Three letters, no definition. AGI has become the most consequential acronym in technology, and nobody can tell you what it means. Not the researchers building toward it, not the companies staking billions on it, not the policymakers trying to regulate it. The term floats through earnings calls, congressional hearings, and arXiv papers with the confidence of something settled. It is not settled. It is not close to settled.

OpenAI defines AGI as "a highly autonomous system that outperforms humans at most economically valuable work." Their partnership agreement with Microsoft reportedly defines it differently: AI systems generating at least $100 billion in profits. One definition is about capability. The other is about revenue. They use the same two words. Sam Altman has called AGI "a very sloppy term," which is a strange thing to say about the stated mission of your company.

Google DeepMind took the most serious shot at resolving this in late 2023, when Meredith Ringel Morris, Shane Legg, and colleagues published a taxonomy surveying nine existing AGI definitions and finding all of them inadequate. Their proposed replacement is a matrix: five performance levels (Emerging through Superhuman) crossed with breadth of generality. Under this framework, current large language models qualify as "Level 1 Emerging AGI." Which tells you more about the framework than about the models.

Dario Amodei at Anthropic rejects the term entirely. He has called AGI "a marketing term" and prefers "powerful AI," which he defines as AI smarter than a Nobel Prize winner across most relevant fields, capable of running autonomously for days. That is a definition with teeth. It is also nothing like the other two.

So we have the three leading AI labs working toward something they cannot collectively name. This is not a minor semantic quibble. Definitions determine timelines, shape investment decisions, trigger contractual clauses, and inform regulation. When someone says AGI is two years away and someone else says it is twenty, they are frequently not disagreeing about progress. They are disagreeing about the destination.

The pattern has a name. Larry Tesler identified it in 1979: "Intelligence is whatever machines have not done yet." Every time AI clears a bar previously considered definitive, the bar moves. The ARC-AGI benchmark went from 0% in 2023 to 85% by December 2024. The response was not celebration but harder benchmarks. GPT-4.5 passed the Turing test in 2025 and it barely made the news. Coding tasks that would have seemed impossible to most researchers five years ago are now routine. The finish line retreats at the speed of approach.

In December 2025, this tension went public in the most entertaining way possible. Yann LeCun declared on a podcast that "there is no such thing as general intelligence" and called predictions of near-term AGI "completely delusional." Within hours, Demis Hassabis fired back, accusing LeCun of confusing general intelligence with universal intelligence. These are arguably the two most qualified people alive to have this argument, and they cannot agree on whether the concept itself makes sense.

Michael Timothy Bennett captured the frustration in an academic paper titled, bluntly, "What the F*ck Is Artificial General Intelligence?". His survey of AGI definitions found them varying on scope, metrics, feasibility assumptions, and whether human parity is even the right target. His conclusion: discussions about AGI risks, timelines, and policy rest on fundamentally incompatible premises.

I think the $100 billion definition is the most revealing one. Not because it is good, but because it is honest. It exists because AGI triggers a contractual clause: if OpenAI achieves it, Microsoft loses access to certain technology. The definition has nothing to do with cognition or capability. It is a legal instrument wearing a lab coat. And yet it governs the most consequential AI partnership in the world. That a financial threshold can sit alongside Turing tests and capability benchmarks under the same label tells you everything about how degraded the term has become.

There is a version of this argument that says none of it matters, that the capabilities are real regardless of what we call them. I have some sympathy for that position. The models are genuinely useful. They write code, summarise research, generate images that would have taken a studio two weeks to produce. Whether that constitutes "general intelligence" is, in some practical sense, beside the point for anyone using the tools today. But the label is not beside the point for the people setting expectations, raising capital, and writing legislation. When a company says it is building AGI, it is making a claim. When that claim has no stable referent, it cannot be falsified. And a target that cannot be falsified is not an engineering goal. It is marketing.

AGI is the only engineering target where the people building it, funding it, and regulating it cannot agree on what it is. We would not accept this in any other domain. Imagine a pharmaceutical company announcing it had cured cancer, but defining cancer as whatever diseases its drug happened to treat. The FDA would have questions. AI has no equivalent authority, no shared specification, no acceptance criteria. It has a phrase that means different things in different rooms and adapts to suit whoever is speaking.

Maybe that is the point. Maybe a fuzzy target serves everyone just well enough: researchers get funding, companies get valuations, politicians get something to regulate, and the public gets a story about the future. The ambiguity is not a bug. It is the product.

Sources: