← AI investigation
Act I · what is AI?

What does “AI” actually refer to, and to whom?

Before any conversation about whether AI is dangerous, valuable, sustainable, or fair, there is a prior question that almost never gets asked in public: what are we even pointing at when we say the word? Eighteen mainstream definitions, no two of them agree on the same set of artefacts.

TopicsRegulationEpistemics
TracksAI
Read in 30 seconds
Regulatory

Converges on "infers outputs"

→ EU AI Act · OECD · NIST · ISO

Broad coverage; differs on whether autonomy + adaptiveness are required

Academic

Russell & Norvig — "rational agent"

→ Even a thermostat technically counts

McCarthy 1955: "any feature of intelligence" · Minsky: "things that would require intelligence if done by men"

Critical

The boundary keeps moving

→ Tesler · Bender · Crawford

Chess, OCR, search ranking — all called AI when invented, demoted to "just computation" once they worked

Three families of definition. They disagree on whether a thermostat, a regression model, a fine-tuned classifier, and ChatGPT are all 'AI' — and the disagreement matters for what gets regulated, funded, criticized, or sold.

00 · The definition problem

Eighteen definitions. No two agree.

A regulatory disagreement on what counts as AI is not a semantic squabble — it determines what gets licensed, what gets banned, what gets reported, what counts as 'AI investment' on an earnings call. The boundary work is the policy work.

Three broad families of definition. The regulatory family (EU AI Act, OECD, NIST, ISO/IEC 22989) converges on a behavioural definition: a machine-based system that infers outputs from objectives. Even within the family they differ on whether autonomy and adaptiveness are gating criteria — the difference between “is a regression model AI?” being yes or no.

The academic family (Russell & Norvig, McCarthy, Minsky) is broader still — Russell & Norvig's “rational agent” technically includes a thermostat as a trivially rational agent (the canonical AIMA example). McCarthy's 1955 Dartmouth proposal is even broader: “any feature of intelligence” that can be simulated.

The critical / sociotechnical family (Bender, Gebru, Crawford, AI Now) flips the frame: AI is not a discrete technology but an assemblage — models + data + labor + institutional deployment + downstream effects. From this angle, “what counts as AI” is itself a power claim about which technical systems get the prestige and the scrutiny of the label.

01 · The disagreement, made visible

Same word. Seven artefacts. Ten definitions. No two columns agree.

The same loan-scoring regression is yes for NIST, partial for the EU, and no for the UK AI Security Institute. Below: the matrix that 'AI policy' is currently built on.

Do the definitions agree on what counts as AI?
Thermostatrule-based controllerRegression modelloan-scoring useBERT classifierfine-tuned, 110MWhisperspeech-to-textDiffusionimage generationLLM chatGPT-4o, ClaudeLLM reasoningo-class, thinkingDEFINITIONscope / framingEU AI Act, Art. 3(1)broad — but autonomy + adaptiveness required·~✓✓✓✓✓OECD (2023)broad — "infers… outputs"·~✓✓✓✓✓NIST AI RMF 1.0broad — no autonomy gate·✓✓✓✓✓✓ISO/IEC 22989:2022standards — engineering definition·✓✓✓✓✓✓China CAC (Aug 2023)narrow — generative services only···✓✓✓✓UK AI Security Institutenarrow — frontier models only····~✓✓Russell & Norvig (rational agent)broadest — anything acting toward best outcome~✓✓✓✓✓✓McCarthy / Dartmouth (1955)broadest — "any feature of intelligence"·✓✓✓✓✓✓OpenAI Charter (AGI)narrow — "outperform humans at most economically valuable work"·····~~Bender et al. — Stochastic Parrotscritical — form, not meaning···~✓✓✓✓ yes~ partial· noSame word. Seven artefacts. Ten definitions. No two columns agree.
A typical loan-scoring regression is yes for NIST and ISO, partial for the EU and OECD (depending on autonomy + adaptiveness gating), and no for the UK AI Security Institute (not frontier) and China's generative-AI rules (not generative). The visible disagreement is the lesson. “We don't do AI” and “we are an AI company” have no shared referent until a column is named.
The EU AI Act and the OECD share the same definitional spine, but the EU adds 'autonomy' and 'adaptiveness' as gating criteria — narrowing what counts in practice.
China's 2023 generative-AI rules cover only public-facing generative services. ChatGPT and Stable Diffusion are in scope; recommender systems and discriminative ML are not.
02 · The boundary keeps moving

Tesler's effect: AI is whatever hasn't been done yet.

Chess, OCR, route-planning, spam filtering, search ranking — all called AI when invented; all demoted to 'just computation' once reliable. The current set (LLMs, diffusion, reasoning) hasn't been demoted yet.

