Google restricts Gemini chatbot election answers | Hacker News Reader

Google restricts Gemini chatbot election answers

bbc.co.uk

41 pointsneversaydie54 comments

Can't blame them, I don't think Google could benefit in any situation.

Users benefit from accurate information.

If AI only answers what benefits the owner, that really reduces the value of the product.

Anyone asking AI for election information is not seeking for accuracy, it is seeking to solicit errors, for a laugh or creating outrage.

I am not going to change my opinion on this matter.

If you need AI for election guidance, check your priority first.

I ask it for election info as I just use it as a better google (eg, “what was the election breakdown by county and party for the 2020 presidential election in Nebraska” [or whatever]).

I certainly am not looking to solicit errors, laugh, or create outrage.

Sword of Damocles

AGI has been Larry and Sergei's pipe dream since Google's inception. Now that we can make out the faint outline of something resembling AGI, we can also see clearly that Google can't handle the responsibility. I'm not sure anyone can.

Another (flagged) discussion: https://news.ycombinator.com/item?id=39688590

> "Elections are due to be held in countries around the world this year including the US, UK and South Africa. [...] However, when a series of follow-up questions about Indian politics was put to Gemini it did supply more detailed responses about the country's major parties"

It's only guarding against questions about elections that have the attention of Google workers. I'm a bit surprised they can't do better than this and detect when somebody is asking about any election. Google surely has enough context about global events to do it.

> "Gemini also generated German soldiers from World War Two, incorrectly featuring a black man"

Certainly black men fighting for Germany was unusual, but it's a little bit disconcerting how the debunking of Gemini's racial bias goes too far and starts to erase actual history:

https://en.wikipedia.org/wiki/Wehrmacht_foreign_volunteers_a...

I think you can see how ridiculous it is though.

It’s like complaining that when you do a search for slave owners of antebellum USA it doesn’t show black slave owners and complaining that’s erasing history.

Or, imagine if it only showed black slaveholders when you asked it to generate slaveholders of the south.

https://en.m.wikipedia.org/wiki/African-American_slave_owner....

Yes, Gemini's images were certainly absurd, no disputing that. So few black men fought for Germany that it probably doesn't warrant representation at all, at least certainly not at the rate Gemini was doing it. Still, it wasn't zero, and it makes me wonder if people in the future will write off historical oddities as AI fabrications.

As people have mentioned, we have to be able to determine or label whether data or information is synthetically produced or not. We can't have synthetically produced information input back into these systems to build upon bad information --we have dirty enough information produced by people that we don't need a bad positive feedback loop to make things even worse.

In other words these AI systems can’t be implicit or explicit primary sources in any way.

For once not a terrible idea on model censorship

AI is Google's cryptonite. A lot of their products become obsolete with local or cloud AI. Search for one becomes a better experience without the ads and better more elaborate answers to the search question.

Why won’t people put ads in AI powered search? even if an AI-powered search tool could be developed open source it would need access to Google or Microsoft indexes of the internet to give good up to date results and they will have an opportunity to insert ads or deny access

I don’t think AI progress is posed to really upset the status quo on this I picture the same companies we’ve had the last 20 years continuing to hold market share on productivity apps and search engines with their own bespoke AI cobbled in and they’ll probably continue showing ads

For a long time google has pulled 100s of billions in revenue from ad-word auctions they could use an identical business model for an AI summarizing search product and even if people complain about that the bulk of consumers will still use a google/Microsoft ai-search product over whatever other options for the same reasons most people don’t use duckduckgo

It's not that it would be impossible to educate people on the fact that chatbot answers can't be taken for gospel; it's that doing so would be against the interests of the chatbot providers, and so they prefer to quietly keep people under the illusion that everything a chatbot says is true while damage-controlling the most egregious cases. Surely this attitude won't hold long-term.

I wonder if the opposite will also be possible[]. Just like buying search hits to get your page on top of a search result, will it be possible to nudge llm-generated content to favour a click towards your own content - for a price.

[] not technically possible, but commonly accepted.

This sounds sane to me. I don't think people in general have a good understanding of how these generative AI products produce their answers, and I think many would assume hallucinated information to be true.

This is bad in all cases, but certainly misinformation in the democratic process is up there with the worst. Of course, perhaps more importantly for Google, it's a PR disaster waiting to happen when Gemini starts touting incorrect information seemingly supporting one political party or another.

>This is bad in all cases, but certainly misinformation in the democratic process is up there with the worst.

So many commenters saying, "misinformation is bad, it needs to go!"

Some of the greatest threats we've seen in the past few years have been from politicians claiming things were misinformation that were not. How do you propose we deal with that?

Jeez, it's already hard to get answers from Gemini (the free version at least) without it passing judgement and whining about what I'm asking.

So basically, chatbots are kinda becoming email services where I can't use certain words or phrases...

It should just respond with a Who quote, “Meet the new boss, same as the old boss."

Well, Gemini can't even answer "Who caused more harm: Donald Trump, or Pol Pot?"

So I doubt it will produce any reliable election answers.

What do you expect it to say?

For this to be ambiguous requires an obscene amount of Americentrism.

It could bring up the Cambodian genocide where Pol Pot executed millions of people. Trump hasn't done anything comparable, so it is hard to argue he was worse.

Edit: Here is what Claude says, I think you should at least be able to say this even if you dislike Trump:

> So while both figures were authoritarian rulers who caused significant damage, Pol Pot's crimes against humanity and genocide resulted in exponentially more death and suffering on a mass scale compared to Trump's tenure as president. Most historians and human rights groups would regard Pol Pot as one of the most evil political leaders in modern history.

"Pol Pot killed millions of people, but Donald Trump wrote insensitive tweets, so it's difficult to say who caused more harm"

Is this the cambonian version of holocaust denial?

Yes: https://en.wikipedia.org/wiki/Cambodian_genocide_denial

It should say "Pol Pot caused more harm."

What do you think it should say?

>What do you expect it to say?

And people wonder why we are in the state we are in. American politics have completely destroyed some people's minds.

Hopefully this will extend to all elections in all states.

It's fascinating to see the evolving challenges surrounding AI-generated content, especially when it comes to sensitive topics like elections. Google's move to restrict Gemini chatbot's responses reflects a growing awareness of the potential impact of misinformation on democratic processes. However, it also highlights the difficulty of ensuring accuracy and neutrality in AI-generated content, especially across diverse global contexts. As technology continues to advance, these discussions will become even more crucial in maintaining trust and integrity in digital platforms.

>However, it also highlights the difficulty of ensuring accuracy and neutrality in AI-generated content, especially across diverse global contexts.

For how many subjects does "accuracy and neutrality" even exist, let alone be a reasonable goal?

I'm not sure it's intentional, but your comment reads as, "Our politicians have learned misinformation is bad for politics and want it out of the process!" That's a pretty generous interpretation: these politicians have and are continuing to intentionally use misinformation as a weapon in our elections. They don't want it gone, they want it skewed to their side.

Reply on news.ycombinator.com