ai training data subpoenas

Summary#

This bill adds a new section to the Copyright Act that creates a special subpoena process to let copyright owners ask AI developers for copies of, or records identifying, works used to train generative AI models. The requestor files a proposed subpoena and a sworn statement of a subjective good faith belief that the developer used the requestor’s copyrighted work. If the paperwork is in proper form, the court clerk must sign and return the subpoena. The developer must then “expeditiously” disclose the requested material. The bill also creates a confidentiality duty for the recipient, a rebuttable presumption if a developer fails to comply, and possible sanctions for bad-faith requests.

Key changes:

• Creates a new Section 514 in Title 17 giving copyright owners a way to get training materials or identifying records from AI developers by subpoena signed by a court clerk.
• Defines covered terms, including “developer,” “generative artificial intelligence model,” “training material,” and “substantially modify.”
• Limits requests to works owned or controlled by the person asking (no fishing for other people’s works).
• Requires a sworn declaration stating a subjective good faith belief and limits use of disclosed materials to protecting the requester’s rights.
• If a developer does not comply, that failure creates a rebuttable presumption that the developer made copies of the work.
• Allows courts to impose sanctions for bad-faith subpoena requests, applying existing Rule 11 standards.

What it means for you#

• Copyright owners and agents

  • You can ask a court clerk to issue a subpoena to an AI developer to get copies or records identifying works you think were used to train a generative model.
  • You must submit a proposed subpoena and a sworn declaration saying you personally believe your works were used and that you’ll use the material only to protect your rights.
  • You cannot use this process to seek other people’s copyrighted works.

• AI developers (companies, researchers, and some government units)

  • If you design, substantially modify, own, or supervise training data for a generative AI model, you are a “developer” under the bill and can receive these subpoenas.
  • You must respond “expeditiously” and provide the requested copies or identifying records unless you successfully resist or negotiate limits.
  • Failure to comply creates a legal presumption that you made copies of the copyrighted work (a presumption that can be challenged).

• Noncommercial end users

  • People who only use generative AI models for noncommercial purposes are explicitly excluded from the definition of “developer.”

• Courts and court clerks

  • Clerks must issue subpoenas that are in proper form and supported by a sworn declaration. Courts handle motions for sanctions and enforcement matters; many procedural rules apply “to the greatest extent practicable.”

• Researchers and third parties

  • The bill may make some training dataset information available to rightsholders, but the requester must keep what they receive confidential and use it only for rights protection.

Expenses#

No publicly available information.

Possible fiscal and private costs (inferred from the bill text):

• Developers may incur costs to search for, assemble, and disclose training data or identifying records. This could affect small companies more.
• Courts and clerks may see increased administrative workload from processing and issuing subpoenas.
• There may be legal costs for both requesters and developers to litigate disputes over compliance, scope, or confidentiality.
• Enforcement costs may arise if courts must resolve noncompliance or bad-faith claims.

Proponents' View#

• The bill appears intended to give copyright owners a fast, formal way to learn whether their works were used to train generative AI models.
• Supporters may argue this increases transparency about training datasets and helps rightsholders enforce their copyrights.
• The rebuttable presumption for noncompliance could make it easier for owners to pursue later legal claims if developers refuse to cooperate.
• Requiring a sworn declaration and allowing sanctions for bad faith seeks to limit frivolous or abusive requests.
• Confidentiality duties aim to protect disclosed materials from broad public release.

Opponents' View#

• One concern is that the bill could force disclosure of proprietary training datasets, including trade secrets, with only a confidentiality rule that limits further sharing by the requester; the bill does not specify how commercial confidentiality or trade secrets will be protected in practice.
• The standard for issuing a subpoena is a requester’s “subjective good faith belief” and a clerk’s ministerial signature if the form is proper. This may permit broad or speculative requests with limited initial judicial review.
• Terms such as “expeditiously” and “records sufficient to identify with certainty” are not defined, leaving uncertainty about what developers must produce and how fast.
• Developers, especially small ones, could face significant compliance costs and operational burdens to search and produce training records.
• The bill applies to state and local government agencies that meet the developer definition, which may raise public-sector workload and information concerns.
• Remedies and enforcement are partly left to existing court rules “to the greatest extent practicable,” so practical reliance on current civil-procedure tools may produce uneven outcomes.