Ask a dozen AI researchers to define artificial general intelligence and you’ll get a dozen answers, most of them reasonable. Ask OpenAI and you get something stranger: a definition that reads like a mission statement but behaves like a clause in a contract.
The company’s core wording has barely moved since 2018 — “highly autonomous systems that outperform humans at most economically valuable work.” Fourteen words, now embedded in a corporate charter, a partnership agreement with Microsoft and a public safety framework. That’s an enormous amount of legal and commercial weight resting on a single sentence.
Here’s what the definition says, what it quietly avoids, and why the exact phrasing keeps showing up in arguments about money, regulation and deadlines.
The sentence at the centre of it
OpenAI’s charter, published in 2018, states that the organisation’s mission is to ensure that AGI — which it defines as “highly autonomous systems that outperform humans at most economically valuable work” — benefits all of humanity. Every later document, from the capped-profit structure to the Preparedness Framework, refers back to that phrase.
Break it into three pieces and the ambition becomes clearer, along with the gaps.
“Highly autonomous”
This rules out the assistant sitting in a chat window waiting for your next prompt. An autonomous system sets sub-goals, runs for hours or days, recovers from its own mistakes and doesn’t need a human babysitting each step. Agentic tool use, long-horizon planning and self-correction all sit under this clause.
“Outperform humans”
Not match. Outperform. The bar is a system that does a task better than a competent professional, not one that merely produces something passable. That’s a meaningful distinction, because a lot of today’s models already beat the median human on narrow benchmarks while remaining obviously unreliable on open-ended work.
“At most economically valuable work”
This is the load-bearing phrase, and the vaguest one. “Most” is doing a lot of work. Fifty-one percent of economically valuable tasks? Ninety percent? Weighted by wages, by hours, by GDP contribution? The charter doesn’t say, and no OpenAI document since has pinned it down with a number.
Where the definition turns into a contract term
In 2019, OpenAI licensed its technology to Microsoft. The deal included an unusual provision: Microsoft’s access applied to technology short of AGI, and the determination of when AGI had been reached rested with OpenAI’s board rather than with either company’s engineers. Reporting has since put a figure on the trigger — a system capable of generating roughly $100 billion in profit — though neither party has published the exact language.
This is why the definition is more than marketing. The moment a model crosses the line, revenue-sharing terms, licensing rights and governance obligations all shift. The AGI definition in OpenAI’s charter is doing more work than most people realise, quietly shaping commercial relationships rather than just describing a research milestone.
Worth noting: after GPT-4 launched in 2023, Microsoft’s own researchers reportedly concluded the model was nowhere near the threshold. OpenAI’s leadership said much the same in public. A definition nobody wants to invoke turns out to be a very comfortable definition to hold.
Inside OpenAI, the term is a moving target
Sam Altman has called AGI “a very fuzzy term” and suggested the phrase has become less useful as it has grown more popular. In July 2024, OpenAI published a five-level framework — conversational, reasoning, agents, innovators, organisations — explicitly as an internal progress tracker rather than a definition of AGI itself. The company placed its own systems somewhere around level two.
That internal ladder is more revealing than the charter. It treats capability as a continuum with named rungs, while the public definition remains a single binary switch. Employees work against the ladder. Contracts and press releases hinge on the switch.
What the definition deliberately leaves out
The charter’s phrasing is narrow on purpose. Compare it with the questions people usually mean when they say “AGI” and the omissions stand out:
- Consciousness or experience. Nothing about sentience, feelings or moral status. A system could satisfy OpenAI’s definition while being, by any philosophical standard, entirely non-conscious.
- Physical capability. The wording is about work, not bodies. A model that never touches a robot arm can still qualify.
- Transfer to new domains. “Most economically valuable work” can be read as breadth across tasks, or as depth on a large but finite list. Those are very different claims.
- Reliability. Outperforming a human on average is not the same as outperforming a human every time, and the gap between those two things is where most real-world deployment problems live.
- Safety. The definition describes capability, not control. Whether the system can be steered, audited or shut down is handled elsewhere — in the Preparedness Framework and the safety protocols, not in the charter sentence.
Why the wording spills into policy
Governments need a threshold to legislate against. If a regulator wants to require pre-deployment testing for “AGI-level” systems, someone has to decide which models qualify, and drafters in Brussels, Washington and London have all borrowed language from the labs. OpenAI’s phrasing has become one of the reference points by default, largely because it was written down first.
The problem is timing. Rules are being written against a definition that its own authors treat as provisional, and the pace of AGI policy news is already outrunning the models it’s meant to govern. A definition flexible enough to survive five years of capability jumps is also flexible enough to be argued about in a courtroom.
How other labs frame it differently
Google DeepMind has leaned toward “general performance across a wide range of non-physical tasks,” with published level definitions that emphasise breadth rather than economic output. Anthropic tends to talk about frontier capability and safety cases rather than a single AGI moment. Academic surveys from the early 2020s often used expert-forecast medians, which produced thresholds ranging from human-level performance on all cognitive tasks to outperforming the median human worker.
None of these definitions contradict each other outright. They simply draw the line in different places, which is exactly why two people can look at the same model and disagree about whether the milestone has been reached.
The same three letters, deadlines 70 years apart
Swap definitions and the forecasts move dramatically. Narrow the goal to “beats experts on most remote knowledge work” and credible estimates cluster in the late 2020s. Widen it to include genuine generalisation, embodied competence and reliable long-horizon autonomy, and the dates slide toward the 2040s or later. The assumptions behind the 2027 and 2100 forecasts are rarely about compute curves — they’re almost always about which definition the forecaster had in mind.
Watch how the dates drift, too. Expert AGI timelines keep sliding for predictable reasons: benchmarks get saturated, definitions get quietly revised, and each generation of models reveals a new category of task that turns out to be harder than it looked. The 2026 AGI prediction round-up is a decent stress test, since several of the loudest forecasts from 2023 are already due for a rewrite.
What actually happens on the day OpenAI says it’s arrived
Practically speaking, the milestone won’t be announced by a benchmark score. It will arrive as a governance event. OpenAI’s board — advised by outside experts under the charter — makes the determination. Once it does, the Microsoft revenue-sharing arrangement renegotiates, the capped-profit structure converts toward its non-profit mission, and the safety team publishes the evaluations that justified the call.
Which means the useful question isn’t “is this AGI?” It’s “who gets to decide, using which evaluations, and what changes the next morning?” Read the charter sentence as a contractual trigger rather than a philosophical claim and the whole debate gets sharper: the interesting details are in the asterisks, the board’s discretion and the words “most” and “economically valuable” left undefined on purpose.