The AI effect — the boundary moves
STILL CALLED AI IN 2026DEMOTED — NOW “JUST COMPUTATION”19601970198019902000201020202030Theorem proving1956→1975Route planning1959→1985ELIZA1966→1985OCR1976→1995Expert systems1980→1992Chess engines1990→1998Deep Blue beats KasparovPageRank / search1998→2010Spam filtering2000→2008Recommendation algorithms2003→2012Statistical translation2006→2018Speech recognition2010→2018AlphaGo2016→2022Large language models2020Diffusion image generation2022Multimodal models2023Reasoning models2024Generative video2024Autonomous agents2025Today's AI is just whatever has not yet been demoted to “the algorithm.”
Larry Tesler: “intelligence is whatever machines haven't done yet.” Each item enters at its inception year above the line, then drops below at its demotion year — the moment it became reliable enough that the public stopped calling it “AI” and started calling it “just computation,” “the algorithm,” or “ML.” The items still above the line in 2026 (LLMs, diffusion, reasoning, agents) are simply the ones that haven't been demoted yet. The methodological consequence: the term “AI” has no stable referent. Any definition pinning to “what AI does today” will be obsolete by the next demotion cycle.
The pattern Larry Tesler named — 'intelligence is whatever machines haven't done yet' — has the documented effect of resetting the AI hype cycle every ~10–15 years.
The 'AI winters' of 1974–1980 and 1987–1993 followed exactly this dynamic — once techniques worked, the AI label was withdrawn and the funding evaporated.
03 · What's being called AI today

Seven things, all called AI.

The figure below is the seven-tier model spectrum from LeResearch's philosophy page. Most public critique is about row 1 (frontier closed LLMs). Most useful AI work happens in rows 4–6 (small open models, classical ML, geostatistics). The marketing label covers all of it.

Seven things, all called "AI"
INSTANCECOMPUTESCALE (RELATIVE)OPENNESSFrontier foundation LLM (GPT-5, Claude, Gemini Ultra)trained for months across thousands of GPUs; what the water critique is actually aboutdata-centerxlclosedMid-size open LLM (Llama 3, Mistral Large)weights released; training cost still massive; inference on a single workstationclusterlpartialVision / multimodal model (CLIP, image-gen)image training sets carry decades of representational bias, calcified in pixel spaceclusterlpartialPhysics-informed neural net (our PI-GNN)millions of params; trains in hours; runs in watts; physics regularizer keeps it honestlaptopsopenClassical ML (random forest, gradient boost)existed for decades; unambiguously “AI” in the taxonomic sense; nobody panicslaptopsopenKriging (geostatistics, 1960s)the workhorse half of computational hydrology depends on; “AI” by the same taxonomylaptopxsopenThe marketing category (“AI-powered”)vibes; may be any or none of the above; sold by the labellaptopxspartial“We don't do AI” has no referent until a row is named.Most critique of AI is a critique of row 1. Most of LeResearch's work happens in rows 4 – 6.
Every row is called AI. They differ on every property that would matter for a rejection or an endorsement: compute scale, training environment, openness, water footprint, relationship to the public record, whose hands are on the governance. Refusing to decompose the category lets a blanket veto pass as a considered position — and lets a blanket enthusiasm pass as the same. Part of what LeResearch exists to do is build the vocabulary that separates a mathematical method from the business model currently monetizing it.

The model spectrum and the definitional spectrum are the same problem viewed from two angles. A blanket veto on “AI” passes as a considered position only because the row being objected to is left unnamed. The same is true of blanket enthusiasm. Part of what LeResearch exists to do is build the vocabulary that separates the mathematical method from the business model currently monetizing it.

04 · The tensions worth tracking

The axes the public conversation routinely conflates.

Six axes of disagreement, six different fights pretending to be the same fight.

  1. 01

    Scope: broad vs narrow.

    EU AI Act covers any inferring system; China CAC covers only generative services; UK AISI covers only frontier models. The same announcement can be "AI policy" in three different senses.

  2. 02

    Criterion: behavioural vs cognitive.

    OECD and EU define AI by behaviour (it produces outputs). DeepMind and McCarthy define AI by cognition (it solves intelligence). The behavioural definitions are more enforceable; the cognitive ones are more aspirational.

  3. 03

    Frame: technical artifact vs sociotechnical assemblage.

    NIST/ISO treat AI as an engineered system to evaluate. Crawford and AI Now treat AI as the labor + data + deployment + harm chain. The same system gets different governance under each frame.

  4. 04

    Capability: pattern matching vs emergent intelligence.

    Bender ("stochastic parrots") and the OpenAI Charter ("highly autonomous systems that outperform humans") cannot both be right about the same models. Most policy is written as if both are.

  5. 05

    Time-stability: fixed engineering definition vs moving target.

    ISO 22989 wants a definition that holds for decades. Tesler's effect guarantees it won't. Any "future-proof" AI rule is making a bet against the AI effect.

  6. 06

    Regulation gating: autonomy + adaptiveness required vs outputs alone sufficient.

    EU AI Act requires both. NIST and ISO require neither. A regression model used to score loan applicants is in scope under one and out under the other.

See also
Upstream
Sibling
  • Act II · Environment →
Maintenance

Last verified against EU AI Act Article 3 + Commission Guidelines (Feb 2025), OECD AI definition (Nov 2023 revision), NIST AI RMF 1.0, ISO/IEC 22989:2022, CAC Interim Measures (Aug 2023), Russell & Norvig AIMA (4th ed), and Bender et al. — On the Dangers of Stochastic Parrots (FAccT 2021) on 2026-04-25.

Earliest expected staleness: EU AI Act enforcement guidance + new state/national definitions (rolling).

This investigation is treated as a living document. The claims marked stale in the receipt cards above are the ones to re-check first. How this is maintained →