DLF Response to Generative AI and D: An Open Letter

[Indraj Gandham]

On Sunday, 12 July 2026 at 20:06:10 UTC, Adam Wilson wrote:
> Kadrey v. Meta is the most recent settled DMCA 1202 case. 
> Unlike in Bartz where an out-of-court settlement was reached, 
> the plaintiffs in Kadrey lost their DMCA 1202 claims. The 
> ruling held that LLM's were far beyond the threshold for 
> transformative fair use. The Kadrey ruling was issued on June 
> 27th 2026. I read the full ruling and after reading it I doubt 
> there will be any successful 1202 cases, the case law building 
> on this particular legal avenue is robust. The judge 
> recommended another claim they could have tried but didn't. The 
> rulings to date make it clear that the underlying technology of 
> LLM's is inherently transformative (The T in GPT stands for 
> "Transformer").

The problem is that LLMs are not *always* transformative. As 
pointed out in the original letter, LLMs can emit significant 
portions of copyrighted code verbatim. Even if you were to 
implement some kind of automated check for unaltered content, it 
is a well-established legal principle that copying non-literal 
elements of software (such as program structure) may also 
constitute infringement. How can you trust an LLM not to 
replicate patterns from existing software in its training set?

As I pointed out in the other thread, I believe that the case law 
(in the United States) is far from settled, and these rulings 
should be treated as preliminary. I also think that it would be 
unwise to rely on blanket arguments; otherwise there would be 
nothing stopping people from developing new models with the 
specific intent to circumvent copyright using "transformation" as 
a shield, irrespective of how transformative the LLM really was 
in that particular instance.

This is not well-trodden ground and I think it would be prudent 
for users of Phobos to exercise an abundance of caution beyond 
the principles outlined in DLF's response, at least for the time 
being.