From export control to thought control

A while back, I wrote about the Anthropic export controls and why the panic at the time was pointing at the wrong risk.

Well, things have moved since then. The ban was partially lifted, but with a little catch: the US government now decides which companies get access. OpenAI, not wanting to find out what would happen to them next, agreed to let the US administration vet its users too.

They started with “no export to foreign nationals” and now we’re already at “you go through us for approval on who gets to use this”.

We’re on a different level here. Export controls are blunt instruments with a legal hook. This is gatekeeping: the government decides who the model serves.

And there are pointers to where this might lead medium- to long-term.

Last July, the White House signed an executive order called “Preventing Woke AI in the Federal Government” (yes, that’s the real name) requiring AI models procured by the government to be “ideologically neutral” and “truth seeking”. What those mean is whatever the administration wants them to.

The cumulative effect of all that is a model that will have learned, through training, feedback loops, and acceptable use policies, what it’s allowed or not allowed to say.

We’ve seen it before: ask a Chinese AI about Tiananmen Square. It doesn’t know, or it deflects. Baidu didn’t delete Tiananmen or add a rule saying “don’t mention June 4th”, there was no need…

The worry here isn’t the US government issuing explicit instructions about what models can say. It’s whether a slow, methodical, version of the same process isn’t already in its early stages. And would we notice if it was?

We mostly apply digital sovereignty to data and cloud services, which are passive. But it applies to AI models too and these models nudge and shape what you think.

Think about that (without AI).

Colin