There's another comment comparing LLMs to shovels, and I think both that and the power tool comparison miss the mark quite a bit. LLMs are a social technology, and the social equivalent of getting your hand cut off doesn't hurt immediately in the way that cutting your actual hand off would. It's more like social media, or cigarettes, or gambling. You can be warned about the dangers, you can see the shells of wrecked human beings who regret using these technologies, but it doesn't work on our stupid monkey brains. Because the pain of the mistake is too loosely connected to the moment of error. We are bad at learning in situations where rewards are immediate and consequences are delayed, and warnings don't do much.
I guess what I'm really saying is that these safety guidelines are not nearly enough to keep us safe from the dangers of AI that they're meant to prevent.
I agree with the thrust of your argument, a minor wording-quibble: LLM's are a falsely-social technology, in the sense that casinos are a false-prosperity technology and cocaine is a false-happiness technology. It exploits the desire without really being the thing.
Safety should go back to being narrowly defined in terms of reducing / preventing physical injury. Safety is not "don't use swear words." Safety is not "don't violate patents." Safety is not "don't talk about suicide." Safety is not "don't mention politics I don't like." As long as we keep broadly defining it, we're never going to agree on it, and it won't be implementable.
But we cannot guarantee those guidelines to always be followed.
Absolving humans of all responsibility for the negative consequences of their own AI misuse seems to the strike the wrong balance for a healthy culture.
I don't think we disagree.
But other things will:
- Liability rules
- Regulations that you get audited on (esp. for companies already heavily regulated, like banks, credit agencies, defense contractors, etc)
If you get the legal responsibility part right, then the education part flows from that naturally.
One day everything works brilliantly, the models are conservative with changes and actions and somehow nail exactly what you were thinking. The next day it rewrites your entire API, deploys the changes and erases your database.
If only there was intellectual honesty in it all, but money talks.
Are all the tool users required to train your safety guidelines and use it in a context that reminds them they are responsible for following them?
Because if no, then no the guidelines are useless and are just an excuse to push blame from the toolmakers to the users.