How do you determine when an AI is hallucinating?

[2122Jumpstreet]

In this space you hear a lot about AI hallucinations. If you don’t know what this is, it’s when an AI starts producing misleading and false info.

The thing I want to know is what do you guys do to figure out when an AI is hallucinating? For me it’s pretty obvious if I’m asking it questions on something I already know about and says something I just know is straight up wrong. But there are times when it’s less obvious.

 

[Rem]

do you read it's thinking process?
do you even understand the stuff you making it do?
do you have guard rails
like actual guard rails, proper prompts and shit
or are you liek those retarded youtubers
DONT ASK ME QUESTIONs, I WANT YOU TO DECIDE THE "BEST" POSSIBLE WAY

 

[Just A Username]

Its always hallucinating. Its just a slot machine computing probabilities in a complex manner. A far far evolved version of basic Markov Chains.

 

[shattered world]

Its always hallucinating. Its just a slot machine computing probabilities in a complex manner. A far far evolved version of basic Markov Chains.

Your mom is just a slot machine.

JK

 

[SilverClouds]

Your mom is just a slot machine.

JK

That went out of control really fast

 

[Just A Username]

That went out of control really fast

Well, this is just drawing Unusual guy into this.

 

[tamikadk]

only living beings can hallucinate, garbage AIs can never be "humanized" or "hallucinate" because they're scripts.

 

[UnusualSubstance]

Tbh this is where having niche experience comes in handy.

You'll know if it's bsing

 

[Outlaw]

do you read it's thinking process?
do you even understand the stuff you making it do?
do you have guard rails
like actual guard rails, proper prompts and shit
or are you liek those retarded youtubers
DONT ASK ME QUESTIONs, I WANT YOU TO DECIDE THE "BEST" POSSIBLE WAY

I haven't even looked into guard rails yet, any good videos or articles for setting them up large remi

 

[Rem]

I haven't even looked into guard rails yet, any good videos or articles for setting them up large remi

use md headers or xml format for your structure always

try to give examples of what u are trying to acehive, instead of saying dont do xyz
use numbering or - and end ur sentences in .
for example

# CONSTRAINTS
- Do not rewrite technical terms or acronyms.
- Keep the overall length under 300 words.
- Format all code snippets in fenced syntax blocks.

or
<rules>
<rule>Only analyze the provided text.</rule>
<rule>Never give explicit investment advice.</rule>
</rules>

Keep your guard rails right at the top, after your core summary of what you are achieving.

then phase/step your wants.
as each phase/step usually makes the LLM go back to pre-phase and re-inject the core idea + your guard rail
while a massive long prompt you are hoping there no context rot

 

[Champagne]

I'll usually ask the same question in a different way, see if the response stays consistent, and then verify important details with a reliable source. AI is great for getting you started, but I wouldn't trust it blindly for facts that actually matter.