Conversations

Note: This part two of three related posts.

In a prior conversation you said something that I have been thinking a lot about: “The organizations opposing AI often seem to assume that using copyrighted material to train AI is fundamentally different from the way human beings learn. But consider a novelist. That novelist may have read thousands of books, absorbed countless literary techniques, learned from other authors’ styles, and incorporated all those influences into an original work. Nobody expects the novelist to compensate every author whose writing influenced his development.”

Here’s my thoughts: In the past, most of what I’ve learned came from authors who were compensated. I learned from school textbooks, newspapers, magazines, and books. In almost 100 percent of these cases, the authors were compensated by me or the school or someone purchasing the media. The same can be said about video/movies either by my direct purchase or through paid advertising, etc. In fact, it is difficult to imagine learning in the past without compensating the creator.

ChatGPT said:

You’ve identified an important weakness in the analogy I offered between human learning and AI training. And I think your observation deserves some exploration because it introduces an economic distinction that my original argument overlooked.

My argument concerned how learning occurs. Yours concerns how the material used in learning is acquired. Those are two quite different questions.

1. Human learning has historically supported a system of compensation

Consider your own experience. Over a lifetime, you’ve acquired knowledge from textbooks, newspapers, magazines, books, television, and movies. In most cases, someone paid for the material. Sometimes it was you, sometimes your school, sometimes an advertiser, and sometimes a library supported by taxpayers. The important point is that there was generally an economic relationship between the consumption of intellectual property and the people producing it. That relationship supported an entire ecosystem of writers, editors, publishers, journalists, filmmakers, and educators.

And there’s another important consideration: The more people who wanted to learn from a particular work, the more economic value that work could potentially generate. A successful textbook might sell hundreds of thousands of copies. A popular newspaper attracted subscribers and advertisers. A novelist whose books influenced generations of writers might continue selling books for decades.

The act of learning wasn’t itself compensated, but access to the material frequently was.

2. AI potentially breaks that relationship

Now consider an AI company that acquires millions of books and uses them to train a model. Once trained, that model can potentially convey information derived from those books to millions of users without those users ever purchasing the original works.

The company may generate substantial revenue from the resulting model, while the original creators receive nothing beyond whatever compensation they received when the training material was initially acquired.

This is economically different from the traditional model.

But there’s a complication.

Suppose a novelist buys a thousand books, learns from them, and writes a successful novel. The authors of those thousand books receive no additional compensation from the novelist’s success.

Now suppose an AI company legally purchases a thousand books, trains a model on them, and sells access to that model.

Why should the second case require additional compensation when the first does not?

The answer isn’t nearly as obvious as some copyright advocates suggest.

3. I think the strongest distinction is scale, not the nature of learning

A human novelist might read a few thousand books over a lifetime. An AI system can process millions of books in a relatively short period. More importantly, that system can subsequently serve millions of people simultaneously. This introduces the possibility that AI could displace some of the very markets that compensated the original creators.

For example, a person who once purchased several reference books might instead ask AI for the information. A student who once bought a study guide might use AI instead.

Of course, technological innovations have displaced existing markets throughout history. The printing press, photography, recorded music, and the internet all disrupted established economic relationships.

And we generally haven’t concluded that preserving the previous business model is sufficient justification for preventing technological progress.

Nevertheless, the economic consequences deserve consideration.

4. But your observation also raises a question about what constitutes compensation

You suggest that nearly everything you learned from involved compensation to its creator.

I think that’s substantially correct, particularly for the period in which you grew up.

But we should distinguish three things:

• Compensation for access: Someone pays to obtain a book, newspaper, or movie.
• Compensation for learning: Someone pays the creator because knowledge was acquired from the work.
• Compensation for subsequent use: Someone pays the creator because that acquired knowledge contributed to another commercially valuable product.

Historically, copyright has primarily protected certain uses of creative expression, not the act of learning from it.

And the second and third categories have generally not required compensation.

That distinction matters because purchasing a book has never meant purchasing an obligation to compensate its author for everything you subsequently learn or create.

It also means that legally obtaining a book and having the legal right to make the copies involved in AI training are separate questions.

5. There’s another possibility that I find particularly interesting

Your observation may actually strengthen a different argument for compensating creators.

Perhaps the appropriate question isn’t whether AI learning resembles human learning. Perhaps it is: Should commercial enterprises be permitted to build enormously valuable products from intellectual property without participating economically in the system that produces that property?

That’s a different question, and potentially a more compelling one. It doesn’t require us to pretend that AI learning is fundamentally unlike human learning. Nor does it require us to accept that every use of copyrighted material should generate a royalty. Instead, it asks whether the economic system supporting intellectual creation needs to evolve alongside the technology.

But there is a danger here, too. If we impose compensation requirements that only the largest AI companies can afford, we could inadvertently protect those companies from smaller competitors. And if the requirements are too restrictive, we could slow innovation without meaningfully improving creators’ livelihoods.

6. One qualification to your historical observation

Although I agree with your general premise, I’d hesitate over the suggestion that virtually all human learning has involved compensation. We also learn through conversation, observation, experimentation, imitation, and the exchange of ideas. Much of our knowledge comes from people who receive no direct payment for teaching us.

Even with books, the connection between learning and compensation can be tenuous. A library might purchase one copy that hundreds of people read. A used book can pass through many hands without generating additional revenue for its author. So the historical relationship between learning and compensation has always been imperfect. What AI changes is the magnitude of that imperfection.

Where I think your argument leads

I see two propositions emerging from our discussion.

First: There is no obvious philosophical reason to treat knowledge acquired by a machine as fundamentally different from knowledge acquired by a human merely because the learning mechanisms differ.

Second: There may nevertheless be legitimate economic reasons to distinguish between the two, particularly when AI systems can reproduce some of the economic value traditionally generated by the original works.

These propositions aren’t contradictory.

And that brings me to something I’d like to ask you.

Suppose an AI company purchased, at ordinary retail prices, every book it used for training. Would that satisfy your concern about compensating creators?

Or do you believe that because the company subsequently earns money from what it learned, the original authors deserve a share of that revenue?

I ask because your answer would help distinguish whether your concern is fundamentally about fair acquisition of knowledge or sharing the economic benefits generated by that knowledge.

Those lead to very different approaches to copyright reform.