ChatGPT found guilty of fabricating cases and citations for a Manhattan lawyer's Federal...

[RandomWAN]

I think chatgpt should be programmed to random give a 25% false answer rate to teach people to fact check

It's not like they're fact-checking without it. Too many people take somebody else's word for it, especially when they're some established institution, just because they have a job title, or some celebrity.

 

[hwertz]

Yes. I ran the scaled-down chatgpt (gpt4all), which is only a 4GB neural network. I decided to give it exercise when I was at a friends watching movies -- so first I did some of the probing what it thought about it's sentience... if you asked one way, it gave a clearly canned response, if you asked another it was giving the troubling sorts of answers you've seen elsewhere. The fact that it uses 0 CPU time *except* when it's printing an answer does allay my concerns that it's sitting there "thinking" about things while waiting for the human to type in questions though.

Then I asked about the movie we were watching ( Cannibal Women in the Avocado Jungle of Death)-- it knew about that which was surprising because it was not exactly a high-profile movie. Asked about the actors and actresses, it knew about them. Asked it some car questions, it knew about them (got the year wrong by 4 years on one though!) How can it have this much info in 4GB? Then I asked about an actresses -- it completely fabricated a birthdate (making here probably 40 years older than she would have been to be given her age in this movie); a spread in playboy, and a movie career through the 1950s, 60s, and 70s -- I googled just in case there was an actress with a similar name, or any actress that had been in the movies listed (like, it got the name wrong but the facts right for SOME actress.) Nope! All hallucinations.

That's an issue with the current AIs -- they will not admit "I don't know", they will just make up an answer if they don't know. Given the legal precedent type info is usually paywalled, I suspect ChatGPT knows about the big cases (like if you asked about cases regarding segregation it would know Plessy vs. Ferguson upholding "seperate but equal" in the 1890s, and Brown vs. Board of Education abolishing it in the 1950s), but probably no info whatsoever about whatever random bits of case law. The term used by AI researchers is a hallucination, both for the "create an image of xyz from scratch", and the making up info (not just having a fact or two wrong, based on ChatGPT reading a source with a few facts wrong, but whole-sale making it up).

I don't know if the AI knows what it doesn't know (as it were) either. I suspect if you said "Please answer 'I don't know' if you don't have information" would work, or if it just drops through "neurons" based on knowledge of similar topics and hallucinates an answer unaware it's actually being made up.

 

[human7]

Thanks, any other references on how to engage this new tech and deploy privately?

How are you using it currently.

Much appreciated,

It's pretty easy to setup. Just download an installer from here (https://github.com/oobabooga/text-generation-webui), follow the prompts as it downloads the software it needs, and then tell it to download a model. It will present to you a web UI that you can access on localhost:7860 in your web browser, or you can use an API to chat with the model.

A good starter model is this one: TheBloke/wizardLM-7B-GPTQ for Nvidia GPU based setup (https://huggingface.co/TheBloke/wizardLM-7B-GPTQ) or TheBloke/wizardLM-7B-GGML for CPUs (https://huggingface.co/TheBloke/wizardLM-7B-GGML), but there are better and bigger models out there. Make sure you check the license before using it in a commercial setting.

 

[godrilla]

The continuation to this story.
https://nypost.com/2023/06/08/humiliated-lawyer-apologizes-over-chatgpt-court-flub/