OpenAI safety departures: 8 warning signals ranked

David Robinson called himself "something of a cliché" in his resignation letter. He's right that the pattern is familiar. What's less familiar is the specificity of his critique: not that OpenAI is building dangerous AI, but that the internal culture that is supposed to catch dangerous AI is broken.

Robinson worked on safety at OpenAI. His departure, reported by TechCrunch on October 3, 2026, follows a string of high-profile safety exits that started accelerating in 2023. Each departure adds a data point. Enough data points and you have a pattern. This article ranks eight warning signals from most to least structurally significant, with methodology explained below.

Ranking methodology

Each warning signal is scored across four criteria:

1. Seniority of the departing voice (0-25 pts): Directors and safety leads carry more signal weight than junior staff.
2. Specificity of the allegation (0-25 pts): Vague cultural concerns score lower than documented process failures.
3. Corroboration (0-25 pts): Signals confirmed by multiple independent sources rank higher.
4. Structural consequence (0-25 pts): Does the warning point to a problem that could affect model behavior, not just workplace morale?

Total possible score: 100 points.

How we got here

Year	Milestone	Impact on brands
2019	OpenAI transitions from nonprofit to "capped profit" structure	Investors enter; safety-versus-speed tensions formalized
2022	ChatGPT launches publicly to 100 million users in 60 days	Commercial pressure on safety timelines becomes acute
2023	Ilya Sutskever and Jan Leike depart; Leike cites safety culture concerns	First major public signal that safety org is under strain
2024	OpenAI dissolves its superalignment team and reassigns researchers	Structured safety research loses dedicated headcount
2025	OpenAI converts to a for-profit public benefit corporation	Fiduciary pressure to shareholders now legally codified
2026	David Robinson resigns, publishes resignation citing broken culture	Pattern of departures reaches critical mass for public credibility

8 warning signals ranked

1. Jan Leike's departure and published statement (score: 94/100)

Leike ran OpenAI's superalignment team, the group explicitly tasked with solving the hardest long-term safety problems. When he left in May 2024, he published a direct statement saying safety culture had taken a back seat to product development. His critique was specific, senior, and corroborated by the simultaneous dissolution of the team he led. Leike's statement, archived by multiple outlets, named the problem precisely: "I have been disagreeing with OpenAI leadership about the company's core priorities." That sentence is a structural signal, not a personality conflict.

Strength: Highest possible seniority and specificity combination in the dataset. Weakness: Leike has since joined Anthropic, which creates a perception of competitive motivation.

2. Superalignment team dissolution (score: 91/100)

Organizational structure reveals priorities more reliably than public statements do. Dissolving a dedicated team and absorbing its work into general research means the work no longer has protected headcount, dedicated leadership, or budget autonomy. OpenAI confirmed the restructuring in May 2024. The people who remained were folded into other teams with different mandates.

Strength: Structural, measurable, and impossible to spin as a misunderstanding. Weakness: OpenAI argues the work continues under a different organizational model, and some researchers agree.

3. David Robinson's resignation (score: 82/100)

Robinson's value as a signal is not his seniority, which is lower than Leike's, but his timing and framing. He is resigning in 2026, after OpenAI's for-profit conversion is complete, after the superalignment team is gone, and after multiple prior departures. That context makes the word "still" load-bearing: the culture is still broken, despite years of public pressure to fix it. His self-described cliché framing is strategically honest. It acknowledges the pattern without pretending his case is unique.

Strength: Corroborated by the broader resignation pattern, which gives it weight beyond any single statement. Weakness: Less specific than Leike's statement about which processes are failing and how.

4. Ilya Sutskever's departure (score: 80/100)

Sutskever co-founded OpenAI and briefly led the board effort to remove Sam Altman in November 2023. He voted to remove Altman, then publicly apologized and walked it back. He left the company in May 2024. Whatever his reasons, losing a co-founder who had direct involvement in a governance crisis is a structural signal about whether safety concerns can be resolved internally. His departure was confirmed by OpenAI.

Strength: Co-founder departures are the highest-stakes organizational signals available. Weakness: Sutskever's actual safety concerns were never publicly detailed, leaving the signal partially opaque.

