ChatGPT Citing Plagiarized Versions of NYT Articles on an Armenian Content Mill | Hacker News Reader

ChatGPT Citing Plagiarized Versions of NYT Articles on an Armenian Content Mill

futurism.com

25 pointsnabla912 comments

In theory if OpenAI decides to respect NYT and never use their content, ignore it completely, how could it know that anything else they see is plagiarizing NYT?

Solving the wrong problem. The big LLM builders need to switch to opt-in (paid licensing deals) instead of opt-out (lawsuits). Exclusive deals if possible for the moat building.

E.g. when running inference, if previous sequence of n tokens can be found in NYT material, decrease the weight of tokens that would result in the sequence continuing to be found from NYT. I imagine this can be done efficiently, if you have the content of NYT available. Presumably that is part of the training material, so this should be possible to do. I think this algorithm could also be improved to deviate more from NYT content.

It would work a bit similarly as llama.cpp's grammar functionality, which permits it to produce only sequences matching a grammar, but in reverse.

It of course remains to be argued if this is still some kind of plagiarization. I mean, it would know exactly what to do to avoid being a direct copy..

Not an opinion on the topic, but technically they could check for copyright issues before inserting it into the training data? How that would work, at those volumes, or if it is feasible, no clue. Have they tried? I doubt it.

They could never economically achieve the required volume of training data from licensed data alone. The model requires plagiarizing the Internet.

What if they do a "clean room approach" of sorts. One non-public LLM is trained on copyrighted material and creates its own articles about stuff. Then a public LLM is trained on these articles. It appears to work for people-based processes.

The problem with that approach is that any biases the original model has will be amplified during the training process, and cause it to have issues.

> It appears to work for people-based processes.

Machines aren't people, and workarounds based on the affordances given to people should not work for machines, because machines should not be given those affordances.

I see no reason to care what affordances someone gives to an (artificial) machine that I don't own.

There are two conditions for AI working at all.

1) Access to all the internet's data without restriction, including all books and research papers.

2) The model isn't controlled and owned by a for-profit company, it has to be an intra-country initiative run for the benefit of mankind.

Anything else is a wild misuse and bastardisation of AI. We're already seeing all the issues inherent in our current approaches.

There could be some methods they can employ, but I don't see any of them being remotely foolproof. E.g. if they ask NYT to provide them with a list of content hashes to check against or the content itself to match similarity, but it's really easy for plagiarisers to use already existing open source LLMs to slightly modify content in different places in automatic fashion. And they would be incentivised to do that as well.

I don't see right now how this battle could be won in theory.

There's going to be bunch of content farms, plagiarisers using their open source LLM automation tools, which essentially launder the information from other sources. Even if you denylist them, it's arbitrary for them to spawn further instances.

Eventually the only way would be to not use the Internet at all, and only use allowlisted sources. But that's quite ridiculous to me.

Not the NYT's problem! Checkmate, AI.

Reply on news.ycombinator.com