Mission Drift in Frontier AI
On Monday, April 27, Microsoft and OpenAI rewrote the contract that has held them together since 2019. Most of the coverage focused on cloud arrangements and revenue percentages, which matter if you care about which cloud provider hosts which AI products, but something else was missing that’s worth noting: the retirement of a clause. The clause was the contractual machinery that connected OpenAI’s mission to its commercial life. Now it’s gone.
The AGI clause
In 2019, when OpenAI was still a nonprofit and Microsoft was about to invest a billion dollars, the two companies wrote an unusual provision into their contract. If OpenAI ever built AI that could outperform humans at most economically valuable work, Microsoft would lose access to it. The technical name for that kind of AI is artificial general intelligence, or AGI. The provision became known as the AGI clause.
The clause was meant to do something specific. OpenAI had been founded as a nonprofit with a mission to make sure transformative AI benefited humanity rather than any single company. Taking a billion dollars from Microsoft created an obvious tension with that mission. The clause was the contractual bridge between the two. Microsoft got commercial access up to the point where AI became transformative. After that, the mission was supposed to take over.
The bridge was already weaker than it looked. According to reporting from The Information, by 2023 the two companies had operationalized AGI in their contract as a system generating one hundred billion dollars in profits, a threshold OpenAI is nowhere near reaching. For the past three years, the clause has been a constraint that may never have operated in practice. But contingent constraints are still constraints. They shape how parties negotiate, what they assume about the future, and what they can be held to. On Monday, the contingent constraint became no constraint at all. AGI is no longer a trigger for anything in the agreement. It’s just a word now.
This is familiar territory for those in the social impact sector. There’s a name in the field for what happens when an institution’s stated mission and its actual operating logic come apart: mission drift, and it’s one of the oldest problems in our work.
A pattern within OpenAI
The clause didn’t get retired in isolation. Within OpenAI, it’s the third mission-era commitment that has been weakened in the past eighteen months.
The first was a hard ceiling on how much money investors could make from the company. When OpenAI took its first round of for-profit investment in 2019, it set a cap. Investors could earn up to a hundred times what they put in. Anything above that was supposed to flow back to the mission. That ceiling is gone. It was removed as part of OpenAI’s October 2025 corporate restructuring.
The second was nonprofit control over the for-profit business. The whole point of OpenAI’s original structure was that the nonprofit, with its commitment to public benefit, would oversee the commercial work. The October 2025 restructuring loosened that oversight. Critics including the OpenAI Files project have argued the nonprofit’s authority was meaningfully diluted. The mission language was preserved. The architecture that gave the mission force wasn’t.
Three commitments designed during the era when AGI was treated as a discrete future event have been retired or weakened in eighteen months. The AGI clause is gone. The profit cap is gone. The nonprofit’s grip on the for-profit has been loosened. That’s a pattern.
Readers from mission-driven organizations will recognize the shape of it. What’s unusual is what was being softened. The capped profit, the nonprofit oversight, the AGI clause: these weren’t standard commercial protections. They were specific commitments designed to address the specific problem of building something the founders themselves believed could be dangerous to humanity.
The other direction
Two weeks before the Microsoft and OpenAI announcement, Anthropic crossed a different threshold.
Anthropic is the company that makes Claude. It was founded by people who left OpenAI because they thought OpenAI wasn’t taking AI safety seriously enough. They built a different governance structure. A small body called the Long-Term Benefit Trust, made up of people whose job is to care about the mission rather than the stock price, picks members of the board. The Trust was designed in 2023 to phase in over time. It started with authority to pick a small number of directors. Eventually it would pick the majority.
That phase-in just completed. On April 14, the addition of one new board member pushed Trust-appointed directors into a majority for the first time.
It’s tempting to read this as the clean counterexample to the OpenAI story. Architecture used versus architecture abandoned. The temptation should be resisted.
The full text of the Trust Agreement hasn’t been published. Anthropic itself has acknowledged that stockholders can amend the Trust and its powers via supermajority, without trustee consent. The Trust didn’t exercise the board authority it already had until this month, despite having held that authority since 2024. The threshold that just got crossed is real. It’s also the first time the structure has actually been used. A governance instrument that hasn’t been tested isn’t yet evidence of constraint. It’s evidence of a possibility.