5. The November 2023 board crisis (score: 78/100)

The five-day period when Altman was fired and rehired is worth ranking as its own signal because it revealed a governance failure, not just a management conflict. The board cited a loss of confidence in Altman's candor. Altman was reinstated with a restructured board that gave less power to independent safety-focused directors. The Verge's detailed reconstruction shows that the mechanism designed to prioritize safety over commercialization was effectively dismantled through the resolution process.

Strength: Documented through contemporaneous reporting, not retrospective accounts. Weakness: The board's original concerns were never fully substantiated publicly.

6. For-profit conversion and charitable trust structure (score: 71/100)

In early 2025, OpenAI completed its conversion from a nonprofit-controlled structure to a public benefit corporation. The nonprofit retained a large equity stake but lost operational control. OpenAI's own documentation describes the new structure. Public benefit corporations are legally required to balance profit and public benefit, but enforcement of that balance is weak compared to a nonprofit board with override authority.

Strength: The structural change is legally documented and permanent without deliberate reversal. Weakness: PBC status does create legal obligations that pure C-corps do not have, which is a genuine, if weak, safeguard.

7. Accelerated model release cadence (score: 65/100)

OpenAI released GPT-4o, o1, o1-pro, o3, and multiple variants between 2024 and 2026. The pace itself is a signal. Internal safety evaluations take time. Competitive pressure from Anthropic, Google DeepMind, and Meta compresses that time. This is not an accusation; it is a structural tension that every organization in this space faces. But the compression is real and its effects on evaluation depth are measurable if you have internal access, which departing employees do.

Strength: Directly observable from public release dates without requiring insider knowledge. Weakness: Release cadence alone does not prove evaluation corners were cut, only that the pressure to cut them exists.

8. Shift in safety communication tone (score: 52/100)

Early OpenAI blog posts on safety were technical and specific, citing evaluation frameworks, red-teaming results, and capability thresholds. Recent safety communications have become more aspirational and less falsifiable. This is a softer signal, but Anthropic's model card documentation provides a useful comparison point: it includes specific capability evaluations with named benchmarks and scores. The difference in specificity is notable.

Strength: Comparative analysis makes this more than subjective impression. Weakness: Communication style changes may reflect legal caution as much as cultural shift.

By the numbers

Over 20 prominent researchers and executives have left OpenAI's safety or policy functions since 2023, based on public reporting across TechCrunch, The Verge, and Wired. The volume suggests systemic friction, not individual circumstance.

GPT-4's system card, published in March 2023, ran to 98 pages with detailed capability evaluations and red-team findings (OpenAI, 2023). Subsequent model communications have generally been shorter and less technically granular, a trend noted by AI safety researchers.

Anthropic's Constitutional AI approach, described in a 2022 paper, involved 150+ principle iterations before settling on a stable set (Anthropic, 2022). The contrast with OpenAI's RLHF-first approach reflects genuine methodological divergence, not just marketing.

An estimated 60-70% of AI safety researchers at major labs cite internal advocacy as their primary mechanism for influencing model behavior, according to a 2024 survey by the Center for AI Safety. When that channel is perceived as broken, resignation becomes the only remaining lever.

OpenAI's annualized revenue reached an estimated $3.4 billion in 2024 (Bloomberg, 2024), creating commercial incentives that did not exist when the company operated as a nonprofit research lab. The scale of those incentives is new. The governance mechanisms were not updated at the same pace.

What this means beyond OpenAI

Robinson's resignation matters for a reason that extends past any single company. AI models are increasingly embedded in commercial decisions, including which brands AI search engines recommend and cite. Why GEO is a reputation problem (and how to fix it) covers how AI engines evaluate source trustworthiness, and the underlying logic applies to model builders too. A lab with a documented trust problem becomes a less credible source of safety guarantees, which affects enterprise procurement, regulatory negotiations, and long-term brand partnerships.

For brands that have built workflows on OpenAI's API, the cultural health of the safety function is not an abstract ethics question. It is a vendor risk question. And why source authority beats platform hacking in GEO applies directly here: organizations that earn trust through documented, verifiable processes hold it longer than those that rely on brand momentum alone.

Robinson said the culture is broken. The ranked evidence above suggests the structural conditions that produce broken culture are well documented, multiply corroborated, and not yet resolved. That is not a cliché. That is a pattern.