NMPA told the Copyright Office AI training fails fair use

The audio task

A music publisher deciding whether to license a catalogue to an AI developer, or to litigate instead, wants to know what its own trade association has formally told the government the legal baseline is. The National Music Publishers' Association filed exactly that argument with the US Copyright Office rather than leaving the question to informal industry comment.

What the documents show

NMPA's own comments filed under Docket No. 2023-6, in response to the Office's Notice of Inquiry, state that a musical work generated principally through AI, with no meaningful human creative contribution, 'should not be copyrightable' and should be treated like a public-domain work, while works where a human uses AI 'as a tool in the creative process' should remain eligible for protection. The submission argues AI training on copyrighted works fails every factor of the US fair-use test, calling training 'fundamentally a commercial endeavor,' and states NMPA opposes an opt-out system for AI training, since 'US copyright law is an opt-in system.' A report on nmpa.org, dated 13 December 2023, quotes the submission's own description of generative AI as, in NMPA's words, 'the greatest risk to the human creative class that has ever existed,' and notes the submission was dated 30 October 2023 and cites a US District Court ruling that month, Thaler v. Perlmutter, in which Judge Beryl Howell wrote that 'human authorship is a bedrock requirement of copyright.'

Rights status

NMPA's filing is advocacy submitted to a regulator, not a court decision, and the Thaler ruling it cites concerned a non-musical AI-generated image with no human author at all, a narrower fact pattern than a human songwriter using an AI tool. NMPA's own submission does not claim AI training on published songs has been judicially declared infringing; it argues the fair-use factors weigh against such use and asks the Office to recommend licensing and record-keeping requirements.

What to check before you use it

This is an editorial checklist. A publisher relying on NMPA's fair-use analysis should check subsequent court rulings in the actual AI-training infringement cases the submission references, since a trade association's regulatory comment is not binding precedent. Anyone quoting NMPA's 'greatest risk' language should attribute it as NMPA's own characterisation rather than as an independently established fact.

• Have any of the AI-training infringement suits NMPA references reached a final ruling since October 2023?
• Does the work in question involve a human author at all, distinguishing it from the Thaler fact pattern?
• Did the Copyright Office's resulting guidance adopt, modify, or decline NMPA's opt-in and record-keeping requests?

NMPA's submission is a clearly argued advocacy document with a specific regulatory audience, and its strongest lines are best read as the association's position on an unsettled legal question, not as a report of a decided one.