LLMs don’t have any awareness of their internal state, so there’s no way for them to see something as a gap of knowledge.
Eggyhead@lemmings.world 3 weeks ago
It annoys me that Chat GPT flat out lies to you when it doesn’t know the answer, and doesn’t have any system in place to admit it isn’t sure about something. It just makes it up and tells you like it’s fact.
kadup@lemmy.world 3 weeks ago
Doorknob@lemmy.world 3 weeks ago
Took me ages to understand this. I’d thought "If an AI doesn’t know something, why not just say so?“
The answer is: that wouldn’t make sense because an LLM doesn’t know ANYTHING - it’s literally just a pile of words
Electricd@lemmybefree.net 3 weeks ago
Thinking model can realize their prediction doesn’t make sense to an extent but yea, it’s not always accurate
figjam@midwest.social 3 weeks ago
Wouldn’t it make sense for an ai to provide a confidence level though?
I’ve got 3 million bits of info on this topic but only 4 of them lead to this solution. Confidence level =1.5%
kadup@lemmy.world 3 weeks ago
It doesn’t have “3 million bits of info” on a specific topic, or even if it did, it wouldn’t be able to directly measure it. It’s worth reading a bit about how LLMs work behind the hood, because although somewhat dense if you’re new to the concepts, you come out knowing a lot more about what to expect when using them, what the limitations actually are and how to use them better if you decide to go that route.
TechLich@lemmy.world 3 weeks ago
You could do this with logprobs. The language model itself has basically no real insight into its confidence but there’s more that you can get out of the model besides just the text.
The problem is that those probabilities are really “how confident are you that this text should come next in this conversation” not “how confident are you that this text is true/accurate.” It’s a fundamental limitation at the moment I think.
JustTesting@lemmy.hogru.ch 3 weeks ago
It’s always funny to me when people do add ‘confidence scores’ to LLMs, because it always amounts to just adding ‘say how confident you are with low, medium or high in your response’ to th prompt, and then you have made up confidences for made up replies. And you can tell clients that it’s just made up and not actual confidence, but they will insist that they need it anyways…
Eggyhead@lemmings.world 3 weeks ago
And you can tell clients that it’s just made up and not actual confidence, but they will insist that they need it anyways…
That doesn’t justify flat out making shit up to everyone else, though. If a client is told information is made up but they use it anyway, that’s on the client. Although I’d argue that an LLM shouldn’t be in the business of making shit up unless specifically instructed to do so by the client.
Electricd@lemmybefree.net 3 weeks ago
It doesn’t store bits of information. All it has are neurons that form a weighted network
figjam@midwest.social 3 weeks ago
Got it do there is nothing resembling context. Thx.
bestboyfriendintheworld@sh.itjust.works 3 weeks ago
Chat GPT makes up everything it says. It’s just good at guessing and bullshitting.
CosmoNova@lemmy.world 3 weeks ago
It doesn‘t know that it doesn‘t know because it doesn‘t actually know anything. Most models are trained on posts from the internet like this one where people rarely ever just chime in to admit they don‘t have an answer anyway. If you don‘t know something you either silently search the web for an answer or ask.
So since users are the ones asking ChatGPT, the LLM mimics the role of a person that knows the answer. It only makes sense AI is a „confidently wrong“ powerhouse.
WanderingThoughts@europe.pub 3 weeks ago
In the end it’s a word generator that has been trained so much it uses facts often enough to be convincing. That’s its basic architecture.
You can ask it to give a confidence level to have an indication of how sure it is of the answer.
Squizzy@lemmy.world 3 weeks ago
It wouldnt finish a lyric for me yesterday because it was copyrighted. I sid it was public domain and it said “You are absolutely right, given its release date it is under copyright protection”
Wtf
int32@lemmy.dbzer0.com 3 weeks ago
yeah, there are guardrails but for copyright, not for bullshit. ig they think copyrighted content is worse than bullshit.
Electricd@lemmybefree.net 3 weeks ago
From a legal standpoint, yes. Look at trump
melroy@kbin.melroy.org 3 weeks ago
It's a feature. Not a bug of LLMs.
Electricd@lemmybefree.net 3 weeks ago
It’s neither. It’s a design flaw. They’re not designed to be able to handle this type of situation correctly
melroy@kbin.melroy.org 3 weeks ago
Llm is just next word prediction. The Ai doesn't know whether the output is correct or not. If it's wrong or right. Or fact or a lie.
So no I'm not spreading misinformation. The only thing that might spread misinformation is the AI here.
Electricd@lemmybefree.net 3 weeks ago
Saying it’s a “feature” makes it seem like it was intended which is clearly not true.
JayGray91@piefed.social 3 weeks ago
Someone I know (not close enough to even call an "internet friend") formed a sadistic bond with chatGPT and will force it to apologize and admit being stupid or something like that when he didn't get the answer he's looking for.
I guess that's better than doing it to a person I suppose.
bestboyfriendintheworld@sh.itjust.works 3 weeks ago
Chat GPT makes up everything it says. It’s just good at guessing and bullshitting.
BlueCanoe@lemmy.ca 3 weeks ago
That’s actually one thing that got significantly improved with GPT-5, fewer hallucinations. Still not perfect of course
Eggyhead@lemmings.world 3 weeks ago
I’m more inclined to believe it’s gotten better at being convincing.
Electricd@lemmybefree.net 3 weeks ago
Did you try it though?
frog_brawler@lemmy.world 3 weeks ago
It’s pretty much the same shit that some people do when they’re put on the spot.
Awkwardparticle@programming.dev 3 weeks ago
It is a system that outputs an answer that is the most probably correct one from what it processes from the inputs. It does not have the concept of creating a lie. It is just a probability machine.
lichtmetzger@discuss.tchncs.de 3 weeks ago
And depending on how OpenAI tweaked it this time it will either realize its mistake after being made aware of it or double down hard on it.
Eggyhead@lemmings.world 3 weeks ago
I sometimes approach this like I do with students. Using your example, I’d ask it to restate the source, then ask it to read the title of that source directly. If it’s correct, I might ask it to briefly summarize what the source article covers. Then I would ask it to restate what it told me about the source earlier, and to explain where the inconsistency lies. Usually by this time, the AI is accurately pointing out flaws in its prior logic. At that point I ask again if it is 100% sure it didn’t make a mistake, and it might actually concede to having been wrong. Then I tell it to remember how and why it was wrong to avoid similar errors in the future. I don’t know if it actually works, but it makes me feel better about it.
Evotech@lemmy.world 3 weeks ago
It doesn’t admit anything, it’s a language machine