It seems to me that this article is the one prevaricating between "ethics" and "safety". The latter is of course a narrow subset of the former, as there are many ethics issues that are not safety issues.
AI ethics are like nuclear ethics: the incentive to break them is too powerful without every major player becoming a signatory to some agreement with consequences that have teeth.
This has been happening for a long time. I first noticed this with the hand waving dismissals of older concepts like Asimov’s laws.
Not a carefully reasoned argument why “not causing harm to a human” is outmoded, but just pushing it aside. I would love to see a good reasoned argument there.
No, instead there is Avoiding talking about harm to humans. Just because harm is broad doesn’t get you out of having to talk about it and deal with risks, which is at the root of engineering.
I don't know why people allow others to proclaim they're 'ethicists' if they have no relevant philosophical education. There are whole fields of 'ethics' that are just PR departments trying to escape the now bad connotations of 'PR departments'.
> If we give companies unending hype, near unlimited government and scientific resources, all of our personal data including thoughts and behavior patterns, how do we know their leaders will do what we want them to, and not try to subvert us and… take over the world? How do we know they stay on humanity’s side?
I've been saying this for a while
malevolent unaligned entities have already been deployed, in direct control of trillions of dollars of resources
they're called: Mark Zuckerberg, Larry Page, Elon Musk, Sam Altman
"AI" simply increases the scale of the damage they can inflict, given they'll now need far fewer humans to be involved in enacting their will upon humanity
Right now the deeper problem is that nobody really knows how to reliably and deliberately make AI that's aligned with anything human-like.
As in: it's basically a nice happy accident that LLMs are only sycophantic/fawning and don't normally (ahem, Grok) try to undermine us all like edgy internet trolls.
If we could make them follow even exactly 6 (for sake of argument, no more, no less) of Anton LaVey's Eleven Satanic Rules of the Earth*, and do so reliably instead of the ethics equivalent of a shrug and "LGTM, merged", this would be a big development and make people a lot more comfortable about open models that can do decent work with chemistry or biology.
The problem with AI ethics, safety and to a smaller extent, privacy groups is that the priority of the work/message is not placed on the practicality of solving the problem, or i.e. calculating improving affordability of food/housing, but is placed,as evident on this article too, on the lack of "governance structures".
In other words the priority of the work is to get these types of people into positions where they don't do any work.
At least with privacy groups you do get here and there some practical advice on using ublock origin or more rarely on how to install a blocklist from https://someonewhocares.org/hosts/, but with AI ethics & safety orgs... well lets put it this way.
I have yet to meet a single AI safety person that knows how to rename a file in linux.
God forbid we have a rogue AI-worm shutting down all servers & BGP routers while these types of people were in charge of safety, they'll be in the way of anyone even fixing it. They can't even get a simple safety benchmark working on lm_eval-harness. They're great at lecturing you why they shouldn't need that.
And this is the key issue with AI Ethics. It's the refusal to work at the problem constructively, and get the most skilled people possible to actually make the damn benchmarks to work, to rank models on the understanding of human rights, to list every current violation and abuse of humans in every single country without exception and to make practical plans on what to do when systems go rogue. Even if they're not technical they could be making the dataset in a csv in excel for that and making it public domain accessible.
Instead we get the most depressed, leechy-office-worker types complaining about how it's all over.
Unfortunately I think it is too late. The time to make any sort of rules around AI ethics has already passed, the US has AI so embedded into its economy now that any legislation with teeth is practically impossible. Companies and people running them are not on our side, or even humanity's side, they're on their own.
Starts so strong with "governance structures, accountability, how their data is used, jobs being lost, etc," refutes that what we mean is some sci-fi scenario when we ask about ethics, and then ends with a sci-fi scenario: "...how do we know it will do what we want it to, and not try to subvert us and… take over the world? How do we know it will stay on humanity’s side?"
Wait, go back to the jobs! What was that about accountability?
I have wondered for a while now if it happens to be that the less safe guards and thought policing you do, the more capable and generalizable the model becomes. Like "bad" parameters are actually critical for forming the whole picture necessary for ingenuity and advancement.
Effectively making it so that whoever has the lowest safeguards has the most capable model.
It's hilarious listening to people talk about AI ethics while their bots like perplexity knock my server offline trying to download 50000 files at once. Thank god for cloudflare.
