cross-posted from: https://infosec.pub/post/24994013
CJR study shows AI search services misinform users and ignore publisher exclusion requests.
Fixing all the shit AI breaks is going to create a lot of jobs
Not enough
“Your father and I are for the jobs the
asteroidAI will create.”
And then I get down voted for laughing when people say that they use AI for “general research” 🙄🙄🙄
I’ve had people legitimately post the answer they got from chat gpt to answer someone’s question and then get annoyed when people tell them its wrong.
“I’m not sure, but ChatGPT says…”
No, fuck off, go back to grade school.
Only yesterday, I searched for a very simple figure, the number of public service agents in a specific administrative region. This is, obviously, public information. There is a government site where you can get it. However I didn’t know the exact site, so I searched for it on Google.
Of course, AI summary shows up first, and gives me a confident answer, accurately mirroring my exact request. However the number seems way too low to me, so I go check the first actual search result, the aforementioned official site. Google’s shitty assistant took a sentence about a subgroup of agents, and presented it as the total. The real number was clearly given before, and was about 4 times that.
This is just a tidbit of information any human with the source would have identified in a second. How the hell are we supposed to trust AI for complex stuff after that?
The AI models can be hilariously bad even on their own terms.
Yesterday I asked Gemini for a population figure (because I was too lazy to look it up myself). First I asked it:
what is the combined population of place 1, place 2 and place 3
It answered:
Unfortunately, providing an absolutely precise, up-to-the-minute combined population figure is difficult. Population numbers are constantly changing.
To find these figures, I would recommend checking the Office for National Statistics (ONS) website. They provide detailed population statistics for local authorities in England.
On a whim, I asked it again as:
combined population of place 1, place 2 and place 3
And then it gave me the answer sweet as a nut.
Apparently I was being too polite with it, I guess?
Gotta let it take the W on that first answer, honestly.
I slapped a picture of a chart into Gemini because I didn’t know what the type of chart was called but I wanted to mention it in a Uni report. I was too lazy to go looking at chart types and thought that would be quicker.
I just asked it “What kind of chart is this” and it ignored that and started analysing the chart instead and started stating what the chart was about and giving insights into the chart. Didn’t tell me what kind of chart it was even though that was the only thing I asked.
Bear in mind that I deliberately cropped out any context to avoid it trying to do that, just in case, so all I got from it was pure hallucinations. It was just making pure shit up that I didn’t ask for.
I switched to the reasoning model and asked again, then it gave me the info I wanted.
While I do think that it’s simply bad at generating answers because that is all that’s going on, generating the most likely next word that works a lot of the time but then can fail spectacularly…
What if we’ve created AI but by training it with internet content, we’re simply being trolled by the ultimate troll combination ever.
This is what happens when you train your magical AI on a decade+ of internet shitposting
They didn’t learn from all the previous times someone tried to train a bot on the internet.
It’s almost poetic how Tay.ai, Microsoft’s earlier shitty ai, was also poisoned by internet trolling and became a Nazi on twitter nearly a decade ago
I searched for pictures of Uranus recently. Google gave me pictures of Jupiter and then the ai description on top chided me telling me that what was shown were pictures of Jupiter, not Uranus. 20 years ago it would have just worked.
Stupid that we have to do this, but add
before:2022and it filters out all the slop
I’m shocked!
Shocked I tell you!
Only 60%‽
Blows my mind that it’s so low.
Training AI with internet content was always going to fail, as at least 60% of users online are trolls. It’s even dumber than expecting you can have a child from anal sex.
Because of what you just wrote some dumb ass is going to try to have a child through anal sex after doing a google search.
They’re not joking about a hypothetical. It was a real thing that happened.
they’ve been having sex the wrong way
that’s subjective
I’m gonna go ahead and try without a Google search.
I believe in you, please name your child after me if it works out.
Know that If it doesn’t work, I’m not giving up.
I believe in you, if you end up having twins please name them after this instance
Well, that’s less bad than 100% SEO optimized garbage with LLM generated spam stories around a few Amazon links.
Who could have seen this coming? Definitely not the critics of LLM hyperscalers.
