man101 posted...
Like it was just making shit up.
The problem isn't that
it
was making shit up. The problem is that it's learning from sources that made shit up. And it has no capacity to differentiate what data is correct and what isn't.
At
best
, you might be able to program an AI to filter out answers depending on how often it reoccurs in sources, but then you're just going to get the
popular
lies.
And you also get the SEO problem of, if I want to increase the odds that your AI will say what I want it to, I just need to make about a thousand different fake articles with slightly varied text, so the AI will read them all as separate sources, compare them against each other, and take whatever it says as fact regardless of how many blatant and ridiculous lies I've packed into it. They all agree, so it
must
be true. And if I'm a corporation that makes a ton of dangerous products, I have a vested interest in using that strategy to convince the AI to tell you that "Explodium Death Gas" is perfectly safe for you to inhale or have in your house, or that it's a 10/10 product that everyone should buy and any negative reviews you might hear are just crazy radical right-wing paranoiacs trying to review bomb.
As you pointed out, AI pulling from Reddit are going to have about the same level of accuracy as your average Redditor. But you'll have a similar problem no matter where you're pulling from, unless you're severely limiting what sources it can pull from, and having a human pre-verify everything for factual accuracy in advance (and even then, you're still going to have human bias gumming up your works).
Every AI is basically as stupid as the stupidest people you allow to feed it.
AI could theoretically be useful on a smaller scale (ie, teach an AI to read everything Shakespeare ever wrote, then use it to look up any passage you want in an instant), but then you're mostly just using it as a search engine anyway, and we've already had those for 30+ years.