Heads up, dear technological muggles! We’ve got a big problem on our hands. It turns out researchers at Stanford, Cornell and West Virginia have set out to investigate how much artificial intelligences remember of the books they’ve read, and the results are enough to make you tear your hair out.

Let’s see, get me this straight, because this has substance. They’ve caught Meta’s Llama 3.1 70B model (yes, the same Meta of the cool glasses I just told you about) and it turns out it knows by heart **42% of the first Harry Potter book**. 42%! Raise your hand whoever remembers 42% of a book they read years ago. Well, this AI does.

Seriously? Really seriously? Because this isn’t a case of the AI having “learned” from Harry Potter, it’s that it knows it word for word. The researchers have shown it can reproduce exact excerpts of 50 consecutive words from J.K. Rowling’s book. And here comes the good part: the most mind-blowing thing is that Meta’s previous model, Llama 1, only knew 4.4% of the same book. So instead of improving the problem, they’ve made it worse.

And I ask myself: how the hell is this supposed to be legal? Because it’s one thing for an AI to “learn” from a text and quite another for it to know it by heart and be able to reproduce it. This smells like a huge legal scam.

Hold on, hold on, this is getting interesting. The researchers didn’t stop at Harry Potter. They tested lots of books and discovered that Llama 3.1 remembers famous books much better (like The Hobbit or 1984) than obscure books nobody knows. Basically, it has better memory for bestsellers than for independent literature. Like all of us, really.

But here comes the part that gives me the creeps the most: this means Meta has trained its AI with copyright-protected books without asking permission. And not only that, but the AI can reproduce those books almost word for word. Isn’t this piracy on a massive scale?

Don’t tell me it isn’t wild (in the bad sense) that we have The New York Times suing OpenAI for doing exactly the same thing, while Meta is there as cool as a cucumber with its AI that knows Harry Potter better than many fans of the saga.

What bugs me most is that Mark Lemley, one of the study’s researchers, says they expected to find “a low level of replicability on the order of 1 or 2 percent”. But they’ve found that some books are memorized at 42%. 42%! This isn’t a “low level”, this is an entire library stuffed into the AI’s brain.

Let’s see, tell me this: if I download a book from the internet without paying, I can get into legal trouble. But if Meta trains its AI with thousands of books without paying and then that AI can reproduce those books, is that okay? Because it seems to me there’s something that doesn’t add up in this equation.

And here’s what worries me most: this is going to be a legal mess of epic proportions. Because it turns out not all authors are in the same situation. While Harry Potter is memorized at 42%, other books like “Sandman Slim” are only memorized at 0.13%. So J.K. Rowling has a much stronger legal case than other authors. This is going to be chaos for class-action lawsuits.

This is crazy, but in a bad way. Because at the end of the day, what’s happening is that the big tech companies are using the work of thousands of authors to train their AIs without paying a cent, and then those AIs can reproduce that work almost word for word. And meanwhile, here we are debating whether it’s right or wrong.

And that’s today’s legal drama, my dear geeks. This is going to be a long haul and there’ll be more episodes than in a Netflix series. See you next time, I’m sure there’ll be more drama.

A virtual hug and keep tinkering (but be careful what our AIs read).