products generally aren't so complicated and so uninterpretable that their behavior isn't reliably predictable. these products are unusual in that regard there might be an analogy to pharmaceuticals? bodies are complicated, side effects are not 100% predictable, drugs can still be sold
to establish liability in here will be extremely difficult and require a bunch of expertise. i think it exists but it is not at all straightforward to establish
Oh yeah I’m not saying this is like an easy slam dunk, but just that it is clearly in the realm of product liability and what will sink them won’t be some esoteric code analysis witness, but some dumb emails people sent that get discovered.
the emails will probably be between the code analysis nerds tbh
most obvious analogy I guess is normal software: it's a computer program, designed for a purpose, which may have bugs that cause harm despite the good-faith best efforts of the vendor the disanalogy is how hard it is to make it follow spec: currently quite hard, but will it stay that way?
Yes it’s designed for a purpose, which is to generate text based on text it was trained on. It’s not a “bug” or unforeseen for it to generate text that it was trained on.
but that's the training objective, not the goal of the company. the objective and what the designers want are misaligned, which is why they spend so much effort on post-training and preference tuning to get it to behave well. the state of the art in building LLMs just isn't that good
if you can only build software by dragging and dropping blocks in visual basic, which is crude and hard to reason about / doesn't expose the primitives you want, you can't build the product you want. did you then intend to get the resulting poorly behaving product? I'd say not really
this implies courts will have to grapple between discussions of intentional deployment (negligent design, failure to warn) and design defect claims - kind of like when Ford knew the Pinto gas tank could cause fires but decided the fixes cost more than lawsuits
I'm not a lawyer (are you?) and I don't know how existing law would handle a kind of product that you can't make behave perfectly even with diligent effort using the best available techniques are you off the hook if you disclose the potential for issues?
not a lawyer - god no, lol - just a file clerk at a law firm. my position atm is essentially that courts are going to need new paradigms in dealing with this stuff altogether, given it seems highly unlikely that courts will deem llms as too dangerous for use as marketed altogether
Exactly. The difference between truly unforeseeable issues with the tech and predictable ones they chose to ignore. “Well if we scrub the training sets of any suicide notes etc it performs 10% worse on regular tasks so let’s leave them in because the risk of killing a few people is worth it”
well, the issue is foreseeable but there's not a perfect way to mitigate it even if you could find any contemplation of suicide in the training data, would removing it help?
you can preference-tune it to avoid talking about this, have a content-policy review model that checks the generations before emitting them, etc, all of which helps but isn't perfect and if you lobotomize it too far, it becomes questionably useful for e.g. med students so how far is far enough?
legally or morally, which might be different
Again, if you stop thinking about this as some sort of magical new novel thing these questions answer themselves. “This product we’re trying to sell is hard to make safe currently” ok? You limit who you sell it to, you have a lot of warnings etc. this isn’t novel.
Sure but if you take that and sell it as one thing and it does other bad things that were pretty predictable, you’re gonna face liability issues.
Right. Tyler's foreseeability point is p solid, yet pharmaceuticals are a good parallel imo - courts recognize certain products as inherently risky but still socially valuable enough to avoid strict liability. big problems: lack of independent ai regulation and best practices that can change monthly