Lies and deception: Andreessen Horowitz’s SB 1047 campaign is as misleading as it gets

The venture capital giant has repeatedly peddled untruths in an effort to kill the AI regulation bill

In the fight against SB 1047, California’s effort to regulate the potential catastrophic risks of AI, one voice has been louder than all others: Andreessen Horowitz. The $43 billion venture capital firm, also known as a16z, has pulled out all the stops to kill the bill: hiring one of Gavin Newsom’s closest advisors as a lobbyist, urging voters to lobby against the bill, and encouraging senior politicians to badmouth it.

The company’s campaign has been extraordinarily successful, bringing exceptional attention to a state-level bill and galvanising significant opposition. But that success has been built on misleading claims and outright lies.

Throughout a16z’s efforts to kill the bill, two false claims dominate. Firstly, the company has said that SB 1047 “requires developers to prove that their models can’t possibly be used for any of the defined hazardous capabilities”. That’s simply not true: the bill only ever asked for “reasonable assurance” from developers, and the text at the time a16z made this claim clearly stated that reasonable assurance “does not mean full certainty or practical certainty”. (The bill has since been amended to replace “assurance” with “care”, an even lower standard.)

Andreessen Horowitz’s other big claim is the idea that the bill will kill open-source AI through an emergency shutdown provision. But, as bill author state Sen. Scott Wiener has explained, “the bill has never required the original developer to retain shutdown capabilities over derivative models no longer in their control”, and was in fact amended to make this “crystal clear”. Nevertheless, the claim continues to form a crucial part of the campaign against the bill.

These two falsehoods join a laundry list of other untrue and misleading claims. The company said the bill would “impose civil and in some cases criminal liability on model developers”, without noting that this liability only applied to “people who intentionally lie to the government”. It’s also led people to believe that SB 1047 would be burdensome for startups, despite it being squarely targeted at models that cost more than $100m to train — hardly the domain of small companies. And it’s said that the bill was written without “serious collaboration from AI’s expert developers, investors or researchers”, despite the bill being crafted with input from Dan Hendrycks, an accomplished AI researcher who has developed some of the leading AI benchmarks. More recently, it’s received support from over a hundred AI researchers at OpenAI, Google DeepMind, Meta and Anthropic, as well as a swathe of high-profile AI academics — hardly indicative of an uninformed bill.

And while Andreessen Horowitz may initiate the spread of misleading statements, they soon take on a life of their own, laundered through other, more reputable people until they’re cited uncritically — and incorrectly — as fact. In an op-ed for Fortune, “godmother of AI” Fei-Fei Li falsely argued that the bill’s “kill switch [provision] will devastate the open-source community”. (Notably, Li also failed to disclose that Andreessen Horowitz is an investor in her new $1 billion AI company.)

Even members of Congress have joined in on the lies. In a letter to Gavin Newsom, Reps. Lofgren, Eshoo, Khanna and others claimed the bill “requires firms to adhere to voluntary guidance issued by the National Institute of Standards and Technology, which does not yet exist”. This is, bill cosponsor Sunny Gandhi says, “incorrect on two counts”: such guidance does, in fact, exist — but the bill does not require adherence to it.

Despite the false claims from both Li and the representatives, both were approvingly cited by Rep. Nancy Pelosi as part of the reason for her opposition to the bill (in a statement which itself regurgitates the incorrect idea that the bill is “ill informed”). That was swiftly followed by a letter opposing the bill from OpenAI, which makes few actual arguments against the bill — but cites Pelosi and the other representatives’ opposition as reason to veto the bill.

Andreessen Horowitz is not the only company making things up about SB 1047. In a letter organised by Y Combinator, a prominent startup incubator, dozens of startup founders claimed that the bill “would mean that AI software developers could go to jail simply for failing to anticipate misuse of their software”. That claim is flatly untrue (in Sen. Wiener’s words, it is “categorically false — and, frankly irresponsible”): the bill would have no such implications.

Nor is this the first time Andreessen Horowitz has lied in an effort to avoid AI regulation. Last November, the company sent an open letter to President Biden which claimed that the “‘black box’ nature of AI models” has been “resolved”. This claim, described by Google DeepMind’s mechanistic interpretability lead as “massively against the scientific consensus”, was later retracted by Andreessen Horowitz partner Martin Casado. Yet despite confessing the lie several months ago, Casado still has it pinned to his Twitter timeline. (Andreessen Horowitz and Casado did not respond to a request for comment on this article.)

There is a reasonable case to be made against AI regulation in general, and SB 1047 in particular: Timothy Lee and Dean Ball, for instance, have offered good-faith criticism of the bill. Unfortunately, though, much of the debate does not focus on the substance. Instead, it is full of lies and half-truths, pushed by people who ought to know better but, faced with a choice between truth and their personal profits, choose to prioritise the latter.

This piece was co-written with Garrison Lovely, one of the best AI journalists around. You can subscribe to his Substack here and follow him on Twitter here.