On 5/17/26 03:50, pelzflorian (Florian Pelz) wrote:
Hugo Buddelmeijer via "Development of GNU Guix and the GNU System
distribution." <[email protected]> writes:
Every example of LLM's regurgitating copyrighted text starts with a
prompt that is derived from that copyrighted text.
Actually a New York Times freelance journalist using “AI” apparently was
tricked into copying from a similar book review in the Guardian [1].
Presumably this could happen for code.
Interesting, thanks.
It would be nice to have a bit more detail about what the New York Times
journalist actually did. For all we know, the AI agent explicitly
searched for related articles online, directly adding the copyrighted
text to the LLM's context window, (which was my point).
(As in, I suppose journalists have AI tools that do exactly that:
harvest and summarize relevant information from the internet on the fly.)
Such copying is still bad of course, but it is different from the LLM
regurgitating copyrighted texts by itself, which was the subject of
discussion here.
[1]
https://www.theguardian.com/books/2026/mar/31/the-new-york-times-drops-freelance-journalist-who-used-ai-to-write-book-review