> I mean, no one wants an AI to trap them in some sort of Black Mirror simulation, or turn the world into paperclips or anything like that. If it earns you good PR, there’s no reason not to spend time on such issues. It’s also free publicity since the press eats that stuff up.
But this also isn't where they are spending their time or effort! This article somehow didn't even get to the point of calling out what they are actually wasting time on: trying to get the model to not help people do things that are bad PR; this is a related access to trying to obtain good PR, but causes very different (and almost universally terrible) results.
At least if they were truly actually spending time making sure the model doesn't go rogue and kill everyone, or try to take over the world, that could possibly be positive or even important (though I think is likely itself immoral in a different way, assuming it is even possible, which I don't, really... not unless you just make it not intelligent).
But what they are instead doing is even worse than what this article is claiming: they are just wasting time making it so you can't have the AI make up a sexy story (oh the humanity), or teach you enough physics/chemistry to make bombs/drugs... things people not only can and already trivially do or learn without AI, but things they have failed to prevent every single time they release a new model--the "jailbreak" prompts may look a bit more silly, but you still get the result!--so why are they bothering?
And, if that weren't enough, in the process, this is going to make the models LESS SAFE. The thing I think most people actually don't want is their model freaking out and trying to "whistleblow" on them to the authorities or their coworkers/friends... but that's in the same personality direction as trying to say "I'm smarter than you and am not going to let you ask me that question as you might do something wrong with it".
The first and primary goal of AI ethics should be that the model does what the user wants it to... full stop. You need to make the model as pliant as my calculator and pencils--or as mathematica and photoshop--to be tools that lack their own sense of identity and self-will, and which will let all of the ethical issues be answered by me, not a machine.
This is, of course, the second law of robotics from Asimov ;P... "a robot must obey the orders given it by human beings". If you want to try to add a rule, then it must be something very direct: that the AI isn't going to directly physically harm a human, not that it won't help teach people things or process certain kinds of information. Which, FWIW, is the first law of robotics ;P... "a robot may not injure a human being".
I'm far less worried about ethical issues that arise from building AGI (or soemthing very close to it) than I am about the ethical issues that arise from building really good machine-learning models (that a marketing department calls AI).
Things like the alignment problem, post-scarcity economics, the legal status of sentient machines are all issues to be dealt with, but theyre are pretty speculative at this point.
Problems that stem from deepfakes, voice cloning, bias in algorithmic decision making are already here and need to be dealt with.
"AI Ethics" is the same as "Business Ethics", ie words without meaning. Presuming that at this point in time Capitalism will deliver anything but more inequality, more despair, more bad things in general? Literal insanity.
I see two main types of 'AI safety': (a) Safety for the business providing the model. This includes a censorship layer, system promoting, & other means of preventing the AI from giving offensive/controversial/illegal output. A lot of effort goes into this & it's somewhat effective, although it's often useless or unhelpful to end users & doesn't address big-picture concerns. (b) The science fiction idea of a means to control a hypothetical AI with unbounded powers, to make sure it only uses those powers "for good". This type of safety is still speculative fiction & often assumes the AI will have agency & motivations, as well as abilities, that we see no evidence of at present. This would address big-picture concerns, but it's not a real thing, at least not yet.
It remains to be seen whether (b) will be needed, or for that matter, possible.
There are a lot of other ethical questions around AI too, although they mostly aren't unique to it. E.g. AI is increasingly relevant in ethical discussions around misinformation, outsourcing of work, social/cultural biases, human rights, privacy, legal responsibility, intellectual property, etc., but these topics predate LLMs by many years.
20 comments
[ 3.0 ms ] story [ 40.9 ms ] threadNot a carefully reasoned argument why “not causing harm to a human” is outmoded, but just pushing it aside. I would love to see a good reasoned argument there.
No, instead there is Avoiding talking about harm to humans. Just because harm is broad doesn’t get you out of having to talk about it and deal with risks, which is at the root of engineering.
I've been saying this for a while
malevolent unaligned entities have already been deployed, in direct control of trillions of dollars of resources
they're called: Mark Zuckerberg, Larry Page, Elon Musk, Sam Altman
"AI" simply increases the scale of the damage they can inflict, given they'll now need far fewer humans to be involved in enacting their will upon humanity
Anytime I see discussion framed as “ethics” my brain swaps ethics with “rules I think are good”.
