In 1955 four researchers asked the Rockefeller Foundation to fund a summer project at Dartmouth College working on thinking machines. John McCarthy, a young mathematician there, needed a name for what they proposed to study, and wanted it kept clear of cybernetics, the neighbouring field it would otherwise have been folded into. He wrote down “artificial intelligence”. Seemingly unimpressed, the foundation gave them about half of the $14,000 they asked for. The phrase outlived all four of them.
The proposal ran to 17 pages on what the machines might be made to do, yet left the expression “artificial intelligence” itself undefined. Seven decades later the term AI is on every lip, written into laws from the US to China and across Europe and Asia, and generating commercial deals worth many times the Rockefeller Foundation’s endowment (which today stands at about $6bn). There is still no definition of AI in general use that everyone agrees on.
Strategy is a word that faces the same problem. It is used in every business plan and many job titles, yet very few people could agree on a definition, or even give you one. In the case of “strategy” this does not matter much, because very little relies on the definition itself; the job gets done one way or the other, and nobody is really regulating or monitoring the use of the word (although perhaps they should, given how abused the term is).
When it comes to artificial intelligence a great deal depends on the definition. Whether a system counts as AI decides which laws apply to it and what its supplier or developer has to say about it. Whether it counts as generative AI, or as an agent that makes its own decisions, decides what it is likely to get wrong and whether anyone would even notice. There is a clear tendency: the more abstract the term, the fewer people can define it, yet the more money and regulation rest on it.
In last week's issue the AI label was contested in Sweden, where it kept a hit single off the official chart; in this issue of The Compute we contest the label itself, taking five of these terms: human intelligence, artificial intelligence, generative AI, agentic AI and artificial general intelligence. For each one: we cover the definition, the way the word is actually used, and the practical difference between them.
One acronym, two words, three legal definitions
Human intelligence is the only one of the five with a clear definition, and it comes from a dictionary nearly two centuries old rather than an emerging technology lab. Merriam-Webster defines intelligence as the ability “to learn or understand things or to deal with new or difficult situations”.
In AI the word human is only ever used as an aspiration, or, dare we say it, a target, for machines: human-level performance, reasoning like a human, matching a human expert. None of those phrases says human at what. Every other term in this issue is measured against this one term, so a loose definition of it makes all of them looser.
On the other hand, artificial intelligence now has at least two legal definitions and they do not match. The European Union’s AI Act defines an AI system by its behaviour: a machine that “infers, from the input it receives, how to generate outputs” such as predictions, content, recommendations or decisions. South Korea’s AI law defines artificial intelligence as the electronic implementation of human intellectual abilities such as learning, reasoning, perception and judgement.
The two definitions ask a regulator to do different things. Under the European one, the regulator looks at what the system does: it takes an input and produces an output that was not written or specified in advance. Under the Korean one, the regulator has to decide which human abilities the system has reproduced, which is an opinion rather than an observation and much harder to agree on in a dispute. The Korean definition also covers more ground, so the same product can count as AI under Korean law and not under European law.
The United States federal government wrote a third definition in 2020: artificial intelligence is a machine-based system that, for a given set of human-defined objectives, makes predictions, recommendations or decisions influencing real or virtual environments.
In everyday use the AI label is attached to almost anything with a large language model (the software behind chatbots) or a rules engine inside it, and to plenty of things that contain neither. It shows up in product names, job titles, internal project names and board documents, applied to conventional automation and ordinary statistics as much as to the systems either piece of legislation would recognise.
The difference will start to be felt in 10 days: on August 2, 2026, the European Act begins requiring businesses to tell people when they are dealing with an AI system and to label machine generated content. Whether those regulations apply hinges entirely on whether a product meets the AI definition. Breaking the rule will result in fines of up to €15m or 3 per cent of a company’s worldwide turnover, with the Act’s heaviest penalties reaching €35m or 7 per cent for clearly banned practices.
The plausibility machine
Generative AI has a definition in the Korean law and none in the European Act. The Korean version describes a system that produces text, sound, images or video by imitating the structure and patterns of the material it was trained on. In simpler terms, it predicts what comes next, one piece at a time. Lately the phrase has become a synonym for AI as a whole, and for chat-based products in particular (such as Claude or ChatGPT).
A generative system does not look anything up. It produces the most plausible continuation of what it was given as input, and plausibility is the only thing it optimises for. Ask it for a citation and it will supply one with the right author, journal and year for a paper that does not exist, because a plausible citation is exactly what was requested. Fluency and confidence say nothing about truth, yet they are what most users judge an answer by. Whoever uses the output has to verify it, because the system that produced it never did, and cannot.
Left to its own devices
An AI agent is a system that chooses its own next step: it acts, looks at the result and decides what to do next. Neither the European Act nor the Korean law defines one; between them we have a lexicon of 79 defined terms, yet this one is not among them.
