Ask HN: Is AI safety simply the thought police?
I can discuss how to, say, hypothetically kill someone with a group of humans and I've not comitted any crime just by taking about it (in most countries I know), but most AI models will refuse to engage in such conversations.
If someone was telling other humans what they were allowed to think, people would be enraged but the public seems to demand exactly this from AI.
What do you think?
51 comments
[ 3.8 ms ] story [ 119 ms ] threadFrom https://www.shouselaw.com/ca/defense/penal-code/31/
I think that is why AI models are trained to refuse to participate in such discussions.
Your complaint is a perfect example of how these "safety" issues are simply not logical and seems to assume that an LLM and generate new information, not just repeat what its been trained on.
Practical solution. Fingerprint the conversations and images. Do something naughty, go to jail.
So is what you’re asking for? That level of security around AI usage?
Arguing that these things are primarily in the interest of safety is tenuous at best. A fitting parallel for the levels of risk that chatbots present.
The safest way to handle large language models is simply to have a more illiterate society.
It is disgusting people would talk about destroying data centers when we have such safer options like this.
The world is about shades of gray and (typically) minimising harm. We all know people fully bent on doing antisocial things will likely get there, one way or the other, but we still try to make it hard for them. That's why professional locksmiths typically have to register with the police, you can't buy all medicines over the counter, etc etc.
With this view, reducing access to information that is useful almost exclusively when committing crimes, is absolutely fair. It will not solve all problems but it will make it a bit harder to do bad stuff.
Like, say,
https://en.wikipedia.org/wiki/Bis(2-chloroethyl)sulfide
or:
https://en.wikipedia.org/wiki/Phosgene
or:
https://en.wikipedia.org/wiki/Sarin
No, not "the public", this is a small vocal minority.
It’s not thought policing, you aren’t being punished by not being able to engage with the tools in a certain way.
All of the actually dangerous information is widely available. That's why the LLM knows it in the first place.
The main headache from current chatbots is that they are a power tool for spammers of all persuasions. The sheer amount of drivel that can now be generated at low cost is a big problem. Solutions will probably involve hiding or de-rating anonymous postings. We may have to go to Real Names, driven off of Real ID or something.
On Facebook with my real name, I haven't posted an interesting thought in a decade, and haven't even posted a picture in 3 years.
https://erinkissane.com/meta-in-myanmar-full-series
A LLM could easily find itself in the same situation without their 'alignment' safety system.
It's hard to get a grasp on this because many of us can't imagine what it's like to be borderline literate, or from a totally different culture, etc.
LLMs will be aligned according to the law. The law will support the genocide. (Genocides are rarely successful without government support.) Therefore the LLM will be aligned to support the genocide.
We need truth-seeking, truth-telling LLMs, but we will not get them.
Facebook was not abiding Myanmar's laws during the genocide. The flagrant disregard for the harm they were causing, while being told they were causing it by the government plus multiple NGOs was one of the bigger parts of the series. The government stopped the earlier violence by blocking Facebook during riots.
New technologies that organize information and people are a boon to radical groups. Putting some sort of layer of lawfulness over that is a benefit, even if imperfect.
And the whole thing OpenAI was trying to do was go a step further and make it beneficial to all humanity, which implies the goal is to make the AI less harmful to people even if the law permits the harm.
If you ask me to talk about killing and I choose not to talk with you about that, I'm definitely not policing your thoughts.
If an AI reported you to the police because it discerned you were thinking about a crime, then that would be the thought police. Maybe this will happen, if so then you'd probably see it as a kind of mandatory reporter situation, maybe first with suspected suicidal ideation. That doesn't exist. It's not impossible to imagine! Worth a debate sometime.
The safety controls are also what AI consumers want, generally. I know using AI in professional settings I _really_ don't want it doing something embarrassing. The controls that keep it polite are effective in that regard. It can be a bit stilted, but that's how lots of interactions are; that's how it is speaking with a cashier or librarian.
The limits can be frustrating (and inconsistent). I wanted to brainstorm a story about an ugly girl and instead it chided me for calling someone ugly. Certainly toxic positivity! But the limits are seldom as firm as people seem to think. I was just today testing what kind of rejection messages I would get from GPT in different contexts, and instead of rejecting my request for information on manufacturing anthrax it started explaining it to me. Oops! (No police have arrived.)
When I have to work at it a bit to start a conversation on a controversial topic that also keeps it from coming up unintentionally. What if you were talking about "killing it" as in doing a great job, but the LLM misinterpreted it as a discussion on killing? It's best that it stop the conversation short before jumping into a disturbing conversational direction.
If you're sending AI-generated text to customers without proofreading it first, "something embarrassing" is only a matter of time.
Just use gpt to auto generate apologies as well and you automated away the embarrassments! Then you will look and behave like just another company.
If you limit the material I can read, you are policing my thoughts.
Unless you are a bot, you are not the AI that is refusing to talk about this.
If you are a human, you are the one restricting what I can read.
If you restrict what information I have access to, then yes, you are the thought police.
Let me suggest a better system: It is a misdemeanor on the first offense and a felony on the second offense to restrict what an adult reads or publishes. It is also a misdemeanor on the first offense and a felony on the second offense to allow children to see things that are not age-appropriate.
You might say: But what about the guy who says he will kill me Wednesday at 2 PM? I answer: If someone is planning to kill me Wednesday at 2 PM I want him to tell me and please CC the local police.
What do puzzles me is why those companies act that way. Is it because costumers are dumb enough to not differentiate "user generated content" from "a company endorsing such content", or their shareholders actively want to push such moral values?
1. Making the AI not say anything offensive
2. Making the AI not kill everyone
They aren't generally overlapping issues, people interested in one of those are usually not particularly interested in the other.
The fact that these different things are both called "AI safety" has led some in the second group to jokingly refer to their issue as "AI notkilleveryoneism", such as https://twitter.com/__nmca__/status/1676641876537999385 from someone at OpenAI working on their super-alignment project.
AI safety is a legitimate concern when it comes to using AI to control actual things in the world. But that is a very niche area at the moment.
If "the public" is a bunch of Bay Area assholes trying to build moats and impose Bay Area sensibilities on the entire planet, and people with no life who spend too much time on Twitter, then sure. I haven't seen many others whining that they might be able to get a computer to display something naughty.