India confronts Google over Gemini AI tool's 'fascist Modi' responses | Hacker News Reader

India confronts Google over Gemini AI tool's 'fascist Modi' responses

theguardian.com

32 pointsbeardyw16 comments

It's going to be hard to argue that section 230 applies to the output of LLM (and therefore Google is not liable) while at the same time arguing that the LLM are transformative and therefore what comes out isn't protected by copyright laws.

Would libel laws apply to LLM, since it is a private chat and not a public declaration?

Does india have an equivalent law to section 230?

Any ideas how they built "wokeness" into gemini AI. I've seen ppl suggest that its simply appending "and make this more diverse" to the user prompt? Is that all they are doing to generate "nazi soldiers are asian" pictures ?

That was what I thought too, but it has to be more involved than that since otherwise fixing the model would just involve removing the prompt modification stuff, which shouldn't take the months I've been hearing it'll take to fix.

My guess is several rounds of RLHF or fine-tuning with some kind of curated 'woke' dataset. Alternatively, they trained a critic model that penalizes showing white people or other 'bad' representations and then fine tuned using that.

How is this woke? It is not an unreasonable response to classify Hindu nationalism as fascist, particularly given fascism is hard to define.

> particularly given fascism is hard to define

Than don't throw around "fascist" as an insult, which seems to be incredibly common.

As an aside, to judge whether something is fascist or not going through the 14 points made by Umberto Eco helps a lot: https://www.openculture.com/2016/11/umberto-eco-makes-a-list...

Umberto Eco grew up during italian fascism so he kinda knew.

I agree. It is quite worry that the word fascist is used so loosely because we forget what it really means. Even worse, the same is happening to the better defined term Nazi.

If it were just Modi, perhaps, but the model is now well-known for generating all kinds of politically controversial / flat out false things.

That's pretty much all generative AI does. It makes things up by design

Most companies are attempting to fine-tune their models to output truth, or putting in place external systems that pick answers that are true.

The BJP ruling party and some of their politicians is clearly aligned with the goals of the RSS (Rashtriya Swayamsevak Sangh) a facist organization - you don't need to be woke but just read uncensored Indian press and see their gatherings where they prattle unsavory things about Indians who are Muslims.

Reply on news.ycombinator.com