But those links are Googled after the model started to answer, they are not the links to the training data
Imagine an artificial “librarian” that read all the books and spits hallucinated quotes for you
But doesn’t let you enter the library, open a single book or even see the sources for those hallucinated quotes
But instead Googles some sources based on hallucinations after generating them ;-)
It’s better than nothing but you can Google them, too, while training data (the library) is completely hidden from you, even the public domain parts of it - zero attribution
There should be at least some correlation. When building the model they give more weight to some pages (e.g. Wikipedia) which have bigger trust (pagerank?). And when they provide links in answers, those matches are listed first which have better pagerank for the query.
So if it sources something in Wikipedia, it is more likely to provide Wikipedia as a trusted source for it.
The problem is when an answer is hallucinated, false, it may provide a source for it which contains the invalid info.
Comments
Secifically in Google AI Overview I always see links to sites where the information is sourced from.
Or at least some of the sites, if the same info is sourced from 100 pages then it only shows 2 or 3, maybe the ones with the biggest PageRanks.
Yep, that’s true
But those links are Googled after the model started to answer, they are not the links to the training data
Imagine an artificial “librarian” that read all the books and spits hallucinated quotes for you
But doesn’t let you enter the library, open a single book or even see the sources for those hallucinated quotes
But instead Googles some sources based on hallucinations after generating them ;-)
It’s better than nothing but you can Google them, too, while training data (the library) is completely hidden from you, even the public domain parts of it - zero attribution
There should be at least some correlation. When building the model they give more weight to some pages (e.g. Wikipedia) which have bigger trust (pagerank?). And when they provide links in answers, those matches are listed first which have better pagerank for the query.
So if it sources something in Wikipedia, it is more likely to provide Wikipedia as a trusted source for it.
The problem is when an answer is hallucinated, false, it may provide a source for it which contains the invalid info.
Yep, a few non-profits work on direct training data attribution:
OlmoTrace, Guide Labs with Clarity and a few more
Labs train the model with attribution baked-in and they say the bigger the model - the more interpretable it becomes
Pretty sure it’s the future