In ordinary use, the term “agentic” is applied to almost any process with several steps and possibly an AI model somewhere inside it, including processes where a person wrote every step in advance.
The difference comes down to when the sequence of steps gets decided: in ordinary automation before the work starts, in agentic AI while the work runs.
A fixed script does the same work every task, so the cost per task is known before it starts. An agent’s cost depends on how many steps it takes and how much material it has generated and is carrying along by then, neither of which is known in advance; the question moves from what the software subscription costs to what each run or iteration costs, and the real answer is a range… or rather “it depends”.
A scripted process that goes wrong usually stops, and visibly so, together with the associated notifications and error codes. An AI agent reads its own previous output as the input to its next step, so an early mistake does not stop the work; it becomes an input for the rest of the task to run on.
Scripts are approved once, because everything they will do is known at the moment of approval. An agent has to be allowed a class of actions in an order nobody can write down in advance, which is a level of delegation and empowerment most organisations have never had to design for.
A two-month study running for seven decades
Artificial general intelligence (AGI), a machine as capable as a human, is the end of the line (is it?), and there is very little to work with: no agreed definition, no agreed test, no consensus on whether current AI technologies and methods will lead us anywhere near it. It has started appearing in timelines, public statements and fundraising material, yet a claim built on the term cannot be checked against anything, it can only be believed… just like faith.
Somehow that has not kept AGI out of commercial and legal documents. Microsoft and OpenAI rebuilt their partnership around the term on October 28, 2025, with Microsoft’s exclusive rights and the revenue share between the two companies both expiring when AGI arrives. Unable to define the word, they agreed that once OpenAI declares it has arrived, an independent panel of experts will have to rule on whether it actually has.
McCarthy and his co-authors gave themselves a summer. Their proposal asked for “a 2 month, 10 man study of artificial intelligence”, on the assumption that every feature of intelligence could be described precisely enough for a machine to copy it. The description is still missing. The word is written into regulations and agreements now, and it is still doing the job it was given in 1955, which was to stand in for something nobody could yet define.
A note to readers
This week we announced Tokenando Junior, an initiative run by our high-school interns to take AI awareness and education beyond the boardroom and into the classroom. The first drop is a free online session, “AI Literacy for students”, taking place on Thursday July 30, 2026 at 18:00 Barcelona time; the registration form is open. More on the wider initiative in the coming weeks.
And one more thing: AI strategy did not make the list above, mostly because nobody can define it either; it is, however, a conversation we are always happy to have. Reply to this issue if you are interested in having a chat.
Wirelessly yours,
Ziad Matar
Co-founder, Tokenando
Editor-in-chief, The Compute
After Hours
The week’s discovery was Sons of the East, a Sydney indie-folk trio we found through “Inappropriate Behaviour”, and a day at the beach introduced the Cuban salsa of Rey Ruiz into the playlist rotation. The World Cup kept its grip on our screens right through Sunday’s final, with a loud celebration of Spain’s win and a little victory lap around the neighbourhood. With the football gone, The Bear has finally taken over the evenings, exactly as promised two issues ago, while The Odyssey is still waiting for us on the big screen… maybe this weekend.
References
“A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence”, 1955, reprinted in AI Magazine: https://dl.acm.org/doi/10.1609/aimag.v27i4.1904 (FLAGGED: may be behind a login wall, click-check before publishing)
“The Birth of Artificial Intelligence”, Dartmouth Alumni Magazine: https://dartmouthalumnimagazine.com/articles/birth-artificial-intelligence
The Rockefeller Foundation, endowment figure: https://www.rockefellerfoundation.org/about-us/rock-faqs/
EU AI Act, Article 3 (definitions), European Commission AI Act Service Desk: https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-3
EU AI Act, Article 50 (transparency obligations): https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50
EU AI Act, Article 99 (penalties): https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-99
15 U.S.C. 9401(3), National AI Initiative Act of 2020 (United States definition of artificial intelligence)
Framework Act on the Development of Artificial Intelligence and Establishment of Trust (South Korea), translation, Center for Security and Emerging Technology: https://cset.georgetown.edu/publication/south-korea-ai-law-2025/
Merriam-Webster, “intelligence”: https://www.merriam-webster.com/dictionary/intelligence
Microsoft, “The next chapter of the Microsoft–OpenAI partnership”, October 28, 2025: https://blogs.microsoft.com/blog/2025/10/28/the-next-chapter-of-the-microsoft-openai-partnership/
OpenAI, “The next chapter of the Microsoft–OpenAI partnership”: https://openai.com/index/next-chapter-of-microsoft-openai-partnership/
Music Business Worldwide on the Sverigetopplistan exclusion: https://www.musicbusinessworldwide.com/ai-created-track-blocked-from-swedens-official-charts-after-racking-up-millions-of-streams-on-spotify/
The Compute, Issue XI, “Human after all”: https://tokenandoai.substack.com/p/human-after-all