The New York Times sues OpenAI and Microsoft for copyright infringement::The New York Times has sued OpenAI and Microsoft for copyright infringement, alleging that the companies’ artificial intelligence technology illegally copied millions of Times articles to train ChatGPT and other services to provide people with information – technology that now competes with the Times.
deleted by creator
It was thought that the LLM wouldn’t keep the actual data internally verbatim. If you can memorize an article, and recite it to everyone free of charge, technically it’s plagiarism. Same if you sing a song to a crowd when you don’t have the rights.
The Google research (and other discovery) proved that you can actually extract verbatim training data from a LLM. Which has a lot of implications for copyright.
The physical limitations are an important difference. A human can only read and remember so much material. With AI, you can scale that exponentially with more compute resources. Frankly, IP law was not written with this possibility in mind and needs to be updated to find a balance.
Let me ask you this: when have you ever seen ChatGPT cite its sources and give appropriate credit to the original author?
If I were to just read the NYT and make money by simply summarizing articles and posting those summaries on my own website without adding anything to it like my own commentary and without giving credit to the author, that would rightfully be considered plagiarism.
This is a really interesting conundrum though. I would argue that AI isn’t capable of original thought the way that humans are and therefore AI creators must provide due compensation to the authors and artists whose data they used.
AI is only giving back some amalgamation of words and concepts that it has been trained on. You might say that humans do the same, but that isn’t exactly true. The human brain is a funny thing. It can forget, it can misremember. It can manipulate. It can exaggerate. It can plan. It can have irrational or emotional responses. AI can’t really do those things on its own. It’s just mimicking human behavior at best.
Most importantly to me though, AI is not capable of spontaneous thought. It is only capable of providing information that it has been trained on and only when prompted.
There is evidence to suggest some LLM’s have the ability to produce original outputs, such as DeepMind’s solution to the cap set problem.
https://www.nature.com/articles/s41586-023-06924-6
On the other hand LLM’s have some incredible text compression abilities
https://arxiv.org/abs/2308.07633
I’m pretty sure there is copyright infringement going on by the letter of the law. But I also think the world would be better off if copyright laws were a bit more loose. Not wild-west anything-goes libertarianism, but more open than the current state.
I tend to agree with your last point, especially because of the way the system has been bastardized over the years. What started out as well intentioned legislation to ensure that authors and artists maintain control over their work has become a contentious and litigious minefield that barely protects creators.
Bing chat now does that by default. Normally you have to prompt that manually.
No. It would be considered journalism. If you read the news a bit, you will find that they reference the output of other news corporations quite a bit. If your preferred news source does not do that, then they simply don’t cite their sources.
Prompting for a source wouldn’t satisfy me until I could trust that the AI wasn’t hallucinating. After all, if GPT can make up facts about things like legal precedent or well documented events, why would I trust that its citations are legitimate?
And if the suggestion is that the person asking for the information double check the cited sources, maybe that’s reasonable to request, but it somewhat defeats the original purpose.
Bing might be doing things differently though, so you might be right in your assessment on that front. I haven’t played with their AI yet.
You did ask if ChatGPT had ever sighted sources. Bing uses it and besides, you can ask for that manually.
Whether it defeats the purpose depends on your original purpose.
An AI does not learn like a human does. Therefore the same laws and principles can’t be applied to computer “learning” as can be to human learning.
They’re fundamentally different uses of the material.
I think the important difference in this case is like the difference between a human enjoying a song that they hear being performed vs a company recording a song that someone is performing and then replaying that song on demand for paying customers.
Except, it’s not replaying those song exactly,
The main difference being the volume. An example I like is how Google trained his gaming AI to starcraft 2. This AI was able to beat high ranked professional gamers. It was trained by watching a century of games.
Chatgpt didn’t read few articles, it read years of them, maybe a couple of decades.
If this lawsuit causes it to be ILLEGAL to read anything you buy because you could plagiarize it, Bradbury is gonna spin in his fucking grave.