Buried in OpenAI’s corporate paperwork is a sentence that has shaped tens of billions of dollars of investment: if the company ever declares it has built artificial general intelligence, that system gets carved out of its commercial agreements with Microsoft. Not licensed. Not sold. Not handed over to the partner that has committed more than $13 billion.
That sentence is the load-bearing wall of OpenAI’s AGI policy. People argue about the safety frameworks, the charter’s promise to benefit “all of humanity,” and the boardroom explosions. The carve-out is the bit with legal teeth, and it explains a lot of otherwise strange behaviour: why the definition stays conveniently vague, why the declaration power sits with a nonprofit board, and why every restructuring announcement gets read as a signal about the timeline.
The two documents doing the real work
OpenAI’s AGI policy isn’t one artifact. It’s a stack of them, and two pieces carry most of the weight.
The charter and the AGI clause
The 2018 charter describes a mission and caps investor returns. The mechanic that matters sits further down: technology that qualifies as AGI is “excluded from IP licences and other commercial terms” with Microsoft, and those terms only cover pre-AGI systems. For seven years that was the most consequential undefined term in the industry.
Then came the late-2025 restructuring. The declaration power moved to the board of the OpenAI Foundation, with an independent expert panel involved in reviewing any such call. According to reporting on the deal, Microsoft’s licence now runs to 2032 and extends to systems OpenAI declares as AGI. The trigger didn’t get sharper. It got softer and pushed six years into the future.
The Preparedness Framework
Published in December 2023 and substantially rewritten in April 2025, this is the document that governs deployment decisions rather than corporate ones. It sorts catastrophic risk into tracked categories (chemical, biological, radiological and nuclear; cybersecurity; persuasion; and, added in 2025, model autonomy) and requires evaluations before a model ships.
The definition is itself a policy instrument
OpenAI has used the same fourteen words since 2018: “a highly autonomous system that outperforms humans at most economically valuable work.” The phrase is broad enough that almost any product could be argued either way, and narrow enough to exclude care work, craft and most scientific creativity, which don’t usually count as “economically valuable.” I’ve picked the wording apart elsewhere; the short version is that OpenAI’s AGI definition is doing a job that has nothing to do with philosophy.
Vagueness pays. A tightly specified definition becomes a contract term someone can litigate. A fuzzy one stays a press release.
How the risk scorecards work in practice
Four levels, four categories
Every tracked category gets a rating: Low, Medium, High or Critical. Low means no meaningful uplift to a bad actor. Critical means the model could plausibly do the work of a skilled expert on its own, whether that’s designing a viable pathogen, running an intrusion campaign end to end, or manipulating a person at scale. The rise of reasoning models and long-running agents is exactly what made the autonomy category necessary, a shift worth understanding in context: OpenAI’s AGI research on reasoning models and agents explains why single-turn evaluations stopped being sufficient.
What a “Critical” rating is supposed to trigger
On paper, a Critical finding halts development and deployment until safeguards bring the risk down to High or below. In practice, the April 2025 revision moved sign-off from the board of directors to company leadership and dropped the requirement that directors review safety reports before a model ships. It also promised public “Safeguards” reports, but only when a Critical capability is found. Silence is the default setting.
The deeper gap is that the framework offers scores without binding rules. Nothing public says “at capability X, action Y is mandatory, and here is who verifies it.” Google DeepMind’s Frontier Safety Framework has the same shape. Anthropic’s Responsible Scaling Policy is more explicit about pausing, yet it still rests on the company’s own judgement of when a threshold has been crossed.
Who actually decides
This is where the policy meets the org chart. A CEO who controls deployment timelines, a Safety Advisory Group that reports internally, and a board with commercial exposure to the outcome are not three independent checks. The staff departures of 2024 thinned out the internal sceptics who would have pushed back hardest on a launch decision: the superalignment team dissolved, Jan Leike went to Anthropic, Ilya Sutskever left to start his own lab.
The uncomfortable question isn’t whether OpenAI’s safety staff are sincere. Most of them plainly are. It’s whether a structure where the people who decide a model is safe to ship also profit from shipping it holds up under the pressure of a build-out measured in hundreds of billions. The five-rung infrastructure plan behind that build-out has no published schedule, which conveniently means no missed milestones.
Where the policy is genuinely thin
- Self-declared AGI. The company decides when the term applies. No regulator, no court, and until recently no third party had a formal role in checking.
- Undefined thresholds. Risk levels exist, but the mapping from measured capability to required consequence stays internal.
- Conditional disclosure. System cards have grown longer on capabilities and thinner on evaluation methodology, and Safeguards reports only appear when a finding is serious enough to demand one.
- Commercial gravity. Every quarter a frontier model stays in the lab costs revenue. Policy that depends on restraint is fighting a permanent headwind.
- Jurisdictional fog. A US company, a UK safety institute, EU AI Act obligations and voluntary commitments to all three. None of them bind the AGI declaration itself.
The other half of the policy: Washington
AGI policy at OpenAI is not only about evaluations. The company’s submissions to the 2025 US AI Action Plan argued for federal rules that preempt a patchwork of state laws, for fair use in training data, and for government-backed export of American models through procurement and diplomacy. That’s a commercial strategy wearing industrial-policy clothes, and it matters because the legal environment determines which safety commitments survive contact with a board meeting. Watch which proposals get adopted. The ones that raise compliance costs for smaller rivals tend to be the ones OpenAI pushes hardest.
Signals worth tracking over the next year
Set the AGI launch dates aside for a moment. They’ve moved repeatedly, and the pattern is instructive on its own, as this survey of every AGI timeline OpenAI has published makes clear. Four concrete things will tell you more than any keynote:
Whether the independent expert panel gets named, with real technical people on it rather than a letterhead. Whether a Safeguards report is ever published for a model that actually ships, as opposed to one that never leaves the lab. Whether the next system card includes evaluation methods or just results. And whether the fourteen-word AGI definition quietly changes in a filing, which is how policy usually shifts at OpenAI: not with an announcement, but with a revised sentence nobody reads until the lawyers do.

