Search engines don't replicate the content, they index and point to it. When search engines have been caught replicating content they have been sued or had to work out licenses.

Ya that worked out real well for Genius.

Search engine caches work on an opt-out basis, which was found legal in Field v. Google Inc.

I mean, generative AI basically involves turning content into vectors in order to try and find related content. If that's not an index structure i don't know what is.

The technical definition of how something works is often irrelevant in the eyes of the law. What matters is the output, and its effects.

The output of a search engine is nothing like the output of generative AI.

Most ML are certainly compression systems, as you imply. But reproduction and redistribution of copyright material is quite different than generating a service that allows someone to find said material. The question about caching is a different one, but I doubt it falls under the same intent, as Google is incentivized to cache content to better facilitate the search __and redirection__ service, not to facilitate distribution.

How do they make the index without ingesting a copy of the content?

Storing a copy of the content isn't the relevant part. It's what that copy is used for.

In a search engine, it's used to direct traffic to the original author. In Midjourney, it's used to create an alternative to the original author.

I certainly use the blurb in search results at times to get an alternative to the original author (especially if the blurbed content is paywalled).

In the US, that blurb is covered by fair use. Google doesn't regurgitate the entire published work.

These many AI lawsuits will likely settle whether training LLMs falls under fair use.

> Google doesn't regurgitate the entire published work.

Wrong. Here's Google's cache of an entire paywalled article currently on the front page of HN: https://webcache.googleusercontent.com/search?q=cache:c-xOMj...

You can get these by clicking the three dots next to any search result, expanding the options in the pop-over, and then clicking "cached."

I mean, it seems like casds of AI regurtitating entire works verbatim is pretty rare. It happens but is unusual.

And also not really what this is about. If AI was just another libgenisis it would be a lot less controversial.

The deal with search engines was always that you would get traffic out of then.

Use my content, to get people to me. Google's snippets kinda broke that deal and people have indeed complained about that, but otoh you can still technically opt out of being indexed.

It's not clear how you opt out of LLMs.

Google does transfer value back to the websites, by sending them traffic.

However, Google does get a lot of criticism when they do slurp up content and serve it back without sending traffic back to the websites! Yelp and others have testified to Congress complaining about this!

You can remove your work from a search index anytime you want. Not so with work included in a training set