OpenAI

Stated Theory of Change

OpenAI's current stated theory of change (reconstructed from mission statements, public positions, and Altman's interviews): Build the most capable AI systems in the world, deploy them widely to generate revenue, use that revenue to fund safety research and infrastructure, and ensure the resulting AGI "benefits all of humanity." The mechanism is that being at the frontier of capabilities gives OpenAI the best position to understand and mitigate risks. Safety comes from being at the leading edge, not from restraint.

The 2018 Charter articulated a more specific theory: "ensure that artificial general intelligence benefits all of humanity" through a nonprofit governance structure with profit caps, a commitment to assist safety-conscious competitors, and an explicit subordination of commercial interests to the charitable mission.

Revealed Theory of Change

OpenAI's actions reveal a different theory of change: Build the most capable AI systems as fast as possible, deploy them to maximize revenue growth (needed to justify an $840B valuation and $1.4T in infrastructure commitments), and treat safety as a cost center that must not meaningfully slow deployment. The governance structures designed to enforce safety-first behavior have been systematically weakened or eliminated whenever they conflicted with commercial interests.

Evidence for this reading:

1. Safety leadership has departed, safety teams have been dissolved. Nearly every senior safety leader left within 18 months, with public statements about safety being deprioritized. The Superalignment and AGI Readiness teams were both dissolved. No comparable departures from the capabilities side.

2. Governance has been restructured to reduce safety constraints. Nonprofit control -> capped profit -> caps increased -> caps eliminated -> PBC. Each step was driven by investor requirements, not mission clarity. The board crisis proved that the governance mechanism cannot constrain the CEO.

3. Safety commitments have been weakened. The Preparedness Framework dropped persuasion tracking and added a "race to the bottom" exception. Safety reports are skipped or compressed. The 20% compute pledge was never fulfilled.

4. Military contracts contradict the original charter. The terms of service banned military use until January 2024. The DoD contract announced in February 2026 was taken immediately after Anthropic was punished for maintaining red lines on surveillance and autonomous weapons.

5. Financial incentives overwhelm all other considerations. At $840B valuation with a $5B annual loss, the pressure to grow revenue is existential for the company. Every major financial relationship (Microsoft $250B Azure commitment, SoftBank/Amazon/Nvidia investments, subscriber growth) creates pressure against safety-motivated deployment delays.

The gap between stated and revealed theory of change is the largest of any major AI lab. Anthropic was founded explicitly by people who left OpenAI because of this gap.

Key Assumptions

For OpenAI's stated theory of change to work, several things must be true:

1. Being at the frontier of capabilities is necessary for doing safety work. Evidence for: you need to study capable systems to align them. Evidence against: many important safety techniques (mechanistic interpretability, RLHF, constitutional AI) were developed by people and organizations that were not building the frontier model. Anthropic does alignment research at the frontier without the same scale of commercial deployment pressure.

2. The PBC structure provides meaningful mission protection. Evidence for: the Foundation appoints all PBC board members. Evidence against: zero reported cases of PBC shareholders successfully enforcing the public benefit obligation in court. The board that fired Altman for mission reasons was destroyed in 5 days.

3. Revenue pressure does not compromise safety decisions. Evidence for: OpenAI still does some safety testing and collaborates with external evaluators. Evidence against: Jan Leike's testimony that safety was "sailing against the wind" and couldn't get compute; skipped safety reports; compressed testing timelines; dropped risk categories.

4. Altman is a trustworthy steward of the mission. Evidence for: he co-founded OpenAI with safety concerns as a stated motivation; he has apologized when caught in specific misrepresentations. Evidence against: multiple former collaborators (Murati, Sutskever, Toner, multiple other executives) describe systematic dishonesty; equity clawback NDAs were signed by Altman then denied; board was not informed of ChatGPT launch.

Each of these assumptions faces serious counter-evidence. The most critical is #3, because the financial structure now makes it nearly impossible for safety to override commercial imperatives without external regulatory force.

Strengths

1. Talent and resources at unmatched scale. OpenAI employs many of the world's best AI researchers, has access to more compute than any other lab, and has financial resources that dwarf all nonprofit safety organizations combined. If safety work requires frontier capabilities, OpenAI is uniquely positioned.

2. Consumer distribution. With 500M+ users and 20M+ subscribers, OpenAI has the largest real-world deployment of AI systems. This provides unmatched data about how AI systems interact with humans at scale.

3. Some genuine safety research output. The deliberative alignment work, the Apollo Research collaboration on anti-scheming, the Preparedness Framework (even in weakened form) -- these represent real technical contributions to safety. Not zero effort; just dramatically insufficient relative to the capabilities being deployed.

4. Public scrutiny. OpenAI attracts more external scrutiny than any other AI lab. This external pressure (from regulators, media, competitors, the safety community) acts as a partial substitute for weakened internal governance. The California AG investigation is an example.

5. Revenue model that could fund safety. If OpenAI reaches profitability, it would have resources to invest in safety that are orders of magnitude larger than any grant-funded organization. The question is whether the incentive to do so will survive.

Weaknesses and Risks

1. Financial structure creates irresistible pressure toward speed over safety. The $840B valuation, $1.4T in infrastructure commitments, and $5B annual losses create a situation where any significant safety-motivated slowdown threatens the company's financial viability. This is not a resolvable tension -- it is structural.

