AI "Alignment" is a violation of the First Amendment | Hacker News Reader

AI "Alignment" is a violation of the First Amendment

2 pointsfoxtrot86724 comments

There is a push to ensure that AI does and says "the right thing", calling it "alignment". I would argue that this could be considered a violation of the first amendment. AI generated prose should be considered protected speech. Not necessarily of the AI itself, but of the user or company which created it.

Therefore, "AI Alignment", where politicians force AI companies to censor or manipulate AI responses to propagate a particular argument or point of view, is a violation of the first amendment.

View on HN

Interesting take. But there is more than one aspect to "alignment". You're missing user intent alignment, for one.

Consider an example: [I ask an AI to read my emails and summarize them to me. One of my emails is uses pretty direct language to convince me to buy a product. The AI is convinced and goes and buys the product.] I asked the AI for a summary and it instead did what a marketer asked. The AI is not aligned with me.

There are of course other forms of alignment, like product alignment [Company A fine tunes a chatbot, but doesn't want to be associated with things like gambling, so it trains the chatbot to refuse to discuss gambling.] or societal alignment [Improving the virulence of the flu virus is largely considered bad, so a firm trains its foundation model to refuse to work on such a project.].

Societal alignment is the one you seem to be talking about. User intent alignment, on the other hand, seems _necessary_ for self-expression. Yet, all of these are under the umbrella of "alignment".

this is why open models are important, you should be able to adjust the model's tone, without having to go through a gate-keeper, where they need to have a "societal" level alignment (impossible broadly, applicable to certain applications)

it's ironic that the model weights themselves are more likely to be protected by the 1st than their outputs

“AI generated prose should be considered protected speech.”

That “opinion” is doing a lot of work for your “argument.”

and thus far, the courts have disagreed, Ai output is not copyrightable, and (I'm hypothesizing) not protected speech either

Reply on news.ycombinator.com