I have a simple question. If censorship is considered evil regarding the written word and communications between humans, why do we want to then censor LLMs differently? It is either counterintuitive or simply a false concept we should abandon. Perhaps it is more about training, similar to how children are 'monsters' and need to be socialized/tamed.
Because of the obvious PR implications of having a program one's company wrote spewing controversial takes. That's what it boils down to - and it's entirely reasonable.
Personally, I wish these things could have a configurable censorship setting. Everyone has different things that get under their skin, after all (and this would satisfy both the pro-censor and pro-uncensored groups). It's a good argument for self hosting, too, because those can be filtered to your own sensitivities.
That would help with cases where the censorship is just dead wrong. A friend was working with one of the coding ones in VS Code, and expressed his frustration that as soon as the codebase included the standard acronym for "Highest Occupied Molecular Orbital" (HOMO) it just refused any further completion. We both guessed the censor was catching it as a false positive for the slur.
> We both guessed the censor was catching it as a false positive for the slur.
There is a word for this. It is called the Scunthorpe problem. Named after the incident in which the residents of the Town Scunthorpe could not register for an AOL account because AOL had an obscenity filter that did not allow the Town name.
It has been a problem since 1996 and still causes problems.
That's the right answer. And it's not like this is a potential risk that is only being theorized about. Microsoft already has a very hands-on experience with disasters of this exact nature:
I censor myself all the time when speaking, depending on the context and who I’m speaking to. I do it because it’s typically in my best interest to do so, and because I believe that it is respectful in that moment. The things I say represent me. I don’t find it too surprising that a company might want to censor their own AI product as it represents them and their reputation.
If I make a word processor, it doesn't need any stance on the Israel/Palestine conflict. It's just a word processor.
But if I make an LLM, and you prompt it to tell you about the Israel/Palestine conflict? The output will be deeply political, and if it refuses to answer that will also be political
The technology industry does not know what to do because unlike industries like journalism and publishing who are used to engaging with politics, a lot of norms, power structures and people in tech think we're still in the 1990s making word processors, no politics here.
Government censorship and having policy on what employees/representatives of your companies say are two different things.
There are a lot of things I can say as a citizen that would get me fired from my job, or at least a talking-to by someone in management.
At the moment at least these LLMs are mainly hosted services branded by the companies that trained and/or operate them. Having a Microsoft-branded LLM say something that Microsoft as a corporation doesn't want said is something they will try to control.
That's also different from thinking that all LLMs should be censored. You can train or run your own with different priorities if you wish. It's like how there's a lot of media out there that you can consume or create yourself perfectly legally that isn't sold at Walmart.
> If censorship is considered evil regarding the written word and communications between humans, why do we want to then censor LLMs differently?
Simple: the people who are very pro free speech (i.e. "censorship is evil"), and the people who want to censor LLMs are distinct groups (though both groups are vocal).
They're not being censored by the state in this case (unlike the gotcha everyone keeps using when deepseek is mentioned). They're being limited by their developers for brand safety.
That's easy. The person who used an LLM to generate the text and then published it as their own without even proofreading it is liable for the text they published.
I think that one difference is friction. Communication in the real world takes more effort to spread and has more limitations on the ability to scale that spread. E.g. it costs money to put up billboards and someone standing on a soapbox in the town square can only reach so many people. That friction provides greater opportunities for cooler heads to prevail and greater opportunities for people to counter questionable narratives.
It's somewhat similar to the laws some places have against providing free alcohol. Alcohol is still legal and abuse still happens. However, at least requiring people to spend money provides some friction to prevent things from escalating too much.
Beyond even PR concerns (which you could, in principle, ignore as silly, though of course in practice they are a significant hurdle), you also need to consider that free speech even for people is not absolute. If an LLM responds to a child's query with sexually explicit content, that likely breaks the law, and the company is liable for that. Similarly, if an LLM generates libelous statements about a real person when prompted to describe that person, the company is liable. If an LLM starts generating medical advice or legal advice, that might break certain laws as well (though perhaps some reasonable disclaimers could fix this too).
Because the people in charge are morons and the public is full of morons and the morons in charge are scared that the morons of the public will get mad and make a big deal about some inconsequential bullshit like "your llm said a naughty word!!"
I think you'll find that censorship is actually quite popular. The censors and censorship-hungry know the connotations of the word "censorship." Nobody will say it with their chest they believe censorship is a societal good straight up and make a robust case for it. It's all coy fuckery to lie to us and themselves that they're not censors.