What we can say is that Anthropic has moved board control under its own published rules rather than around them. Whether the rules constrain anything the company wouldn’t have done anyway is a question the next twelve months will start to answer. The OpenAI move is a clear retirement of architecture. The Anthropic move is, for now, an activation that hasn’t yet been load-tested.
Two of the most consequential AI labs are responding to the same moment with opposite contractual gestures. Whether either gesture produces accountability is itself part of the story.
What contracts do that speeches don’t
Even as the contractual commitments have been weakened, public rhetoric about AI risk is at least as loud as it has ever been.
Anthropic’s CEO spent January at Davos warning that artificial general intelligence could cause civilization-level harm. He has been publishing essays predicting that powerful AI will arrive within the next two years. OpenAI’s leadership has been making similar predictions, more cautiously phrased but still extraordinary by any normal standard. The story that AI is heading toward a transformative event, one that will require new kinds of governance, is alive and growing.
Contracts and speeches do different work.
Contracts bind. Speeches don’t.
When OpenAI’s mission commitments lived in contractual machinery, they had teeth, or at least the appearance of teeth. The profit cap limited what investors could earn. The nonprofit board had real authority over the company’s leadership, and exercised it briefly in November 2023 before being reversed within days. The AGI clause threatened to constrain Microsoft’s access if AGI was declared, even after the 2023 profit-threshold operationalization rendered that constraint mostly theoretical. As those commitments have been moved out of contracts and into mission statements, blog posts, and CEO essays, they have lost even the contingent force they had.
Mission drift is what happens when an institution’s revenue or funding pressures begin to compromise its founding commitments. The deeper problem is older. Chris Argyris and Donald Schön distinguished between an organization’s espoused theory, the logic it communicates to the world, and its theory-in-use, the actual map driving decisions. Organizations fail to learn, they argued, when they ignore the gap between the two.
A parallel distinction lives in evaluation practice. A theory of change describes how change actually happens: what produces what, under which conditions, with what contributions from outside the organization. A theory of action describes only what the organization does. Most organizations think they have the first. What they have is the second, mislabeled. The mission language names outcomes the work is supposed to produce. The architecture is what produces them, or fails to. Without architecture, the mission language is a theory of action claiming credit it cannot earn.
The OpenAI case shows what happens when the architecture and the language come apart visibly. The capped profit, the nonprofit oversight, the AGI clause: these were the architecture. They were the mechanism by which the espoused theory was supposed to discipline the theory-in-use, but that architecture has been weakened. The mission language hasn't changed. The same words mean different things depending on whether they sit in the contract or in the press release. What used to be binding is now positioning.
The Google Approach
Google DeepMind offers a third pattern. When Google acquired DeepMind in 2014, the founders insisted on an AI ethics board as a condition of sale. The board was discussed but never properly stood up. What Google DeepMind has instead is the Frontier Safety Framework, an AGI Safety Council, a set of AI Principles. These are internal policy documents, not contractual constraints. They live entirely on the espoused side of the Argyris and Schön distinction. There's no architecture to retire because none was built.
What’s left
The mission commitments most worth taking seriously are the ones that have survived a real attempt to remove them.
OpenAI’s didn’t. The architecture is gone. The mission paragraph remains. Anthropic’s haven’t been tested. The Trust has authority on paper, but hasn’t yet been pressed against a decision the company would have made differently without it. Until it is, the structure is a hypothesis about constraint, not a demonstration of one.
The thing that gets quietly retired is rarely the language. The language stays. What gets retired is the machinery that made the language do anything.
For anyone in the social impact sector, the question to ask about a mission, your own or someone else’s, isn’t whether the language is good. The question is what architecture supports it, what that architecture binds, and what happens when someone with power tries to remove it.
That’s the part to watch.
Anthralytic helps mission-driven organizations clarify and amplify their impact. We work at the intersection of evaluation, AI governance, and the systems that shape both.
