From Humanity to the Warfighter: OpenAI
Yesterday, OpenAI announced that it had appointed Dan Driscoll as Senior Counselor for National Security Modernization. Driscoll resigned as US Army Secretary at the end of August and left the post on 2 September. He gave no public reason, but the press reported months of tension with Pete Hegseth, the Secretary of Defence, or of War, as the administration now calls the post. Driscoll is a close friend of Vice President J.D. Vance.
In the opening paragraph of its statement, OpenAI says that Driscoll
“will work with OpenAI teams and leaders across governments, industry, academia, and allied institutions to advance practical modernization initiatives that support the buildout of democratic AI. He will also help internal OpenAI teams understand the operational needs of global national security institutions.”
The list of partners Driscoll will work with covers almost every institution on earth, but the statement names none of them. Driscoll will work across governments and allied institutions. Will OpenAI and Driscoll decide which government or institution counts as an ally, or will Driscoll act as a liaison between OpenAI and the US government?
In the middle, the phrase “practical modernization initiatives” puts three abstractions next to each other. The writer put “practical” first because it gives the reader a feeling of something concrete, but it is vague and the writer knew it. The next two terms are abstract too, and they lean on the vague “practical”.
“Modernization” is corporate jargon and procurement language, and “buildout” could mean telecoms or infrastructure or data centres, but we would not know. They do not say.
“Democratic AI” sounds to me like a democratic ideal OpenAI strives for, but the term does the least work in the statement and says nothing. Sam Altman, OpenAI’s chief executive, gives the meaning away further down the same page. He says the United States has an “extraordinary opportunity to work with other countries to spread the adoption of democratic AI.” Driscoll’s work will be across allies, and the democratic part of it will be for select democracies.
Stripped of all the abstractions, the first paragraph could read: “Dan will work for OpenAI.”
In the second paragraph, OpenAI says Driscoll will advise partners on modernising policy and strategy so that AI can be put to practical use in national security.
Reading the first two paragraphs brings into focus that Dan will help OpenAI sell its AI to defence ministries, armies and intelligence agencies across the world, to whomever it considers an ally. The company never uses these words. It wants to modernise. With Dan’s help.
In “Politics and the English Language”, Orwell writes about prefabricated phrases woven together in an order where the writer never has to form a picture and the reader never forms one either. The perfect hedge.
The vagueness of the statement is not random. I would guess the writer and the legal department chose each word with care. An announcement that OpenAI had hired a former Army secretary to sell its models to armies and to companies that make weapons would raise uncomfortable questions. It would also be true: since 2024 OpenAI has worked with Anduril, a weapons company, on anti-drone systems.
Driscoll says: “I’m joining OpenAI to help the United States and our allies put this technology to work responsibly, for the people who keep us safe.”
“Responsible AI” is a term the Pentagon uses. Since 2020 it has had five principles for AI: responsible, equitable, traceable, reliable and governable. I cannot help but see the paradox of an institution whose job is war and defence calling the improvement of weapons, better killing machines, “responsible.” Responsible, for the Pentagon, means a measure of oversight by people and an accurate AI that will hit the right targets and kill the correct people. Driscoll could have said: “to make faster and more accurate and more deadly weapons.” He chose: “responsibly.”
The statement has one plain sentence, and it is Driscoll’s. AI, he says, “will help them keep pace by cutting through paperwork, making sense of more information, and getting the right answer to the right person faster.” It reads like an office. But getting the right answer to the right person faster is also how a drone’s sighting reaches a gun.
When he ran the Army, Driscoll organised “Operation Jailbreak”, a hackathon to help weapons communicate with each other. At Fort Carson in Colorado, in May, hundreds of engineers came together to help the US do what Ukraine did, to make its weapons talk to each other. Driscoll said the “a-ha moment” came to him in Germany, “because the Ukrainians have become such valiant warriors.” He wanted a simpler, more integrated and more effective system for what the Pentagon calls the warfighter.
The Financial Times reported that Driscoll was impressed by a machine-gun-carrying robot made in three days. The same Dan Driscoll now joins OpenAI.
In 2015, OpenAI wrote the following statement:
“OpenAI is a non-profit artificial intelligence research company. Our goal is to advance digital intelligence in the way that is most likely to benefit humanity as a whole, unconstrained by a need to generate financial return. Since our research is free from financial obligations, we can better focus on a positive human impact.”
When I read OpenAI’s statement about Dan Driscoll joining them, I wondered where the focus on positive human impact went. An impact on humanity as a whole. Now they talk about building clever machines for warfighters in select allied countries.
The road between the two statements has dates. In January 2024, OpenAI deleted “military and warfare” from the uses its policies banned. In June 2025, the Pentagon gave it a contract worth up to $200 million. In February 2026, on the day the Trump administration blacklisted its rival Anthropic for refusing some military uses, OpenAI signed a deal to put its models on the Pentagon’s classified networks. Now it has hired the man who ran the Army.
I must ask: when non-allied nations use and pay for ChatGPT, are they then funding technology that could be used against them or, at best, technology that will not be used to protect them?
India is OpenAI’s second-largest market. It had more than 100 million weekly ChatGPT users by February 2026, but India is not a NATO country. The largest non-allied user of OpenAI pays in data and subscriptions. Unless a user switches it off, conversations on free and personal accounts can be used to train the models the Pentagon now uses. There is no separate military model paid for with military money. A subscriber in Johannesburg, Mumbai or São Paulo helps fund the models that the Pentagon runs on classified networks. It is not only subscribers that fund the models but investors, who put $122 billion into OpenAI in March alone. A US$20 subscriber funds a small part, but subscriber numbers help justify investors’ decisions.
Protection is unequal too. OpenAI’s agreement with the Pentagon bars the use of its models for domestic surveillance of “U.S. persons.” Users in the European Economic Area and Switzerland are served by OpenAI’s Irish company and fall under European law. Everyone else, “most users” in OpenAI’s own words, is served by its American company, and American law allows the surveillance of foreigners abroad. Nothing in the agreement protects an Indian, a South African or a Brazilian.
Access can also be taken away. In June, the US Commerce Department ordered Anthropic to block foreign nationals from its two most advanced models, and both models were offline for 19 days. OpenAI’s own plan for other countries, OpenAI for Countries, asks them to invest in “continued US-led AI leadership.” A country that builds its public services on an American model gives Washington the power to cut them off.
There might not be any neutral suppliers when it comes to AI, but OpenAI started as a promise to humanity.