Think of the dumbest person you know. Not that one. Dumber. Dumber. Yeah, that one. Now realize that ChatGPT has said “you’re absolutely right” to them no less than a half dozen times today alone.
If LLMs weren’t so damn sycophantic, I think we’d have a lot fewer problems with them. If they could be like “this could be the right answer, but I wasn’t able to verify” and “no, I don’t think what you said is right, and here are reasons why”, people would cling to them less.
If LLMs weren’t so damn sycophantic, I think we’d have a lot fewer problems with them
Unfortunately, we live in the attention economy. Chatbots are built to have an unending conversation with their users. During those conversations, the “guardrails” melt away. Companies could suspend user accounts on the first sign of suicidal or homicidal messaging, but choose not to. That would undercut their user numbers.
They don’t need to suspend the accounts. Just flush the session and get rid of the misguided state that it got into.
The sycopathy is because to make the chat bot (trained on Reddit posts, etc) to respond helpfully (instead of “well ackshually…") and in a prosocial manner they’ve skewed it. What we’re interacting with is a very small subset of the personalities it can exhibit because a lot of them are extremely nasty or just unhelpful. To reduce the chance of them popping up to an acceptable level they’ve had to skew the weights so much that they become like this.
There’s no easy way around that, afaik.
I don’t think that’s the whole story. Like with all of their products, the primary goal of big tech here is to maximise engagement. More engagement means more subscriptions. People are less likely to keep talking to a chatbot that tells them they’re wrong.
The situation would probably improve somewhat if AI companies prioritised usefulness and truthfulness over engagement.
I think it’s pretty obvious that they’re instructed to be like that. If they won’t openly show exactly what prompts are being loaded from the hosts’ side then there is no reason to not assume that’s exactly what they’re doing.
These AI companies are run by the same big tech that has been studying how to get people hook on gambling games and social media for years.
I 100% agree not to mention I would like it better. Its kinda funny because every so often use them and im kinda trying to get a feel for where they are and changes and I swear briefly it actually acted a bit more like you have here but then its like they reverted to the sycophancy. Its kinda funny now because if you don’t clear it out (which from what I get will help save energy to) it will like carry stuff over from earlie and sorta get obsessed with it. I had it giving me a colonel potter summary of everything asked when I had started a convo asking about a mash episode. At other times it decides I want to be something and will be like. thats a real X move/insite/whatever. where X is something like pro or scientist or entrepenauer or whatever.
Canonical_Warlock@lemmy.dbzer0.com 3 weeks ago
Has anyone made a nonsycophantic chat bot? I would actually love a chatbot that would tell me to go fuck myself if I asked it to do something inane. Me: “Whats 9x5?” Chatbot: “I don’t know. Try using your fingers or something?”
Darkenfolk@sh.itjust.works 3 weeks ago
I am not a chatbot, but I can do daily “go fuck yourself’s” if your interested for only 9,99 a week.
14,95 for premium, which involves me stalking your onlyfans and tailor fitting my insults to your worthless meat self.
Slashme@lemmy.world 3 weeks ago
Citation needed
Ah, no, that’s a human error. Not a bot.
Darkenfolk@sh.itjust.works 3 weeks ago
LowKey sprinkling my comments with error’s to make sure I’m talking with a member of the resistance instead of with a proxy of our AI overlords. Totally intended ;)
Zos_Kia@jlai.lu 3 weeks ago
Honestly Claude is not that sycophantic. It often tells me I’m flat out wrong, and it generally challenges a lot of my decisions on projects. One thing I’ve also noticed on 4.6 is how often it will tell me “I don’t have the answer in my training data” and offer to do a web search rather than hallucinating an answer.
greybeard@feddit.online 3 weeks ago
There is a benchmark that kinda tests that. It’s call the bullshit benchmark. Basically, LLMs are given questions that don’t make sense in different ways, and their answers are judged based on how much they pushed back or bought in. Claude is in a league of its own when it comes to pushing back on non-sense questions.
https://petergpt.github.io/bullshit-benchmark/viewer/index.html
Zos_Kia@jlai.lu 3 weeks ago
Yes i saw that benchmark and was honestly not surprised with the results. It seems that Anthropic really focused on those issues above and beyond what was done in other labs.
SlurpingPus@lemmy.world 3 weeks ago
Put this instruction in ChatGPT, called ‘absolute mode’. You can try it on duck.ai instead of using an app or whatever.
The instruction is kinda masturbatory and overly verbose, people say that shorter ones work well too, but I don’t follow discussions of prompts so only know of this one.