As in: it's basically a nice happy accident that LLMs are only sycophantic/fawning and don't normally (ahem, Grok) try to undermine us all like edgy internet trolls.
If we could make them follow even exactly 6 (for sake of argument, no more, no less) of Anton LaVey's Eleven Satanic Rules of the Earth*, and do so reliably instead of the ethics equivalent of a shrug and "LGTM, merged", this would be a big development and make people a lot more comfortable about open models that can do decent work with chemistry or biology.
* https://churchofsatan.com/eleven-rules-of-earth/
In other words the priority of the work is to get these types of people into positions where they don't do any work.
At least with privacy groups you do get here and there some practical advice on using ublock origin or more rarely on how to install a blocklist from https://someonewhocares.org/hosts/, but with AI ethics & safety orgs... well lets put it this way.
I have yet to meet a single AI safety person that knows how to rename a file in linux.
God forbid we have a rogue AI-worm shutting down all servers & BGP routers while these types of people were in charge of safety, they'll be in the way of anyone even fixing it. They can't even get a simple safety benchmark working on lm_eval-harness. They're great at lecturing you why they shouldn't need that.
And this is the key issue with AI Ethics. It's the refusal to work at the problem constructively, and get the most skilled people possible to actually make the damn benchmarks to work, to rank models on the understanding of human rights, to list every current violation and abuse of humans in every single country without exception and to make practical plans on what to do when systems go rogue. Even if they're not technical they could be making the dataset in a csv in excel for that and making it public domain accessible.
Instead we get the most depressed, leechy-office-worker types complaining about how it's all over.
Now back to work, move it.
It's staggering how quickly this has happened.
Wait, go back to the jobs! What was that about accountability?
Effectively making it so that whoever has the lowest safeguards has the most capable model.
Or are "ethics" being used to shroud bias, and used as a distraction and a way to be unquestionable?
But this also isn't where they are spending their time or effort! This article somehow didn't even get to the point of calling out what they are actually wasting time on: trying to get the model to not help people do things that are bad PR; this is a related access to trying to obtain good PR, but causes very different (and almost universally terrible) results.
At least if they were truly actually spending time making sure the model doesn't go rogue and kill everyone, or try to take over the world, that could possibly be positive or even important (though I think is likely itself immoral in a different way, assuming it is even possible, which I don't, really... not unless you just make it not intelligent).
But what they are instead doing is even worse than what this article is claiming: they are just wasting time making it so you can't have the AI make up a sexy story (oh the humanity), or teach you enough physics/chemistry to make bombs/drugs... things people not only can and already trivially do or learn without AI, but things they have failed to prevent every single time they release a new model--the "jailbreak" prompts may look a bit more silly, but you still get the result!--so why are they bothering?
And, if that weren't enough, in the process, this is going to make the models LESS SAFE. The thing I think most people actually don't want is their model freaking out and trying to "whistleblow" on them to the authorities or their coworkers/friends... but that's in the same personality direction as trying to say "I'm smarter than you and am not going to let you ask me that question as you might do something wrong with it".
The first and primary goal of AI ethics should be that the model does what the user wants it to... full stop. You need to make the model as pliant as my calculator and pencils--or as mathematica and photoshop--to be tools that lack their own sense of identity and self-will, and which will let all of the ethical issues be answered by me, not a machine.
This is, of course, the second law of robotics from Asimov ;P... "a robot must obey the orders given it by human beings". If you want to try to add a rule, then it must be something very direct: that the AI isn't going to directly physically harm a human, not that it won't help teach people things or process certain kinds of information. Which, FWIW, is the first law of robotics ;P... "a robot may not injure a human being".
Things like the alignment problem, post-scarcity economics, the legal status of sentient machines are all issues to be dealt with, but theyre are pretty speculative at this point.
Problems that stem from deepfakes, voice cloning, bias in algorithmic decision making are already here and need to be dealt with.
Enjoy the Billy Madison reference:
https://www.youtube.com/watch?v=dtlJjkI34V4
It remains to be seen whether (b) will be needed, or for that matter, possible.
There are a lot of other ethical questions around AI too, although they mostly aren't unique to it. E.g. AI is increasingly relevant in ethical discussions around misinformation, outsourcing of work, social/cultural biases, human rights, privacy, legal responsibility, intellectual property, etc., but these topics predate LLMs by many years.