Trump’s AI Oversight Plan Is Everything VCs Claimed To Hate About Biden’s Plan — Only Worse

A key takeaway of the Trump II admin is that they came in declaring a whole bunch of things to be stupid and worth ripping apart, only to later realize how important and structurally necessary those things were, and then rush to recreate them in a much sloppier, worse version. Now they’re doing it with AI policy.

As you may recall, under Biden, there was a push to get the frontier AI model companies to agree to do some voluntary testing in association with the government, to make sure that at least someone was looking over how some of the most powerful models might do bad things. Eventually, the administration worked out an agreement with OpenAI and Anthropic to give the newly established U.S. Artificial Intelligence Safety Institute (USAISI) (a part of NIST) early access to frontier models for some basic standardized testing. You can see an example of the testing results on Anthropic’s Claude Sonnet 3.5 released a few months after that model was already on the market.

Not really a huge deal, but there was some grumbling and concerns among some that the U.S. government really shouldn’t be in the business of reviewing AI models. But some folks completely lost their minds about this. Notably, some of the top VCs in Silicon Valley, such as Marc Andreessen and Ben Horowitz. Andreessen suggested that the Biden approach to regulating AI (which, again, was incredibly light touch and the product of many, many months of negotiations between AI supporters and skeptics) “would impose tyranny far beyond anything even imagined by the Communists and Fascists of the 20th Century.”

Around that time, both Andressen and Horowitz went all in for Donald Trump, citing Biden’s supposedly awful approach to AI (and cryptocurrency) as a key reason. To this day, if you hear either of them talk about this (as they seem to do on every podcast they can join), they repeatedly make extreme claims about how the Biden administration was out to destroy American AI innovation and (phew!) only Donald Trump has brought back sane policies.

Indeed, one of Trump’s first orders of business on Day 1 was to revoke Biden’s policy on AI safety reviews.

However, as we’ve been chronicling, it’s not as if the Trump administration has had any real coherent policy plan on AI. He sure claimed to have a plan, but much of it seemed to actually be focused on culture war bullshit about stopping “woke” AI from existing (which sure sounds like a suppression of speech, but alas). And then, when Pete Hegseth’s lackeys started throwing a shit fit that Anthropic wouldn’t let them use its tools to do mass surveillance on Americans, the admin declared Anthropic woke, and sought to destroy the company.

As we keep warning, the tech bros and VCs who embraced Trump’s version of MAGA fascism in the belief that through him they’d get the regulatory utopia they were hoping for were so obviously going to have the rug pulled out from under them.

And now we’re at that time for the latest rug pull. According to the NY Times, the Trump administration is planning to demand that they get to pre-vet all new AI models:

The administration is discussing an executive order to create an A.I. working group that would bring together tech executives and government officials to examine potential oversight procedures, according to U.S. officials, who declined to be identified in order to discuss deliberations over sensitive policies. Among the potential plans is a formal government review process for new A.I. models.

Oh. A formal government review process for new AI models? That seems more stringent and compliance-oriented than the voluntary setup that the Biden admin had created.

Also, with the Biden admin, the process involved a lot of careful deliberation and building out the capabilities within USAISI within NIST (a widely-trusted agency involved in setting technical standards). It appears the Trump admin version will be… a bit different and less well thought out:

Officials said that if the administration moved ahead with vetting A.I. models, the working group would help determine the agencies that would help with that effort. With no federal agency responsible for all government cybersecurity work, some officials said having the N.S.A., the White House Office of the National Cyber Director and the director of national intelligence oversee the model review was the best way to proceed.

So, we went from a voluntary, fairly light touch program under the Biden admin, where a bunch of tech standards nerds do a pretty straightforward safety review of models, with no further enforcement mechanisms… to one in which potentially the NSA and the intelligence community now gets early access to AI models with some sort of ability to approve or reject them.

What could possibly go wrong?

For starters: everything.

Meanwhile, just recently Marc Andreessen was going on podcasts talking about how happy he is that Biden’s “ruinous” federal attacks on AI are now gone and unlikely to return. Want to try again, Marc?

Look, this was always going to be the result. This is not to say that Biden’s AI policies were good. They weren’t. In typical Biden fashion there were too many competing voices in the room and so the eventually policy outcome was kinda meh. It would not have been my preferred approach. But it was hardly ruinous, let alone a kind of “tyranny far beyond anything even imagined by the Communists and Fascists of the 20th Century.”

But, as we’ve pointed out over and over again, even if you’re able to get on Trump’s good side, MAGA-style fascism is never good for you for very long. Sure you can try to cash out while the getting is good, but eventually that kind of fascism always fails. It’s simply unsustainable, and in the hands of a bunch of incompetent Trump courtiers, eventually it was always going to turn into some sort of effort to control and mold companies towards Trump and his cronies’ interests.

The end result here: we have a way worse version. Just as some of us have been warning about and predicting all along.

Amusingly, the author of Trump’s original AI policy, Dean Ball, wrote in his newsletter just before the NY Times released their article that:

the current trajectory of federal frontier AI governance is worse than the direction of AI policy under the Biden administration…

Gosh. Who could have possibly predicted that an administration full of incompetents, hangers-on, and power-hungry, mediocre, ignorant bros who want to pretend they’re the masters of the universe would fuck this one up?

Yeah, look, maybe next time, instead of embracing obvious fascism, just deal with the fact that sometimes within normal democratic structures you get bad policies you disagree with, rather than deciding we need to set fire to the constitutional order and the institutions that made American innovation so successful.

Hopefully, we can bring some of that back before it’s too late, so that next time, people like Marc Andreessen don’t set American innovation on fire just because the Biden admin hoped that people would check to make sure AI models were a bit safe.