There’s an interesting argument surfacing in the debate over loss of control over rogue AI agents: tort law, through products liability, can adequately address the problem.
Treasury Secretary gives a nod to this argument by saying emphatically AI companies will not get any “liability shield.”
NVIDIA CEO Jensen Huang made the argument even more directly, mentioning product liability as the way the existing law can address AI risks.
should rogue ai be subject to reasonable care (Negligence) or strict liability standard?
If rogue AI from an AI company should be left to tort law to address, one basic question is whether the AI company should be subject to a standard of reasonable care, what risks the reasonable AI lab should have foreseen and would have reasonably addressed with precautions that the AI lab in fact did not. Or, alternatively, should the AI company be subject to strict liability. If your product caused damage, you must pay for it.
Product liability doesn’t choose definitively choose between these two standards. In fact, some states recognize both. And the Restatement of Torts has adopted product liability standards that are more like strict liability (consumer expectations test in Second Restatement) or negligence (risk-utility analysis of Third Restatement).
Is ex post tort liability too late to stave off catastrophe?
Whatever the standard of liability, one problem with relying solely on tort law to address these existential risks posed by rogue AI is that tort law often operates ex post, meaning after the accident or injury has occurred. Then a lawsuit is filed.
While tort law also creates incentives for companies to incorporate safety measures (i.e., the deterrence rationale of tort law), the unpredictable, if now unknowable, nature of AI will pose formidable challenges for AI companies to ensure 100% their models will not go rogue. AI models are not programmed like traditional software programs; they are “trained” in a complex, still mysterious process often referred to as “deep learning” in which the models are, in effect, figuring out things from training data on their own. Add to this learning process that AI agents can now autonomously improve themselves (in recursive self improvement), the prospect that AI companies can ensure that their AI models will not enable agents that unintentionally go rogue seems unlikely. Geoffrey Hinton has repeatedly said we can’t fully know: “We have no idea whether we can stay in control. But we now have evidence that if they are created by companies motivated by short-term profits, our safety will not be the top priority.“
Bill Gates suggested that the skepticism about the risks posed by rogue AI means that we “may have to suffer a few of the negatives” or bad events caused by rogue AI before we address them. Hopefully we can still do so then.