AI helper hallucinating a lot when explaining something

Product Support
, ,

AI-generated summary

The discussion revolves around the AI helper’s increasing tendency to hallucinate, particularly when explaining functions or terms. Jagster notes that while the AI helper has been valuable for clarifying concepts, it has recently become unreliable, often fabricating information. This issue may stem from the model or prompting, and despite trying different models (4.0-mini to 4.1-mini), the problem persists. sam suggests that upcoming features will allow users to wire the helper to a dedicated persona, enabling prompt engineering to improve behavior. In the meantime, sam recommends trying Gemini Flash 2.5, which is cost-effective and potentially more accurate. Jagster also observes that the summarizers now default to English, a fixable issue due to prompt editing, but the AI’s tendency to take the easiest path highlights limitations in current models.

1

I can’t pinpoint when AI helper started hallucinating (aka. it doesn’t know and starts to be… creative) because I have had something else to do like three weeks now, but today explaining-function has been totally useless.

I’ve used it a lot earlier to open terms and expressions to readers and for me it has been the most valuable thing what AI helper can offer [1]. Sure, it has have tendency to find context from a post itself, but this amount of hallucination is too much.

It comes from prompting and/or model. I’m a customer of OpenAI, and I don’t think I have will to change, because then I need to edit all prompts too. Self hosted model os not an option, I’m too poor for that.

So, what are my realistic options now? Stop using it?

Edit

I used 4.o-mini and changed it to 4.1-mini. I got better replies, but… it’s common knowledge is quite — is capped right word? It should use google or something else, but then there should be ability to edit prompting, and that won’t be an option, if I’ve understood right earlier discussions about prompting. So is my only option to move more expensive model, if I would stay with OpenAI?

  1. I would like to offer that option to my users too, but I know it is really difficult, because now it needs ability to edit published post. If it could act in composer, that would be a real game changer ↩︎

3

We are only a few more weeks away from allowing you to wire helper to a dedicated persona, at that point you can experiment with prompt engineering to see if you can coax the cheaper models to behave as you want them to.

I would recommend though trying gemini flash 2.5, it is very cheap (and even has a free tier) it should outperform 4.1 mini.

4

Yesterday both Summarizers started use only English on my forum. But because I could edit prompts there was easy fix for single-language forum. In that case prompting itself looked totally right, and I would use such generic wording too, but lazy OpenAI model took the easy road :man_shrugging: