The true meaning of principled engagement with the government has become the AI industry’s most contested question, with OpenAI claiming to have achieved it through a Pentagon deal and Anthropic having demonstrated what it looks like when principled engagement fails to reach agreement. The debate between these two definitions of principle will shape how AI companies govern themselves and negotiate with governments for years to come.
Anthropic’s definition of principled engagement was specific and consistently applied. The company engaged with the Pentagon in good faith over many months, offering comprehensive support for lawful military operations while excluding two uses it considered ethically unacceptable. This was engagement — sustained, serious, and in good faith — but it was engagement with conditions that the administration ultimately refused to accept.
The administration’s refusal was expressed in maximalist terms. President Trump’s ban on all federal use of Anthropic technology, framed in language that cast the company’s ethics policy as politically motivated obstruction, rejected not just Anthropic’s conditions but the premise that AI companies are entitled to impose any conditions at all on government use of their products.
OpenAI’s Sam Altman offered a different definition. His Pentagon deal, announced the same night, claimed to prove that principled engagement is achievable — that AI companies can maintain ethical conditions on government use while still securing contracts and advancing commercial goals. His call for these terms to be offered to all AI companies suggested he sees the deal not as an exception but as a template.
The hundreds of workers who signed solidarity letters with Anthropic have offered their own definition, one that prizes consistency and conviction over contractual cleverness. Anthropic itself, holding its position without adjustment after losing its government contracts, embodies that definition in practice. Whether principled engagement means making a deal or refusing to make one without adequate protections will be debated in the AI industry long after this week’s specific events have faded from daily attention.
