The AI dooooooom thread

[HVC]

It's like someone kicking you in the nuts and when you complain they reply "you told me to left my leg, not to left my leg slowly and not kick you in the nuts. Use better prompts next time"

[Jacob]

@dguller - That depends on the crux of the argument being made, surely?

If the argument hinges on what sort of work in writing the prompts is required to achieve or avoid directly plagiarizing copyrighted material, then yes showing the level of prompt engineering involved is important.

But if the argument hinges on whether OpenAI as a product depends on unauthorized commercial use of copyrighted material, then the level of required prompt engineering to achieve this result may be less relevant.

[Sheilbh]

Does the prompt really matter if it's plagiarizing anyway? I mean if your claim is that it plagiarizes then that's clearly true.

It's one part of the claim. The big thrust is exactly that - you've nicked our copyrighted material in order to build your model. In addition to that your product can be used in effect to fully recapitulate our copyrighted material - which is evidence of the fact that you've ingested (without permission) our content to build your model.

The Bing chat stuff interests me because again there's big implications for Google there (and it's interesting no claim against them, yet).

[The Minsky Moment]

I've been at events and met people from the models and whenever I've asked about copyright they've always said they are very very confident on it being fair use in the US.

Seems like bravado more than confidence.

The fair use factors are:  (1) the purpose and character of the use, (2) the nature of the copyrighted work, (3) the amount and substantiality of the original work that is taken, and (4) the effect of the use upon the plaintiff's commercial market.

(4) is still up on in the air, but the others are likely going against Open AI.  Seems like they are hanging their hat a lot on the "transformative" nature of the use, but how transformative is it to take textual information into a database and spit it back out in response to a user query? They may be counting on the courts to buy into their own marketing hype.

[HVC]

Dumb question, would footnotes save them. I guess it wouldn't make the product look good for the market, but would it cover their ass for copywriter purposes?

[DGuller]

Does the prompt really matter if it's plagiarizing anyway? I mean if your claim is that it plagiarizes then that's clearly true.

There are a couple of trivial prompts I can imagine that would "plagiarize" something.  One prompt would involve asking for a Bing search.  Another prompt would have you input the article in prior prompts, and then ask ChatGPT to relay it verbatim.  Both are extreme examples, but examples nonetheless where the screenshotted output would not be what it appears. 

Another reason prompt matters is that it's not in question that NYT articles were used to train ChatGTP; what matters is whether this kind of verbatim plagiarism is going to happen in practice, without long engineering work to make it do something that appears damning.

[Josquius]

The best plagiarism I've heard of is those image ais that include watermarks (eg getty) in their generated images.

[DGuller]

Dumb question, would footnotes save them. I guess it wouldn't make the product look good for the market, but would it cover their ass for copywriter purposes?

Unless ChatGPT directly reads the articles as a result of the prompt and summarizes them, footnotes don't even seem like something that is possible, if my understanding of LLMs is correct enough.  At its core, ChatGPT is like a human that has a memory with a lot of capacity, but it's not a photographic memory. 

All of your knowledge comes from somewhere, but can you really cite where you got most of it?  Some pieces of knowledge you probably do remember where you got it from, especially the more esoteric knowledge, but most knowledge is something that you've synthesized from many sources, and which doesn't match exactly any one source.

[Sheilbh]

Dumb question, would footnotes save them. I guess it wouldn't make the product look good for the market, but would it cover their ass for copywriter purposes?

No but I think it is something that's planned for the search engine replacements (for Bing and Google) which is an AI powered chatbot where you ask a question, get an answer but it will also basically footnote to the original source(s).

It doesn't get round copyright or address other concerns but it is a step in the right direction - it's also probably to help defend Google and Microsoft from any claims around spreading lies because they'll be able to say individuals could always have clicked and checked the original source.

[Josquius]

We've all been getting prompts to write shit for LinkedIns AI generated articles right?

Be sure to write nonsense.

I wrote a bunch of stuff about potatoes in response to wanting me to write an article on more work relevant topics.

[DGuller]

Dumb question, would footnotes save them. I guess it wouldn't make the product look good for the market, but would it cover their ass for copywriter purposes?

No but I think it is something that's planned for the search engine replacements (for Bing and Google) which is an AI powered chatbot where you ask a question, get an answer but it will also basically footnote to the original source(s).

It doesn't get round copyright or address other concerns but it is a step in the right direction - it's also probably to help defend Google and Microsoft from any claims around spreading lies because they'll be able to say individuals could always have clicked and checked the original source.

It can already do footnotes if you ask it to search the Internet.  It will indeed give hyperlinked footnotes in that case.  However, you're not really going off ChatGPT's "memory", though, you're essentially just asking it to summarize something it just read.

[Syt]

The Austrian unemployment agency has introduced a chatbot based on ChatGPT. It's going about as well as you'd expect.

[Jacob]

Michael Cohen gave his lawyer (false) AI generated case citations and his lawyer filed them: https://arstechnica.com/tech-policy/2024/01/michael-cohen-gave-his-lawyer-fake-citations-invented-by-google-bard-ai-tool/

Oopsies

[crazy canuck]

Another lawyer using artificial intelligence and not realizing nonexisting case citations were being created. This time in BC.

https://globalnews.ca/news/10238699/fake-legal-case-bc-ai/

[Jacob]

AI generated spam is apparently reshaping the internet: https://www.businessinsider.com/ai-spam-google-ruin-internet-search-scams-chatgpt-2024-1