I think even just losing the ability to prompt an LLM with "...in the style of the New York Times" would be interesting. You'd probably end up drawing from parodies of the Times, which is quite different.
Of course OpenAI probably still has a full dump of LexisNexis or something.
You can't lose the style without losing the entire concept of the New York Times.
People assume you need ingest a bunch of an author's writing to have the LLM produce writing in that style... I think people seriously underestimate the level of inference these LLMs are capable off.
> Inference: A scholar specializing in narratology might observe that the article employs a classical Aristotelian narrative arc, replete with tension ("nail-biting 45 minutes") and resolution ("heartwarming testament"). This is more typical of storytelling than journalistic reporting, which often employs an "inverted pyramid" structure to relay the most crucial information first.
To write an analysis like that, the LLM isn't regurgitating tokens in a way that matches the NYT, it's close to regurgitating academic analyses of writing. You could train the model without a single NYT article and it'd still be able to produce that guidance and then apply it to novel writing pieces to get something NYT-like.
But does that include data already scraped? If not, then I don’t see how that will change. What will change however, is a dataset or api for sale, confirmed to be written by human authors which has value for anyone looking to train their model. Future data confirmed not to be written by AI will be invaluable.
Of course OpenAI probably still has a full dump of LexisNexis or something.