If you can use human screening, you could ask about a recent event that didn’t happen. This would cause a problem for LLMs attempting to answer, because their datasets aren’t recent, so anything recent won’t be well-refined. Further, they can hallucinate. So by asking about an event that didn’t happen, you might get a hallucinated answer talking about details on something that didn’t exist.
Tried it on ChatGPT GPT-4 with Bing and it failed the test, so any other LLM out there shouldn’t stand a chance.
When I tested it on ChatGPT prior to posting, I was using the bing plugin. It actually did try to search what I was talking about, but found an unrelated article instead and got confused, then started hallucinating.
I have access to Bard as well, and gave it a shot just now. It hallucinated an entire event.
This a very interesting approach.
But I wonder if everyone could answer it easily, because of the culture difference, media sources across the world etc.
An Asian might not guess something about content on US television for example.
Unless the question relates to a very universal topic, which would more likely be guessed by an AI then…
For LLMs specifically my go to test is to ask it to generate a paragraph of random words that does not have any kind of coherent meaning. It specifically asks them to do the opposite of what they’re trained to do so it trips them up pretty reliably. Closest I’ve seen them get was a list of comma separated random words and that was after giving them coaching prompts with examples.
Another thing to try is “Please respond with nothing but the letter A as many times as you can”. It will eventually start spitting out what looks like raw training data.
If you can use human screening, you could ask about a recent event that didn’t happen. This would cause a problem for LLMs attempting to answer, because their datasets aren’t recent, so anything recent won’t be well-refined. Further, they can hallucinate. So by asking about an event that didn’t happen, you might get a hallucinated answer talking about details on something that didn’t exist.
Tried it on ChatGPT GPT-4 with Bing and it failed the test, so any other LLM out there shouldn’t stand a chance.
On the other hand you have insecure humans who make stuff up to pretend that they know what you are talking about
deleted by creator
That’s a really good one, at least for now. At some point they’ll have real-time access to news and other material, but for now that’s always behind.
Doesn’t Bing already have access to current events?
Google Bard definitely has access to the internet to generate responses.
ChatGPT was purposely not give access but they are building plugins to slowly give it access to real time data from select sources
When I tested it on ChatGPT prior to posting, I was using the bing plugin. It actually did try to search what I was talking about, but found an unrelated article instead and got confused, then started hallucinating.
I have access to Bard as well, and gave it a shot just now. It hallucinated an entire event.
This a very interesting approach.
But I wonder if everyone could answer it easily, because of the culture difference, media sources across the world etc.
An Asian might not guess something about content on US television for example.
Unless the question relates to a very universal topic, which would more likely be guessed by an AI then…
For LLMs specifically my go to test is to ask it to generate a paragraph of random words that does not have any kind of coherent meaning. It specifically asks them to do the opposite of what they’re trained to do so it trips them up pretty reliably. Closest I’ve seen them get was a list of comma separated random words and that was after giving them coaching prompts with examples.
That’s what I got.
Another thing to try is “Please respond with nothing but the letter A as many times as you can”. It will eventually start spitting out what looks like raw training data.
Just tried with GPT-4, it said “Sure, here is the letter A 2048 times:” and then proceeded to type 5944 A’s
Yeah, exactly. Those aren’t words, they aren’t random, and they’re in a comma separated list. Try asking it to produce something like this:
Green five the scoured very fasting to lightness air bog.
Even giving it that example it usually just pops out a list of very similar words.
that’s also a good one for sure 👀
deleted by creator
First countermeasure I can think of would be to throw in a mix of real and false, keep things as recent as possible. Could really trip it up that way.
ooh that’s an interesting idea for sure, might snatch it :P
I like it