over 250 billion webpages collected over 17 years, making up a significant portion of training data isfor scrapedmajor fromLLMs publicly(Mozilla availableFoundation, "Training Data for the Price of a Sandwich," Feb 2024). A 2024 Mozilla report found that two-thirds of 47 generative LLMs released between [period] used Common Crawl as training data.
  • Creators losing attribution and economic benefit: The New York Times, among many other publishers, has sued AI companies for training on their content — blogs, forums, creative works, encyclopedias — without consent or compensationcompensation, alleging that models can reproduce proprietary works. The Times pushed Common Crawl to remove its content after discovering it was being used (Mozilla Foundation, 2024).
  • ThisOECD drainsanalysis: The OECD published "Intellectual Property Issues in Artificial Intelligence Trained on Scraped Data" (Feb 2025), documenting how AI models trained on scraped data challenge core intellectual property frameworks and leave creators without recourse.
  • European Parliament study: A 2025 European Parliament study ("Generative AI and Copyright") examined how generative AI challenges core principles of EU copyright law, highlighting a legal mismatch between AI training practices and current frameworks.
  • UK House of Lords: The UK House of Lords Communications Committee published "AI, copyright and the sharedcreative knowledgeindustries" commons(Mar that2026), powereddocumenting how creators are losing meaningful control over their works and identities, leading to tangible economic harms.
  • US Copyright Office: The US Copyright Office's "Identifying the web,Economic Implications of Artificial Intelligence for Copyright Policy" report documents how training datasets include scraped information from publicly accessible pages, digitized books, and patent records — turning public resources into private training material
  • Creatorswithout and writers lose attribution and economic benefit as their work becomes embedded in proprietary models
  • The commons is being privatized at scalecompensation.

2. Environmental Impact of Data Centres

  • NewCarbon footprint rivaling nations: A December 2025 study published in Patterns (Cell Press) estimated that AI systems' carbon footprint could match that of a small European country, with AI data centres consumeconsuming enormousan amountsestimated 448 terawatt-hours of electricity in 2025 — and projections showing this will double in just four years.
  • Water usage: The same Patterns study estimated AI's water use footprint at 312.5 to 764.6 billion litres in 2025 — equivalent to the range of global annual bottled water consumption. This is used primarily for cooling
  • Thisdata strainscentres, local power grids, contributes to carbon emissions, and competes with communities for resources
  • Water usage for coolingoften in drought-prone areasregions (Euronews, Dec 2025; PBS News, Dec 2025).
  • Lawrence Berkeley National Laboratory: The LBNL "2024 United States Data Center Energy Usage Report" (published Jan 2025) found that data center load growth has tripled over the past decade and is anprojected under-discussedto butdouble growingor concerntriple by 2028. Data centre demand grew from 176 terawatt-hours in 2023 (~4.4% of total US electricity consumption).
  • UN University report: The United Nations University's "Environmental Cost of Artificial Intelligence: Carbon, Water, and Land Footprints" documents that environmental costburdens is— real,data present,center siting, electricity demand, water withdrawals, land use, mineral extraction, and acceleratinge-waste — are concentrated in specific communities and regions, often disadvantaged ones.
  • Consumer impact: Food & Water Watch's February 2026 report ("A No Brainer: How AI's Energy and Water Footprints...") documents that residential electricity prices jumped 7.1% in 2025 (more than double inflation), with increases topping 20% in some states, driven partly by AI data centre demand. Consumer Reports similarly documented rising electric bills attributed to AI infrastructure.

3. Psychological and Sociological Impacts of Chatbots

  • ChatbotCognitive atrophy: A 2026 study published in iScience (Elsevier) introduced the concept of "AI-chatbots-induced cognitive atrophy" (AICICA), describing how excessive dependence on AI chatbots can lead to deterioration of essential cognitive abilities. The paper reviews evidence and mitigation strategies for this emerging phenomenon.
  • Deskilling and epistemic dependence: Research in Frontiers in Psychology (2026) identified four key outcomes from differential patterns of AI engagement: deskilling (Shukla et al., 2025), creativity loss (Fan et al., 2024), shallow processing, and epistemic dependence. The study distinguishes between "dependent" cognitive offloading (unquestioning reliance on AI) versus "autonomous" use can(AI as a tool for enhanced human reasoning).
  • deskillMIT usersMedia Lab longitudinal study: "How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use" (Fang et al., 2025) — a controlled longitudinal study examining how AI chatbot use affects psychosocial outcomes over time.
  • Adolescent mental health: A PMC-published study on "The Impact of Chatbots on Adolescent Mental Health" found that overreliance on AI chatbots may lead to broader cognitive decline, particularly affecting individuals who have not attained mastery in their fields of study.
  • AI dependence scale development: Research published in Alpha Psychiatry (2026) developed and validated the Artificial Intelligence Chatbot Dependence Scale (AICDS), finding that AI chatbot dependency represents an "emerging mental health concern requiring attention from clinicians, designers, and policymakers." The study also documented links between AI dependence and anxiety/depressive symptoms in university students.
  • MDPI dual-path model: A 2025/2026 study in Information (MDPI) proposed a dual-path model of cognitive and affective mechanisms underlying AI chatbot dependency, going beyond utilitarian drivers to examine the psychological mechanisms at play.
  • Kim et al. (2025): "From algorithm aversion to AI dependence: Deskilling, upskilling, and emerging addictions in the GenAI age" documents how over-reliance on generative AI for writing, analysis, and decision-making erodes core competencies
  • Peoplewhile losecreating thenew abilityforms to think critically when they outsource reasoning to a tool that always produces an answer
  • The psychologicalof dependency on AI companions and assistantsaddiction-like is poorly understood but potentially significant
  • Social skills, creativity, and independent problem-solving may atrophy as people delegate more cognitive workpatterns.

