Best Way To Pace Frontier Is To Hold Labs Liable For The Behaviour Of Their Models: Naval Ravikant

AngelList founder Naval Ravikant said the best way to pace frontier AI development is to hold labs fully liable for the behavior of their models, extending that liability to hosts serving dangerous or weakly-guarded open source models and to end users who deliberately jailbreak a model to cause harm. Ravikant's accountability-at-every-layer framework contrasts with the EU's prescriptive, risk-tiered AI Act, and it aligns with a Washington mood in which Vice President JD Vance has questioned whether incumbent labs lobbying for safety rules are protecting the public or raising entry costs for smaller competitors. Economist Tyler Cowen has warned that liability could spiral out of control as autonomous agents multiply, a problem Ravikant's approach sidesteps by keeping liability with whoever controls the model at each stage.

AngelList founder and startup philosopher Naval Ravikant has weighed in on one of the thorniest questions in AI policy right now, one that’s come up as frontier labs increasingly clash with governments https://officechai.com/ai/ai-companies-begging-for-regulation-feels-like-a-trojan-horse-us-vice-president-jd-vance/ over how much oversight AI actually needs. Ravikant’s take is that liability, not new rulebooks, is the mechanism that should keep frontier AI development in check. According to Ravikant, the best way to pace the frontier is to hold labs fully liable for the behaviour of their models. He extended the logic downstream too: for “dangerous” or weakly-guarded open source models, the hosts serving them should bear the liability. And end users aren’t off the hook either, in his view, if someone deliberately jailbreaks a model to cause harm, that responsibility sits with them. It’s a framework built around accountability at every layer of the stack rather than a single centralized rulebook trying to anticipate every possible misuse in advance, an approach that stands in contrast to the EU’s more prescriptive, risk-tiered AI Act. There’s an elegant simplicity to Ravikant’s pitch. Instead of regulators trying to predict every way a model could go wrong and writing rules for each scenario, ahead of time, liability puts the burden squarely on whoever controls the model at each stage. A lab that ships a system knowing it’s inadequately tested would be the one paying for the fallout, which naturally pushes safety work earlier in the development cycle rather than treating it as a compliance checkbox. That’s arguably a more effective incentive than a pre-approval process, because it doesn’t require regulators to be smarter than the labs themselves, it just requires courts to assign fault after something goes wrong. The framing also fits the current mood in Washington. The current administration has been openly skeptical of labs asking for new safety regulation, arguing that the government already has substantial criminal and regulatory power https://officechai.com/ai/dario-amodei-pretending-to-be-a-perfect-little-angel-by-asking-for-ai-regulation-donald-trump/ over these companies without needing new legislation. Vice President JD Vance has gone https://officechai.com/ai/ai-companies-begging-for-regulation-feels-like-a-trojan-horse-us-vice-president-jd-vance/ a step further, suggesting that when the biggest incumbents in the room are the ones lobbying loudest for safety rules, it’s worth asking whether that regulation protects the public or simply raises the cost of entry for smaller competitors trying to catch up. Existing product liability, negligence, and consumer protection law already give courts plenty to work with when a company’s product causes harm, and there’s a reasonable argument that AI doesn’t need a bespoke legal regime layered on top before anyone’s even sure what the actual failure modes look like at scale. Economist Tyler Cowen has floated even more radical ideas in this vein, at one point suggesting that as autonomous agents multiply https://officechai.com/ai/ai-agents-should-be-governed-by-laws-written-by-ai-economist-tyler-cowen/ , liability could spiral out of control if every offshoot of an offshoot gets traced back to the original model maker, and that some kind of containment on how far liability travels may be needed. Ravikant’s split-the-difference version sidesteps that problem by keeping liability close to whoever has the most control at each point, the lab for their own frontier model, the host for a loosely-guarded open weight model, and the individual for deliberate misuse. It’s the kind of answer that doesn’t require anyone to agree on how dangerous AI ultimately becomes, since a strong liability regime works whether the risk turns out to be modest or severe, and it leaves labs with a very direct reason to get their models right before, rather than after, they reach the public.