Magistrate Judge van Keulen agreed with Anthropic that it should be allowed discovery into Concord Music’s attempt to get Claude to produce infringing lyrics.
The Judge ruled: “The Court hereby GRANTS Anthropic’s request and orders Publishers to disclose to Anthropic (a) the total number of the Publishers’ (or their agents’) prompt-output pairs that are encompassed in the set of 15,000 prompt-output pairs; (b) the number of such prompts in that set that resulted in the output of song lyrics; and (c) the number of such prompts in that set that resulted in the invocation of Claude’s guardrails or otherwise failed to elicit song lyrics.”
But the Judge also ruled that Anthropic cannot ask Concord Music about these numbers during depositions: “Consistent with and for the reasons set forth in the Court’s order at Dkt. 478, Anthropic may not ask questions about these numbers in any deposition, whether pursuant to Federal Rule of Civil Procedure 30(b)(1) or 30(b)(6).”
What appears to be developing in these copyright lawsuits is a possible distinction between (1) infringing outputs generated by the plaintiffs’ own prompting and potentially adversarial tactics and (2) infringing outputs generated by users of the AI model in real life.
Anthropic is angling to show that Concord Music had to use extreme efforts and prompting to generate a relatively small number of allegedly infringing lyrics. That type of evidence might help to show that Claude’s guardrails are reasonably effective for the average user, and that Concord Music’s examples do not accurately reflect use of Clause in real life. For example, let’s imagine it took 1,000 prompts to Claude to even generate 1 allegedly infringing lyric.
