Those claiming AI training on copyrighted works is “theft” misunderstand key aspects of copyright law and AI technology. Copyright protects specific expressions of ideas, not the ideas themselves. When AI systems ingest copyrighted works, they’re extracting general patterns and concepts - the “Bob Dylan-ness” or “Hemingway-ness” - not copying specific text or images.
This process is akin to how humans learn by reading widely and absorbing styles and techniques, rather than memorizing and reproducing exact passages. The AI discards the original text, keeping only abstract representations in “vector space”. When generating new content, the AI isn’t recreating copyrighted works, but producing new expressions inspired by the concepts it’s learned.
This is fundamentally different from copying a book or song. It’s more like the long-standing artistic tradition of being influenced by others’ work. The law has always recognized that ideas themselves can’t be owned - only particular expressions of them.
Moreover, there’s precedent for this kind of use being considered “transformative” and thus fair use. The Google Books project, which scanned millions of books to create a searchable index, was ruled legal despite protests from authors and publishers. AI training is arguably even more transformative.
While it’s understandable that creators feel uneasy about this new technology, labeling it “theft” is both legally and technically inaccurate. We may need new ways to support and compensate creators in the AI age, but that doesn’t make the current use of copyrighted works for AI training illegal or unethical.
For those interested, this argument is nicely laid out by Damien Riehl in FLOSS Weekly episode 744. https://twit.tv/shows/floss-weekly/episodes/744
My point is, that the following statement is not entirely correct:
One obvious flaw in that sentence is the general statement about AI systems. There are huge differences between different realms of AI. Failing to address those by at least mentioning that briefly, disqualifies the author regarding factual correctness. For example, there are a plethora of non-generative AIs, meaning those, not generating texts, audio or images/videos, but merely operating as a classifier or clustering algorithm for instance, which are - without further modifications - not intended to replicate data similar to its inputs but rather provide insights.
However, I can overlook this as the author might have just not thought about that in the very moment of writing.
Next:
While it is true that transformer models like ChatGPT try to learn patterns, the most likely token for the next possible output in a sequence of contextually coherent data, given the right context it is not unlikely that it may reproduce its training data nearly or even completely identically as I’ve demonstrated before. The less data is available for a specific context to generalise from, the more likely it becomes that the model just replicates its training data. This is in principle fine because this is what such models are designed to do: draw the best possible conclusions from the available data to predict the next output in a sequence. (That’s one of the reasons why they need such an insane amount of data to be trained on.)
This can ultimately lead to occurences of indeed “copying specific texts or images”.
It doesn’t matter whether I directly prompted it for it. I set the correct context to achieve this kind of behaviour, because context matters most for transformer models. Directly prompting it do do that was just an easy way of setting the required context. I’ve occasionally observed ChatGPT replicating identical sentences from some (copyright-protected) scientific literature when I used it to get an overview over some specific topic and also had books or papers about that on hand. The latter demonstrates again that transformers become more likely to replicate training data the more “specific” a context becomes, i.e., having significantly less training data available for that context than about others.