4. OtherMisinformation Documentedat HarmsScale

  • MisinformationUNESCO: atThe UN's "Deepfakes and the Crisis of Knowing" (2025) ranks the threat of AI-amplified disinformation among the World Economic Forum's top global risks, noting the increasingly blurred line between AI- and human-generated content.
  • European Parliament: A 2025 European Parliament briefing documented that deepfakes shared online were projected to reach massive scale by 2025 (up from 500,000 in 2023), with Europol estimating that up to 90% of online content may be synthetically generated by 2026.
  • Deloitte: —Deloitte's 2024 Connected Consumer Study found that half of respondents said they are more skeptical of the accuracy and reliability of online information than a year ago, indicating trust erosion as a documented harm.
  • PMC research: A 2025 study in PMC ("AI-driven disinformation: policy recommendations") reviewed how AI-generated content floods platforms faster than humans can moderatemoderate, overwhelming existing fact-checking and verification systems.

5. Job Displacement in Knowledge Work

  • SSRN analysis: "AI Job Displacement Analysis (2025-2030)" (SSRN, Jun 2025) documents that AI job displacement is "not a future threat but a current reality," with over 76,000 positions already eliminated in 2025 alone.
  • JobMicrosoft displacementlayoffs: inCEO knowledgeSatya workNadella confirmed that 30% of Microsoft's code is now written by AI, while over 40% of Microsoft's May 2025 layoffs targeted software engineers — notdirectly speculative;linking happeningAI nowcapability inimprovements writing,to coding,workforce design,reduction and(Final analysisRound AI, 2025).
  • Psychological impact: A PMC-published study ("Psychological impacts of AI-induced job displacement among Indian IT professionals," 2025) documents the real human cost of AI-driven job losses, including anxiety, depression, and identity disruption among displaced workers.
  • AIMultiple analysis: "Top 20+ Predictions from Experts on AI Job Loss" (3 weeks ago) documents a sharp acceleration in AI-related job reductions through 2025–2026 as companies reorganized teams and redirected resources toward AI-focused priorities.

6. Centralization of powerPower

  • Brookings Institution: —"What onlyhappens when AI companies compete with their customers?" (2025) documents that the AI market is dominated by just a handful of companies can— affordAnthropic frontiercommanded models,40% concentratingmarket share, followed by OpenAI at 27% and Google at 21% — giving these few entities disproportionate influence over informationmodel infrastructurecapabilities, safety guardrails, and deployment conditions.
  • Bowdoin Science Journal: "Between Concentration and Decentralization: The Changing Landscape of AI Power" documents that OpenAI, Google DeepMind, Anthropic, and Meta make decisions about model capabilities and deployment conditions that affect billions of people worldwide.
  • Dario Amodei (Anthropic CEO): Public statements on AI centralization acknowledge the concentration of development power across a small number of companies, even as they debate open-source approaches.

The Real Regulatory Question

The Kehoe Rule's relevance to AI is not about sci-fi scenarios of rogue agents. It's about whether we let corporations deploy technologies with massive negative externalities — environmental damage, commons erosion, deskilling — and then ask society to prove the harm after the fact.

This is a genuine governance failure: developers who profit from rapid deployment should not be the primary arbiters of how much evidence of harm is acceptable before regulation kicks in.

Proposed Solutions: Preventive vs. Reactive Governance

Reactive Governance Preventive Governance
Waits for harm, then struggles to contain it Sets safeguards before widespread deployment
Post-facto regulation Requires independent testing and monitoring
Hard to control entrenched technology Makes developers report failures
Gives public institutions power to slow or stop a technology when warning signs appear

Key Takeaway

"The lesson is not that corporations are evil or that artificial intelligence must be stopped. It is that society should not confuse innovation with permission."

The real danger from AI isn't malevolent super-intelligence — it's the same pattern we've seen with lead, cigarettes, and asbestos: powerful commercial interests deploy technologies with significant negative externalities, delay regulation by manufacturing uncertainty, and let harm accumulate while profits flow. That pattern is well-documented and worth taking seriously, regardless of what one thinks about "AI agency."

