The EU’s AI Act is in force, yet the Commission is still explaining what an “AI system” is. Your board duties are determined by a term everyone uses and no-one defines.
1. The EU AI Act in force: regulating nothing and everything
On 2 August 2026 the Act’s transparency duties commenced, and the Commission can now fine a provider of a general-purpose model the higher of 15 million euro and 3% of its total worldwide annual turnover1.
Every obligation in the Act hangs on one phrase, and Article 3(1) is the only attempt at defining it:
An “AI system” means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments.
Regulation (EU) 2024/1689, Article 3(1)
This “definition” substitutes a proxy — inference — which inherits the entire problem: nothing in the Regulation says how much inferring makes a system intelligent, or what level of autonomy.
This vacuum is now yours to stand in. Whether you’re part of a corporation, a government body, or a charity, if you use artificial intelligence you must first decide whether the Regulation reaches you at all. You must decide it with no acceptance criterion for what counts as intelligent. The Regulation’s own scope is therefore undefined. And as it is currently phrased, it regulates… nothing and everything.
The Commission has already had to explain itself. It published separate guidance on what an “AI system” is, and that guidance concedes that “no automatic determination or exhaustive lists of systems that either fall within or outside the definition of an AI system are possible”. It declines to draw a bright line, and by its own first pages it binds no-one2.
Whose question this is. What the Regulation requires of you is a question for your counsel. What artificial intelligence actually means is not a legal question at all, and counsel cannot answer it. That one is ours.
2. The Act inherited the problem; it did not create it
The problem is that we cannot yet characterize in general what kinds of computational procedures we want to call intelligent
John McCarthy, What is Artificial Intelligence? 2007
If it works, it isn’t AI
Edward Feigenbaum, 1988
John McCarthy, who coined the term “artificial intelligence”, admitted that a general definition of intelligence is very difficult, and lamented that “as soon as it works, no one calls it AI anymore”.3 AI as a discipline itself has never defined its own subject4. The textbooks use “intelligence” on every page and define it nowhere. Instead, textbooks take the view that artificial intelligence is defined as the collection of methods for solving “difficult” or “human-level” problems5 — and what counts as such has changed countless times since 1950. AI’s goalpost has moved with each problem solved: checkers, chess and Go among the board games, and Jeopardy! beyond them. Each time the machine appeared “more intelligent”, with no consensus on what that meant.
Seventy years have produced no shortage of answers. What they have never produced is a general definition. And now that LLMs exceed human performance on a range of domains increasing by the day, the “difficult” and “human-level” descriptors no longer apply.
3. Sufficient, not necessary conditions
But not just AI. For half a century, SETI has searched for extra-terrestrial intelligence — without having to define what it is: it searched for a narrowband radio signal such that “no natural process is known to produce”. That is an operational definition: it fixes not what the thing is, but what would count as evidence of it. Turing’s test makes the same manoeuvre: it defines what counts as evidence for intelligence — a sufficient condition. In short, each names how the capacity for intelligence can be demonstrated, and none says what is being satisfied.
Consider what each answer actually fixes: SETI fixes a context: a radio band, a signal no star produces, and a sender worth answering. Turing fixes another: a typed conversation, a fixed span of minutes, and an interrogator to be fooled. The IQ test fixes a third: one adult, literate and timed, working in the language the test was written in. Each is a sound test of something. The something is different in every case.
One commonality to all answers: They provide sufficient conditions. IQ tests, Turing’s test, SETI’s radio signals, and their many successors. Necessary conditions, however, have never been agreed, in millennia.
As a consequence, “intelligence” defaults back to psychology, an anthropocentric discipline, which treats it anthropocentrically, defines it as a single general factor g, measured using ‘quotient’ tests built exclusively for adult humans.
A measure built for adult humans cannot be used elsewhere. It cannot score a toddler or a musical genius, let alone an octopus, a bee colony, a swarm of drones, or Claude Opus 5. That is no defect in IQ tests. They set a sufficient condition in a narrow context and fail to provide a universal criterion.
So the argument was never about intelligence. It was about which context counts, conducted by people who were stating no context at all.
4. What Hilbert’s question actually settled
None of this is unprecedented. Mathematicians worked with algorithms for a thousand years and never defined one: an algorithm was an effective procedure, a mechanical method, a recipe a clerk could follow without insight — all of it intuitive, none of it precise. The gap went unnoticed until a question was asked that could not be answered without closing it. In 1928 Hilbert posed the Entscheidungsproblem: is there an algorithm that decides whether any given statement of first-order logic is provable? Nobody could say yes or no, because nobody could say what an algorithm was.
Hilbert’s question was answered twice over in 1936, by Alonzo Church in April and by Alan Turing in May, each working independently of the other. But they did not define effective computation — and never claimed to. They described what a human clerk can do with a pencil, a paper tape, and a finite set of rules, and they fixed those constraints as the criterion6. Turing, Church, Kleene, and Post each fixed their own — Turing machine, λ-definability, general recursiveness, and finitary combinatory processes were arrived at independently — and every answer is of the same kind: a procedure is effective if, and only if, it can be carried out by this. The four were later proved to pick out one and the same class of functions. This is our model of an answer.
