Wealthy tech entrepreneurs including Elon Musk launched OpenAI in 2015 as a nonprofit research lab that they said would involve society and the public in the development of powerful AI, unlike Google and other giant tech companies working behind closed doors. In line with that spirit, OpenAI’s reports to US tax authorities have from its founding said that any member of the public can review copies of its governing documents, financial statements, and conflict of interest rules.
But when WIRED requested those records last month, OpenAI said its policy had changed, and the company provided only a narrow financial statement that omitted the majority of its operations.
“We provide financial statements when requested,” company spokesperson Niko Felix says. “OpenAI aligns our practices with industry standards, and since 2022 that includes not publicly distributing additional internal documents.”
OpenAI’s abandonment of the long-standing transparency pledge obscures information that could shed light on the recent near-implosion of a company with crucial influence over the future of AI and could help outsiders understand its vulnerabilities. In November, OpenAI’s board fired CEO Sam Altman, implying in a statement that he was untrustworthy and had endangered its mission to ensure AI “benefits all humanity.” An employee and investor revolt soon forced the board to reinstate Altman and eject most of its own members, with an overhauled slate of directors vowing to review the crisis and enact structural changes to win back the trust of stakeholders.
Access to OpenAI’s conflict-of-interest policy could show what power the new board really has over Altman and his outside pursuits, which include personal investments in numerous startups pursuing AI projects and a nuclear reactor maker. His day job and personal projects intermingling played some role in board members’ distrust, according to people involved in the situation but not authorized to discuss it. In 2019, while Altman was at the helm, OpenAI signed a nonbinding letter of intent to buy $51 million of AI chips from Rain, a startup in which he has invested more than $1 million, WIRED reported last month. OpenAI hasn’t moved forward with a purchase. Felix says Altman is transparent with the board about his investments and follows a process for managing potential conflicts.
Some sunlight on OpenAI’s governing documents could reveal whether it has made revisions to stabilize an unusual corporate structure and potentially pacify backers such as Microsoft. The company’s founding bylaws, publicly available via its 2016 application to the Internal Revenue Service for tax-exempt status, indicate how a fraction of the board could take control and push out Altman. OpenAI’s filings to the IRS through 2022 reported that no “significant changes” had ever been made to its governing documents. But the company almost certainly made updates after Altman’s return to allow it to give a nonvoting seat on the nonprofit board to Microsoft, whose CEO, Satya Nadella, complained publicly that he had been blindsided by Altman’s firing. Any additional changes made at that time remain a secret.
WIRED’s request for the documents promised in OpenAI’s IRS filings fell to the counsel for its nonprofit to decide on. Their denial and new policy of withholding those documents extends an existing trend of dwindling openness at a project founded on transparency. OpenAI once published extensive detail about its AI inventions but has more recently guarded the technical details and data behind its most famous tool, ChatGPT. Felix, the company spokesperson, says San Francisco-based OpenAI discloses all the material required by the IRS and California’s attorney general and that it regularly publishes information about its research and safety work, while also making its research freely available in the form of tools such as ChatGPT.
OpenAI’s declining openness has been most notable since 2019, when the nonprofit created a for-profit subsidiary to house most of its AI development and draw outside investment. That opened the way for OpenAI to hitch its future to the largesse of Microsoft, one of the tech giants it was founded to challenge, and also to shroud its finances. OpenAI cofounder turned competitor Elon Musk said at a New York Times event in November that his former company should be called Super-Closed-Source-for-Maxiumum-Profit-AI.
OpenAI’s original nonprofit organization—and its board—retain ultimate control of OpenAI’s activities and technology. Like all US nonprofits, it must publicly share upon request a copy of its annual report to the IRS and indicate in those submissions whether any additional documents like its bylaws or similar or related documents such as a governing constitution or conflict of interest policy were also available to the public during the last year.
Some notable nonprofits do that, like the Bill & Melinda Gates Foundation, which publishes its bylaws and rules on conflicts and workplace relationships, but it’s not standard practice.
“It is not common for organizations to make their governing documents or internal policies public,” says Rick Cohen, chief operating and communications officer for National Council of Nonprofits, an advocacy group.
Yet for seven consecutive years, from its founding through 2022, OpenAI stated in its annual IRS filings that it made those submissions as well as other files available “upon request.” It’s unclear if anyone ever took OpenAI up on the invitation in the years through 2022—OpenAI won’t say.
Last month, after two days of waiting on OpenAI communications staff to fulfill an emailed request for its governing documents, conflict rules, and financial statements, WIRED rang the doorbell outside OpenAI’s San Francisco headquarters on December 14 asking to see all those documents. A receptionist said over an intercom that wouldn’t be possible, hung up, and didn’t reengage. OpenAI’s IRS filing for 2023 that would reflect it has changed its previous policy isn’t due until later this year.
To encourage oversight of nonprofits, US tax law requires them to at least make their annual reports to the IRS, form 990s, available for public inspection at their offices the same day they were requested if they are not posted on an organization’s website. OpenAI doesn’t post its reports on its website but still didn’t provide them when WIRED showed up in person. IRS documentation says violators can face fines of $20 a day, up to $10,000, but the agency declined to comment about OpenAI, citing confidentiality provisions of tax law. OpenAI hasn’t been accused of wrongdoing, and Felix says OpenAI’s reports are available online through government and research databases.
Even when nonprofits tell the IRS their internal documents are open to all, they can be difficult to access. “When I have asked a group for forms it says are public, I basically never get them unless I cc the IRS and the organization’s general counsel,” says Brian Galle, a Georgetown University law professor specializing in nonprofit issues.
The financial statement OpenAI shared in response to WIRED’s request excludes the results of OpenAI’s “affiliated entities,” most crucially the for-profit unit that sells ChatGPT and other services, citing a desire to protect trade secrets. The statement, which covers 2022, shows just $44,000 in revenue and $1.3 million in expenses. That’s accurate for the nonprofit, but OpenAI overall reportedly generated hundreds of millions of dollars in sales last year and spends even more on high-end computers and top-flight researchers.
Investors in OpenAI’s for-profit arm and employees of the company likely have access to some of the records the company now declines to share publicly, though they are bound by nondisclosure agreements covering internal files. Because OpenAI’s nonprofit takes in virtually no public support, it lacks the incentive that some nonprofits have to be more transparent so as to loosen the purse strings of donor or grant makers.
OpenAI might win more loyalty from its customers or trust from regulators by opening up, but Altman has said the company didn’t lose any clients during its two weeks of high drama and generally has been greeted warmly by political leaders. The company may feel that the ongoing board review into OpenAI’s governance and what is shared from its findings will be enough.
Without a reversal in OpenAI’s policy and a follow-through on its promises, so much about an increasingly influential organization may never become publicly known, like whether the new board amends the conflict-of-interest policy to better wrangle Altman and other executives. Back in 2022, it may have been possible to get the answer with just the ring of a doorbell.
https://www.wired.com/feed/rss
Paresh Dave