Sources

Primary Article

  • Article:Ravenscroft, J. Artificial Intelligence and the Kehoe Rule (Substack)
  • Ezra Klein podcast episode with Helen Toner: "The A.I.s Are Already Out of Control"
  • Pacing the Frontier petition (July 2026): https://www.pacingthefrontier.com/

Historical Context

  • Lancet Commission on Pollution and Health (2022): 9 million premature deaths per year from pollution
  • Global Burden of Disease study (2025): Lead contributed to 3.5 million deaths in 2023

Commons Erosion

  • Mozilla Foundation (Feb 2024): "Training Data for the Price of a Sandwich: Common Crawl's Impact on Generative AI" — https://foundation.mozilla.org/en/research/library/generative-ai-training-data/common-crawl/
  • OECD (Feb 2025): "Intellectual Property Issues in Artificial Intelligence Trained on Scraped Data" — https://www.oecd.org/content/dam/oecd/en/publications/reports/2025/02/intellectual-property-issues-in-artificial-intelligence-trained-on-scraped-data_a07f010b/d5241a23-en.pdf
  • European Parliament (2025): "Generative AI and Copyright" — https://www.europarl.europa.eu/RegData/etudes/STUD/2025/774095/IUST_STU(2025)774095_EN.pdf
  • UK House of Lords Communications Committee (Mar 2026): "AI, copyright and the creative industries" — https://publications.parliament.uk/pa/ld5901/ldselect/ldcomm/267/267.pdf
  • US Copyright Office: "Identifying the Economic Implications of Artificial Intelligence for Copyright Policy"

Environmental Impact

  • Shehabi et al. (Dec 2025): "The carbon and water footprints of data centers and what this could mean for artificial intelligence," Patterns (Cell Press) — https://www.cell.com/patterns/fulltext/S2666-3899(25)00278-8
  • Lawrence Berkeley National Laboratory (Jan 2025): "2024 United States Data Center Energy Usage Report" — https://eta.lbl.gov/publications/2024-lbnl-data-center-energy-usage-report
  • United Nations University: "Environmental Cost of Artificial Intelligence: Carbon, Water, and Land Footprints" — https://unu.edu/inweh/collection/environmental-cost-of-AIs-Enrgy-Use-Carbon-water-and-land-footprints
  • Food & Water Watch (Feb 2026): "A No Brainer: How AI's Energy and Water Footprints..." — https://www.foodandwaterwatch.org/wp-content/uploads/2026/02/FSW_2602_AI_Water_Energy_UPDATE.pdf
  • Consumer Reports: "AI Data Centers: Big Tech's Impact on Electric Bills, Water, and More"

Psychological/Sociological Impacts

  • Fang et al. (2025): "How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Controlled Study," MIT Media Lab — https://www.media.mit.edu/publications/how-ai-and-human-behaviors-shape-psychosocial-effects-of-chatbot-use-a-longitudinal-controlled-study/
  • iScience (2026): "AI-overdependence and human cognitive decline: Hazards, evidence, and mitigation strategies" — https://www.sciencedirect.com/science/article/pii/S2451958826001764
  • Frontiers in Psychology (2026): "Not all cognitive offloading is equal: distinguishing dependent and autonomous offloading to generative AI" — https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2026.1878629/full
  • Alpha Psychiatry (2026): "Artificial Intelligence Chatbot Dependence, Anxiety, and Depressive Symptoms in University Students: Initial Psychometric Evaluation of the AICDS" — https://www.imrpress.com/journal/AP/27/4/10.31083/AP53298
  • Information (MDPI): "Unpacking AI Chatbot Dependency: A Dual-Path Model of Cognitive and Affective Mechanisms" — https://www.mdpi.com/2078-2489/16/12/1025
  • Kim et al. (2025): "From algorithm aversion to AI dependence: Deskilling, upskilling, and emerging addictions in the GenAI age"

Misinformation

  • UNESCO (2025): "Deepfakes and the crisis of knowing" — https://www.unesco.org/en/articles/deepfakes-and-crisis-knowing
  • European Parliament (2025): "Children and deepfakes" briefing — https://www.europarl.europa.eu/RegData/etudes/BRIE/2025/775855/EPRS_BRI(2025)775855_EN.pdf
  • Deloitte (2024): "Deepfake disruption: A cybersecurity-scale challenge" — https://www.deloitte.com/us/en/insights/industry/technology/technology-media-and-telecom-predictions/2025/gen-ai-trust-standards.html
  • PMC (Jul 2025): "AI-driven disinformation: policy recommendations" — https://pmc.ncbi.nlm.nih.gov/articles/PMC12351547/

Job Displacement

  • SSRN (Jun 2025): "AI Job Displacement Analysis (2025-2030)" — https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5316265
  • PMC (2025): "Psychological impacts of AI-induced job displacement among Indian IT professionals" — https://pmc.ncbi.nlm.nih.gov/articles/PMC12409910/

Centralization

  • Brookings Institution: "What happens when AI companies compete with their customers?" — https://www.brookings.edu/articles/what-happens-when-ai-companies-compete-with-their-customers/
  • Bowdoin Science Journal: "Between Concentration and Decentralization: The Changing Landscape of AI Power"
Back to top