Sam Altman returns to OpenAI board after external review

On Friday, 8 March 2024, OpenAI announced that Sam Altman would return to its board of directors. The decision followed a review by WilmerHale of his removal as chief executive on 17 November 2023 and subsequent rehiring. The board expressed confidence in Altman and Greg Brockman continuing to lead the organisation.

OpenAI's publication about the review says the law firm conducted dozens of interviews, examined more than 30,000 documents and evaluated several corporate actions. It also says the removal did not arise from concerns about product safety, the pace of development, finances or communications to investors, customers or partners.

One qualification is essential: the public received a summary of findings published by OpenAI, not WilmerHale's full report. Its conclusions can therefore be reported with attribution; the underlying testimony, documents and reasoning that were not published cannot be treated as publicly inspected. The case teaches readers to examine the architecture of a corporate investigation, not just the size of a number.

What WilmerHale reviewed

The new board created a special committee comprising Bret Taylor and Larry Summers and retained the firm on 8 December 2023. Its mandate was to review events surrounding Altman and Brockman's removal from the board and Altman's dismissal as chief executive. According to the summary, employees, current and former directors and advisers cooperated.

WilmerHale found a breakdown of trust between the former board and Altman. It concluded that the original 17 November announcement accurately recounted the decision and its rationale, that the directors believed they were mitigating internal management challenges and that they did not anticipate the destabilisation that followed. It also found that the board had broad discretion to dismiss Altman, although his conduct did not mandate removal.

That wording is narrower than “Altman was cleared.” Saying that conduct did not require removal does not mean the review rejected every communication problem; the summary retains its conclusion about lost trust. Likewise, saying product safety did not cause the decision does not certify that every product was safe. It identifies the cause of a governance event, not a technical assessment of each model.

The review also identified a procedural failure: the board acted on an abbreviated timetable, without advance notice to key stakeholders, without a full inquiry and without giving Altman an opportunity to address its concerns. That sequence separates authority from process. A body can possess formal power to make a decision and still exercise it with inadequate information or preparation.

Who commissioned, who investigated and who decided

“External” describes the law firm; it does not make the entire decision system independent. A board committee was the client; WilmerHale assembled the record and reported findings; the special committee made a recommendation; and the board decided to endorse the return. Those four roles should not be compressed into “an auditor reinstated Altman.”

To assess independence, ask who selected and paid the investigator, who defined the mandate, whether it could reach every document and witness, who received the report and what was published. The summary says the committee provided resources and authority for a comprehensive review, but it does not provide the complete report, appendices, transcripts or document-selection criteria. That limitation does not invalidate the work; it sets the boundary of public verification.

Document count indicates scale, not quality by itself. Thirty thousand files may contain duplicates or peripheral material; one conversation may be decisive. A review is judged through coverage, chain of custody, corroboration, right of response and the trace from evidence to conclusion. Without the report, the public knows the declared volume and the summary, not that entire chain.

The board that was actually announced

OpenAI added three new directors: Sue Desmond-Hellmann, former chief executive of the Bill & Melinda Gates Foundation; Nicole Seligman, former executive vice president and general counsel at Sony; and Fidji Simo, chief executive and chair of Instacart. Altman returned as a director.

Bret Taylor was not one of the new March appointments: he already chaired the initial board formed after the crisis, alongside Larry Summers and Adam D'Angelo. The difference changes the reading of the transition. The three existing members oversaw the review and remained; the three additions broadened experience in health, nonprofits, law, consumer technology and corporate governance.

The company also announced new governance guidelines, a strengthened conflicts policy, an anonymous reporting line for employees and contractors, and more committees, including Mission and Strategy. A governance measure is not assessed by its label: readers need to know who receives alerts, what independence they have to investigate, how retaliation is prevented and what information reaches directors.

Why OpenAI's structure magnified the crisis

OpenAI did not operate as a conventional company whose board looked only to profitability. The structure announced in 2019 placed the commercial partnership under the nonprofit board's control and put the mission of ensuring artificial general intelligence benefits all humanity first. Investors and employees accepted that this obligation could take priority over their financial interests.

That design gave the board extraordinary power and a difficult responsibility. It had to oversee mission and safety while the commercial organisation raised capital, hired staff and served customers. The dismissal showed that a legal architecture can grant authority without automatically supplying a good process for using it.

Governing AI systems involves more than deciding whether a model passes tests. It requires allocating information rights, defining recusals, documenting disagreements, preparing succession and deciding which incidents reach directors. When information rests on personal relationships, lost trust can become an institutional crisis before a shared record exists.

A record for reading any internal review

A useful record has seven fields. Client: who commissioned and paid. Mandate: which questions were included and excluded. Investigator: expertise and possible conflicts. Evidence: documents, interviews and access. Contradiction: who could respond. Published product: full report, summary or release. Consequence: decisions and controls adopted.

Then copy the exact conclusion before paraphrasing it. Here, product safety did not cause the removal; lost trust did; the board had discretion; Altman's conduct did not mandate dismissal; and the process was abbreviated. Turning that into “nothing happened” removes half the finding.

Questions the document does not answer should also be recorded. The summary does not give every episode that damaged trust, each former director's full position or the documentary analysis connecting facts to conclusions. It does not publish a general assessment of OpenAI's safety. Recognising those absences does not insinuate an answer; it avoids pretending to know one.

From crisis to testable control

For the reforms to have substance, the board needs indicators: time taken to investigate reports, recorded recusals, deployment decisions escalated, risk follow-up, succession exercises and completion of corrective actions. A reporting line without traceability is a mailbox; a committee without independent information is a name.

Altman's return resolved a personnel decision, but did not by itself prove that tensions among mission, safety and growth had been solved. The durable lesson is methodological: when an organisation announces an “independent” review, separate who investigated, who decided and which document can be opened. Only then can a reader distinguish a proved fact, an attributed conclusion and an area that remains outside public view.