5 ms·

The lawsuit itself (which arstechnica links to): https://nytco-assets.nytimes.com/2023/12/NYT_Complaint_Dec2023.pdf https://nytco-assets.nytimes.com/2023/12/NY

The lawsuit itself (which arstechnica links to): https://nytco-assets.nytimes.com/2023/12/NYT_Complaint_Dec2023.pdf https://nytco-assets.nytimes.com/2023/12/NYT_Complaint_Dec20... From page 30 and onwards has some fairly clear examples on how ChatGPT has an (internal) copy of copyrighted material which it will recite verbatim. Essentially if you copy a lot of copyrighted material into a blob and then apply some sort of destructive compression to it. How destructive would that compression have to be for the copyright no longer to hold? My guess it would have to be a lot. As I see it the closeness of OpenAI may be what saves it. OpenAI could filter and block copyrighted material from the LLM from leaving the web interface using some straight forward matching mechanism against the copyrighted part of the data set ChatGPT has been trained on. Whereas open source projects trained on the same data set would be left with the much harder task of removing the copyrighted material from the LLM itself.

jprete 3y ago
The answer to the "closedness" is externally controlled audits.
vanchor3 3y ago
> Essentially if you copy a lot of copyrighted material into a blob and then apply some sort of destructive compression to it. How destructive would that compression have to be for the copyright no longer to hold? My guess it would have to be a lot. I imagine the goal is closer to "enough that no one notices we stole it", either in a way that it's not easily discoverable or even when directly analyzed there's enough plausible deniability to scrape by.
Jackson__ 3y ago
I wonder how they got these results, seeing as they are not showing any of the usual UI's (i.e. ChatGPT/Copilot). It makes it difficult for me to ascertain whether it is repeating from it's training data, or they committed the same mistake as the OP article of using Copilot, which ends up googling(binging?) the article first, before replying.

jprete 3y ago

The answer to the "closedness" is externally controlled audits.

vanchor3 3y ago

> Essentially if you copy a lot of copyrighted material into a blob and then apply some sort of destructive compression to it. How destructive would that compression have to be for the copyright no longer to hold? My guess it would have to be a lot. I imagine the goal is closer to "enough that no one notices we stole it", either in a way that it's not easily discoverable or even when directly analyzed there's enough plausible deniability to scrape by.

Jackson__ 3y ago

I wonder how they got these results, seeing as they are not showing any of the usual UI's (i.e. ChatGPT/Copilot). It makes it difficult for me to ascertain whether it is repeating from it's training data, or they committed the same mistake as the OP article of using Copilot, which ends up googling(binging?) the article first, before replying.