Federal court says training AI on books is fair use, but pirating them isn't. The Anthropic decision reshapes copyright law for the AI industry, with implications for authors and content creators.
I wonder if authors going forward could simply be given the opportunity to grant permission to access their books for AI. Not talking about simply purchasing the book but deciding whether or not to restrict the book in perpetuity. Might sound like self sabotage but some authors may choose that. Or do publishers ir authors set a particular price for AI access. This ruling makes sense but it may also open up new business models for publishing that put more money in the pockets of authors.
This kind of makes sense. But it doesn't take into account that AI could now be used for pirating books. There's nothing to stop someone from prompting "Claude, get me the text of the first chapter of XXXX".
It's not legal to go to a public library and photocopy a book, and I imagine if a library enabled this, it would be liable. Why should the AI companies not be liable too?
Well part of why that doesn’t work is AI is actually transformative and is unlikely to be able to actually recall correctly word for word. Model makers are also putting safeguards there to prevent that happening in the first place.
I will say that the thing that I'm struggling with the most from the decision is that Judge Alsup is saying that the entirety of "Claude" is the equivalent of one "student".
No matter how many GPUs.
No matter how many Models.
The trained model as a whole - Claude - is a single learner. Kind of warps my mind.
That is a good way to put it. Mind-warping. Not sure how the judge is going to rule otherwise? But also it doesn’t make intuitive sense to call Claude singular?
Nate. I'm really enjoying your podcast overviews of your written articles. However, this seems to prevent me from doing what I previously loved which was getting the Substack iOS app to read me your articles while I'm driving in the car (or working out, walking the dog, cooking dinner, etc). Have I missed something in the app UI or does the presence of a podcast block text-to-speech for the article? Would love to know.
I had to look this one up! It looks like Substack only permits one audio track per post at this time, so that’s what’s going on. Not sure I have an easy way to fix that one.
That said, you can get iPhone to read it to you using the accessibility settings and selecting read the screen.
Thanks for taking the time to look into this, Nate. And to explore options. I get being able to try alternative text to speech approaches and I might take a look at elevenlabs. However, the reality is that nothing is likely to be quite as easy to use as the built in Substack feature. So if you are able to put the audio elsewhere that would be wonderful.
This makes so much sense to me. It's how humans learn, too. We read great authors, or study favorite artists who painted what we love, or learn to mimic Eddie Van Halen as kids when we wanted to play guitar in a band.
What the judge didn't understand is that a human tha pays and read a book has limited bandwidth in generating content whilst a company has unlimited bandwidth. So it is unfair for a company to just pay for 10 pound bound where they can product millions per seconds similar to it whereas a human probably can make use of it a few times per day.
Much less worried about copyrights than than how AI be used to end human rights. (I know, different topic, but when else will my meatbag brain get to make that quip?)
I wonder if authors going forward could simply be given the opportunity to grant permission to access their books for AI. Not talking about simply purchasing the book but deciding whether or not to restrict the book in perpetuity. Might sound like self sabotage but some authors may choose that. Or do publishers ir authors set a particular price for AI access. This ruling makes sense but it may also open up new business models for publishing that put more money in the pockets of authors.
Not sure how that would be enforceable but I bet we have conversations around it yeah
This kind of makes sense. But it doesn't take into account that AI could now be used for pirating books. There's nothing to stop someone from prompting "Claude, get me the text of the first chapter of XXXX".
It's not legal to go to a public library and photocopy a book, and I imagine if a library enabled this, it would be liable. Why should the AI companies not be liable too?
Well part of why that doesn’t work is AI is actually transformative and is unlikely to be able to actually recall correctly word for word. Model makers are also putting safeguards there to prevent that happening in the first place.
I will say that the thing that I'm struggling with the most from the decision is that Judge Alsup is saying that the entirety of "Claude" is the equivalent of one "student".
No matter how many GPUs.
No matter how many Models.
The trained model as a whole - Claude - is a single learner. Kind of warps my mind.
That is a good way to put it. Mind-warping. Not sure how the judge is going to rule otherwise? But also it doesn’t make intuitive sense to call Claude singular?
wonderful update ... I always look forward to your content. Thank you!
glad it was helpful!
Nate. I'm really enjoying your podcast overviews of your written articles. However, this seems to prevent me from doing what I previously loved which was getting the Substack iOS app to read me your articles while I'm driving in the car (or working out, walking the dog, cooking dinner, etc). Have I missed something in the app UI or does the presence of a podcast block text-to-speech for the article? Would love to know.
I had to look this one up! It looks like Substack only permits one audio track per post at this time, so that’s what’s going on. Not sure I have an easy way to fix that one.
That said, you can get iPhone to read it to you using the accessibility settings and selecting read the screen.
If you want to get slightly fancier you can also get an audio file very quickly via eleven labs: https://elevenlabs.io/blog/how-to-convert-text-to-mp3
Makes me wonder if I should stick the audio files somewhere
Thanks for taking the time to look into this, Nate. And to explore options. I get being able to try alternative text to speech approaches and I might take a look at elevenlabs. However, the reality is that nothing is likely to be quite as easy to use as the built in Substack feature. So if you are able to put the audio elsewhere that would be wonderful.
Keep up the great work.
I will see what I can do
Thank you sir
Great breakdown. Very informative and insightful. Thank you!
glad it helped! It’s a big deal
This makes so much sense to me. It's how humans learn, too. We read great authors, or study favorite artists who painted what we love, or learn to mimic Eddie Van Halen as kids when we wanted to play guitar in a band.
Of course AI would learn the same.
yeah it’s always felt like the rational approach to how we handle AI legality—but I’ve learned not to expect the rational approach from the courts lol
Not lately, for sure.
What the judge didn't understand is that a human tha pays and read a book has limited bandwidth in generating content whilst a company has unlimited bandwidth. So it is unfair for a company to just pay for 10 pound bound where they can product millions per seconds similar to it whereas a human probably can make use of it a few times per day.
Loving your content, Nate. Thanks.
Could this also cover music?
So what options do authors have to expliciy prohibit ingestion by AI companies?
And btw - how could anyone know or control whether information was recreated verbatim?
Much less worried about copyrights than than how AI be used to end human rights. (I know, different topic, but when else will my meatbag brain get to make that quip?)