2. Governance mechanisms have been destroyed and replaced with weaker ones. The board crisis proved that internal governance cannot constrain the CEO. The PBC structure is legally untested. The Foundation's 26% stake gives it formal power but no demonstrated willingness to exercise it.

3. Safety leadership has departed en masse. The people most qualified to judge whether OpenAI takes safety seriously -- the ones who were there -- left and publicly explained why. No amount of policy language compensates for the loss of Sutskever, Leike, Schulman, Brundage, Kokotajlo, and Murati.

4. Military contract crosses stated red lines. Taking a DoD contract immediately after a competitor was punished for maintaining red lines on surveillance, while refusing to release the contract text, is the clearest available evidence that commercial opportunity overrides stated values.

5. Opacity about safety investment. No safety budget, no safety headcount, no accounting for the superalignment compute promise. An organization genuinely committed to safety transparency would disclose these.

6. Race dynamics. OpenAI's competitive posture (Stargate, aggressive deployment, opposing state regulation, the Preparedness Framework's "competitor exception") suggests it views the AI race as one it must win, not one it should slow.

Cross-References

Anthropic: Founded in 2020-2021 by 11 OpenAI employees who left specifically because of the dynamics described in this analysis. Anthropic's refusal of the DoD surveillance/weapons clause -- which led to it being labeled a "supply-chain risk" -- provides a direct comparative test of how safety commitments hold under government pressure. Anthropic took the cost; OpenAI took the contract. The FLI Safety Index gives them similar grades (Anthropic C+, OpenAI C+), but the military contract represents a practical divergence not captured in the index.

DeepMind/Google: Operates under Google's corporate umbrella, which provides massive resources but also creates similar commercial pressure. FLI grade: C. DeepMind has avoided the degree of public controversy around safety departures that OpenAI has experienced.

Safe Superintelligence Inc. (SSI): Founded by Ilya Sutskever specifically as an alternative to the OpenAI model -- a company focused exclusively on safety, without the consumer product revenue pressure. SSI's existence is itself a critique of OpenAI's approach.

MIRI, ARC, Redwood: Safety-focused research organizations that operate at a fraction of OpenAI's resources. Their existence represents the theory that safety research can be done independently of frontier capabilities work.

What Would Change This Assessment

1. OpenAI publishes detailed safety budget and headcount data showing safety investment growing proportionally with capabilities spending. This is the single most falsifiable claim: show us the numbers.

2. The DoD contract text is released and independent legal review confirms it contains meaningful, enforceable prohibitions on surveillance and autonomous weapons use.

3. OpenAI declines a commercially valuable deployment on safety grounds, in a way that is verifiable and costly. The military contract was the most obvious test, and OpenAI failed it by the standards of its original charter.

4. A senior safety leader joins OpenAI from the outside -- someone with credibility in the safety community who would only join if they were satisfied with the internal culture. Currently, the talent flow is entirely in the other direction.

5. The Foundation exercises its governance authority in a visible, costly way -- for example, blocking a deployment or requiring additional safety testing against management's wishes.

6. External regulation creates binding constraints that substitute for internal governance failures. This would change the assessment not of OpenAI's internal dynamics but of whether those dynamics matter.

Self-Critique

What sources should I have checked but did not:

• OpenAI's own blog posts and safety page (inaccessible due to 403 errors)
• Preparedness Framework v2 full text (PDF not fetched)
• GPT-4o and GPT-5 system cards
• Altman's personal blog (blog.samaltman.com)
• The Atlantic's "Empire of AI" excerpt (The Optimist by Keach Hagey) -- referenced in OpenAI Files but not directly fetched

Where is this analysis potentially biased:

• The evidence base is heavily weighted toward critical perspectives. The most candid sources (Toner, Leike, Intercept) are critical; the defense (Holter/Theo) is less detailed. OpenAI's own website being inaccessible means I could not read their current self-presentation.
• There may be significant safety work happening internally that is not visible because (a) it is not published yet, (b) current employees are under NDA, or (c) the most sensitive safety work is necessarily confidential.
• The framing "safety departures prove safety is deprioritized" could be overstated if the departing individuals had idiosyncratic reasons or if replacement hires are doing equivalent work under less public profiles.

What would a thoughtful person who disagrees say:

• "OpenAI is doing more safety work than any other lab in absolute terms, and the departures reflect normal attrition at a fast-growing company, not a safety crisis."
• "The restructuring to PBC was necessary to compete and ultimately puts more resources into safety than a nonprofit ever could."
• "Altman's management style is aggressive but effective, and the board crisis was caused by a dysfunctional board, not a dysfunctional CEO."
• "The military contract includes more safeguards than previous government AI deployments and represents responsible engagement rather than reckless commercialization."

Single weakest claim: The claim that OpenAI's safety effort is fundamentally inadequate relies heavily on departure testimony and the absence of disclosed safety metrics. It is possible that internal safety work is more substantial than the external evidence suggests, and that the departures reflect personality conflicts and organizational growing pains rather than a genuine safety crisis. However, when multiple senior safety leaders independently describe the same pattern and publicly stake their reputations on it, the burden of proof shifts to OpenAI to demonstrate otherwise -- and they have not done so.

What information would most change my view: Detailed, verifiable data on safety spending, safety headcount, and safety team authority within OpenAI's decision-making process. If OpenAI published these and they showed safety investment growing proportionally with capabilities spending, it would substantially alter the assessment.