Are frontier models really too dangerous?
To an increasingly authoritarian regime, the opportunity to shape how models describe the world may be too good to pass up on.
Link: OpenAI will delay GPT-5.6 after Trump administration request, by Hayden Field in The Verge
I’ve got (at least) two worries about the story that the Trump Administration halted the release of models from both Anthropic and OpenAI.
Anthropic recently pulled its Fable model release in response to the government. Now it turns out that OpenAI has done something similar:
“The Information reported that OpenAI CEO Sam Altman told employees Wednesday in a company Q&A that it would release GPT-5.6 in limited preview form — granting access only to a small group of enterprise customers — in compliance with a request from the federal government. During that preview period, the Trump administration itself would reportedly approve access for customers on a case-by-case basis.”
In some ways, what a coup for the AI industry. This technology is so powerful that the government doesn’t think anyone should have it — and when it does inevitably release into the public’s hands, what a valuable product that will be. Get the tech that’s too dangerous to be released! This magical product can be yours for an unbelievable price!
So one worry is that this is, in essence, great marketing for these vendors.
But it’s worth remembering that these AI models are black boxes that respond to information queries in opaque ways. The more people rely on them for knowledge, the more powerful the models become. The argument being presented is that they can be used in ways that might present traditional security threats — but consider that some versions of the truth, to the wrong kind of authoritarian-minded government, might also be considered a threat. (Remember that “extremism on migration, race, and gender” and hostility to “traditional American views” are now considered markers of domestic terrorism.)
This is a golden opportunity, in other words, to hit pause on frontier model releases, at a time when models are becoming more prevalent, in order to make sure models are shaped to represent a certain version of the world. The administration has already signaled a willingness to do this; there is nothing to say they aren’t. The only way to prove that they aren’t is to open source not just the models but the training process and make the whole thing transparent and verifiable. The industry is a long way off from doing that.