The lesson is not that mathematicians have agreed about the nature of effective computation. The lesson is: you do not have to define what a thing is in order to settle what would count as an instance of it. Fix the acceptance criterion, and a question nobody could answer is settled.
5. Intelligent at what?
Here we get to the core of our argument: There is no such quantity as intelligence simpliciter. What an agent has is a capability displayed in a given context, a setting, and the context is not merely a detail of the measurement. It is an elementary part of what is being measured. Put plainly:
| Question | Answer |
|---|---|
| I(agent) → ℕ | no answer: there is no such quantity |
| I(agent, context) → ℕ | a number, and what every benchmark reports |
Let us be precise. Russell and Norvig supply the setting5: an agent acts in a context, and a context is an environment together with a goal. Examples:
- A chess game and winning it
- A round of Tetris and the lines cleared before the stack tops out
- A court hearing and advising the client
- A photograph of a skin lesion and a correct diagnosis
- Your company’s payment traffic and the fraud caught without blocking good customers
There is no such quantity as intelligence simpliciter
Name the context and “how intelligent?” has an answer. Without context, the question becomes meaningless: are fish more intelligent than birds? has no answer, for it depends entirely on whether the problem involves swimming or flying.
Every benchmark sets a context: The coding benchmark score7, the bar-examination score, the medical-licensing score, an IQ test: each fixes an environment and a goal, and each reports how one agent did inside it. That is a real measurement, and it is worth having. What it is not is a property of the system that travels with it out of that context.
6. What a board does on Monday
This is not pedantry. A contract, a standard, and a statute are definitions with teeth, and governance is made of them.
You cannot audit “safety” without a criterion for safe. You cannot safeguard “sensitive information” without pinning down that type of data. The same legislature managed it once already: nobody convenes a working group to decide whether a medical record is health data, because the GDPR settled the question. For “intelligent”, nothing has been settled.
The “Certification Theatre”. A system can pass every test put to it and meet no definition at all.
This is where it costs you money. Your vendor benchmarked in one context. You deploy in another. The score does not travel, and nothing in the certificate says so.
So put three questions to the supplier, and to your own people, about every system you are asked to approve. In which environment was this measured? Against which goal? By which measure, and who chose it? Leave any of the three unanswered, and you have been shown a number that means nothing where you intend to use it.
None of that is a theory of intelligence, and none of it needs to be. It is a filter, and a board can apply it in one meeting. Procure functions, not adjectives.
7. What this does not settle
Naming the context does not settle the philosophical question, and it is not the only criterion intelligence will need. Those arguments are centuries old, and they will outlast this Regulation. Your exposure, however, will not wait for them.
One lesson has held across decades of my own experience: in software engineering, in artificial intelligence, in governance, and in the philosophy of computer science. Every unresolvable argument arises from different people using one word for different things. A board cannot afford that luxury.
Ask your vendor this week, which context their benchmark was run in, and whether it is the one you deploy in. Tell me what came back.
Notes
- Regulation (EU) 2024/1689, Article 101(1), which empowers the Commission to fine a provider of a general-purpose AI model “not exceeding 3 % of their annual total worldwide turnover in the preceding financial year or EUR 15 000 000, whichever is higher”. Article 113 applies Chapter V from 2 August 2025 but excepts Article 101, which therefore takes effect on the Act’s general application date, 2 August 2026. The transparency obligations of Article 50 sit in Chapter IV and apply from the same date.
- European Commission, Commission Guidelines on the definition of an artificial intelligence system established by Regulation (EU) 2024/1689 (AI Act), C(2025) 5053 final, Brussels, 29 July 2025. The quoted sentence is paragraph 62; paragraph 7 reads “The Guidelines are not binding. Any authoritative interpretation of the AI Act may ultimately only be given by the Court of Justice of the European Union (CJEU).”
- Bertrand Meyer, “John McCarthy”, Communications of the ACM BLOG@CACM, 28 October 2011, who reports hearing McCarthy say it. This is the earliest documented instance, and it appeared days after McCarthy’s death; he published no such sentence. Widely circulated since through Nick Bostrom, Superintelligence: Paths, Dangers, Strategies, Oxford University Press, 2014, chapter 1.
- Stuart Russell, “Defining Intelligence”, Edge, 2 July 2017.
- Stuart Russell and Peter Norvig, Artificial Intelligence: A Modern Approach, 4th edition, Global Edition, Pearson, Harlow, 2021, ISBN 978-1-292-40113-3.
- A. M. Turing, “On Computable Numbers, with an Application to the Entscheidungsproblem”, Proceedings of the London Mathematical Society s2-42, received 1936 and published 1937, 230–265; the analysis of what a human computer can do is section 9. Identifying that class with effective calculability is the Church-Turing thesis, and it is not a theorem: B. Jack Copeland and Oron Shagrir, “The Church-Turing Thesis: Logical Limit or Breachable Barrier?”, Communications of the ACM 62(1), January 2019, 66–74.
- Artificial Analysis, “Coding agents”, accessed 30 August 2026.
Dr Amnon H. Eden is a computer scientist, the Principal Scientist at the Sapience.org thinktank, and founder of Moneta Sapiens Ltd.
Comments are welcome.
—» dysclaimer «—
