Their argument is that the copying to their training database is “research”. This would be a legal fair use of unauthorised copying. However, normally with research you make a prototype, and that prototype is distinctly different from the final commercial product. With LLM’s the prototype is the finished commercial product, they keep adding to it, thus it isn’t normal fair use.
When a court considers fair use, the first step is the type of use. The exemptions are education, research, news, comment, or criticism. Next, they consider the nature of the use, in particular whether it is commercial. Calling their copying “research” is a bit of a stretch - it’s not like they’re writing academic papers and making their data publicly available for review from other scientists - and their use is absolutely commercial. However, it needs to go before a judge to make the decision and it’s very difficult for someone to show a cause of action, if only because all their copying is done secretly behind closed doors.
The output of the AI itself is a bit more difficult. The database ChatGPT runs off of does not include the whole works it learned from - it’s in the training database where all the copying occurs. However, ChatGPT and other LLM’s can sometimes still manage to reproduce the original works, and arguably this should be an offense. If a human being reads a book and then later writes a story that replicates significant parts of the book, then they would be guilty of plagiarism and copyright infringement, regardless of whether they genuinely believe they were coming up with original ideas.
Their argument is that the copying to their training database is “research”. This would be a legal fair use of unauthorised copying. However, normally with research you make a prototype, and that prototype is distinctly different from the final commercial product. With LLM’s the prototype is the finished commercial product, they keep adding to it, thus it isn’t normal fair use.
When a court considers fair use, the first step is the type of use. The exemptions are education, research, news, comment, or criticism. Next, they consider the nature of the use, in particular whether it is commercial. Calling their copying “research” is a bit of a stretch - it’s not like they’re writing academic papers and making their data publicly available for review from other scientists - and their use is absolutely commercial. However, it needs to go before a judge to make the decision and it’s very difficult for someone to show a cause of action, if only because all their copying is done secretly behind closed doors.
The output of the AI itself is a bit more difficult. The database ChatGPT runs off of does not include the whole works it learned from - it’s in the training database where all the copying occurs. However, ChatGPT and other LLM’s can sometimes still manage to reproduce the original works, and arguably this should be an offense. If a human being reads a book and then later writes a story that replicates significant parts of the book, then they would be guilty of plagiarism and copyright infringement, regardless of whether they genuinely believe they were coming up with original ideas.