Technology

What's the value of old journalism?

It's a product where the value curve is heavily weighted towards recency.

In theory, the greatest value theft is when the AP writes a piece and two dozen other 'journalists' copy the thing changing the text just enough not to get sued. Which is completely legal, but what effectively killed investigative journalism.

A LLM taking years old articles and predicting them until it can effectively learn relationships between language itself and events described in those articles isn't some inherent value theft.

It's not the training that's the problem, it's the application of the models that needs policing.

Like if someone took a LLM, fed it recently published news stories in the prompts with RAG, and had it rewrite them just differently enough that no one needed to visit the original publisher.

Even if we have it legal for humans to do that (which really we might want to revisit, or at least create a special industry specific restriction regarding), maybe we should have different rules for the models.

But to try to claim a LLM that's allowing coma patients to communicate or to problem solve self-driving algorithms or to diagnose medical issues is stealing the value of old NYT articles in its doing so is not really an argument I see much value in.

The solution is imposing to these companies the responsibility of tracking their profit per media, tax them and redistribute that money based on the tracking info. They're able to track all the pages you visit, it's complete bullshit when they say they don't know how much they make for each places their ads are displayed.

but I also don’t like the idea of all the profit going to the ones who didn’t create the product.

Should... should we tell him?

Ai isn't creating the product. It consumed it.

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.

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.

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.

Except the news still needs to come from somewhere. While GPT can "create" things, it's not a journalist. It's just the next step in aggregation skimming money from the actual sources.

This person seems not to know very much about what they are talking about, despite their confidence in saying it.

It looks like they think the reason AI output can't be copyrighted is because it's been "ruled a derivative work" but that's not the reasoning provided which is that copyright can only protect human creativity, and thus machine output without human involvement can't be copyrighted - with the judge noting the line of what proportion of human contribution is needed is unclear.

The other suits trying to claim the models are derivative works are either yet to be settled or in some cases have been thrown out.

Even in one of the larger suits on whether training is infringement regarding LLMs, the derivative claim has been thrown out:

Chhabria, in his ruling, called this argument “nonsensical,” adding, “There is no way to understand the LLaMA models themselves as a recasting or adaptation of any of the plaintiffs’ books.”

Additionally, Chhabria threw out the plaintiffs’ argument that every LLaMA output was “an infringing derivative” work and “constitutes an act of vicarious copyright infringement”; that LLaMA was in violation of the Digital Millennium Copyright Act; and that LLaMA “unjustly enriched Meta” and “breached a duty of care ‘to act in a reasonable manner towards others’ by copying the plaintiffs’ books to train LLaMA.”

• Parts of Sarah Silverman-Led AI Copyright Case Against Meta Dismissed, But Not Core Argument

Social media has really turned into a confirmation bias echo chamber where misinformation can run rampant when people make unsourced broad claims that are successful because they "feel right" even if they aren't.

Perhaps the reason hallucination is such a problem for LLMs is that in the social media data that's a large chunk of their training everyone is so full of shit?