The November 2023 firing was supposed to be the inflection point. Sam Altman, caught being not consistently candid with his board, was ousted from the company he built – only to be reinstated five days later when investors made clear that accountability was optional for men who had made them rich. What looked like a crisis of governance turned out to be the first clear signal of something more important: the rules simply do not apply to Altman in the way they apply to everyone else. The Pentagon deal is the logical conclusion of that signal.
OpenAI Was Never a Safety Organisation – It Was Always Heading Here
The revisionism around OpenAI’s founding requires constant pushback. The company was established as a nonprofit in 2015 explicitly to pursue artificial general intelligence in a way that would benefit humanity rather than shareholders. That mission statement was not marketing boilerplate – it was the justification for recruiting some of the world’s best researchers at below-market rates, for claiming special status with regulators, and for securing billions in philanthropic capital.
By 2019, OpenAI created a capped-profit arm to fund expensive model training. By 2022, it launched ChatGPT and became a commercial juggernaut. By 2024, Altman was pursuing a personal chip manufacturing venture while serving as CEO. And in February 2026, OpenAI agreed to deploy its models on Pentagon classified networks – a move that came, according to Reuters, within hours of rival Anthropic being blacklisted by the Trump administration for refusing the same terms.
That sequence is not coincidence. It is a trajectory. The nonprofit mission was progressively subordinated to commercial scale, commercial scale required government partnerships, and government partnerships inevitably lead to the Department of Defense. Gary Marcus, whose Substack documented Altman’s pattern of dishonesty long before it was fashionable to say so, called the Pentagon agreement the culmination of a pattern that stretches back to the 2023 firing. He is right – and the people who dismissed that criticism as sour grapes owe it to themselves to reconsider.
The Safety Narrative Was the Product Being Sold to the Pentagon
What makes the Pentagon deal genuinely alarming is not that OpenAI signed it – every tech company eventually chases government contracts. What is alarming is the specific asset the deal transfers. The Department of Defense did not acquire access to ChatGPT’s raw capabilities. It acquired OpenAI’s safety brand.
As The Verge reported, the key language in OpenAI’s agreement is any lawful use. That phrase means the Pentagon can deploy OpenAI’s models for anything the law currently permits – and U.S. intelligence agencies have spent decades stretching the definition of lawfully permissible to cover surveillance programmes that would horrify the public if fully disclosed. OpenAI’s red lines against domestic mass surveillance and autonomous weapons are not written as hard contractual prohibitions; they are promises contingent on the government not deciding to reinterpret them.
Altman knew the optics were bad. He said so publicly. He admitted on March 3, according to CNBC, that the deal looked opportunistic and sloppy and that the company shouldn’t have rushed the announcement. The company subsequently amended the contract to add explicit language barring surveillance of US citizens – an admission that the original deal did not include those protections. You do not amend language you did not need in the first place.
From ‘Not Consistently Candid’ to Defence Contractor in 28 Months
The timeline is worth holding in focus. In November 2023, OpenAI’s board concluded Altman was not consistently candid in his communications. According to former board member Helen Toner, speaking to The Verge, the board simply could not believe things Altman was telling them – a description of dysfunction that goes well beyond the usual CEO-board friction. Among the specific failures: Altman withheld his ownership of the OpenAI Startup Fund while presenting himself as an independent director, gave inaccurate information about safety processes on multiple occasions, and allowed the board to learn about major product launches via Twitter rather than direct communication.
OpenAI co-founder Ilya Sutskever testified that Altman had been manipulating executives for over a year before his removal. Fast Company reported in 2026 that Altman’s honesty remained under active legal scrutiny. And yet, by March 2026, Altman was using a staff town hall to lecture employees that they did not get to make operational decisions regarding military AI deployment – a man whose own track record of candour with oversight bodies is at best contested now telling safety researchers their concerns are above their pay grade.
Nearly 900 employees from OpenAI and Google signed an open letter opposing the Pentagon’s demands, according to The Guardian. Several senior safety researchers left the company. The chalk messages appearing outside OpenAI’s San Francisco offices – Where are your redlines? – were not the work of naive idealists. They were the work of people who understood, better than most, what the deal actually meant.
What This Actually Means
The Pentagon deal does not represent a betrayal of OpenAI’s founding mission. It represents the completion of a transformation that was underway long before most observers were willing to name it. The 2023 firing was the last moment the company’s governance structure tried to hold Altman accountable; investor pressure overruled it. The commercial pivot, the capped-profit restructuring, the aggressive product releases over safety team objections – each of these was a step on the same path.
What Altman has built is not, and has not been for some time, a safety-focused AI research organisation. It is a technology platform company pursuing scale, government access, and regulatory advantage – and the Pentagon deal is the most honest thing OpenAI has done in years. It tells you exactly what the company is now, and exactly what the 2015 nonprofit charter was always destined to become. The critics who warned about this were not being paranoid. They were reading the trajectory correctly. The endpoint was always here.
Sources
Gary Marcus Substack |
CNBC |
Reuters |
The Verge |
The Guardian |
Fast Company |
TechCrunch