TRAIN Act
- Sponsor
- Sen. Welch, Peter [D-VT]
- Committees
- Judiciary Committee (primary)
- Last action
- Jul 24, 2025
Bottom line
The TRAIN Act creates a new administrative subpoena power for copyright owners to discover if their works were used to train generative AI models, aiming to increase transparency and accountability in the AI industry.
What it actually does
This bill establishes a new administrative subpoena process, allowing copyright owners to request information from artificial intelligence (AI) model developers about whether their copyrighted works were used in training generative AI models. It mandates that U.S. district court clerks issue these subpoenas upon proper request and creates a rebuttable presumption of copyright infringement if a developer fails to comply. The process is intended to provide copyright holders with a mechanism to investigate potential unauthorized use of their content by AI systems.
Proponents argue
Supporters argue that the bill is a crucial step towards protecting intellectual property rights in the age of artificial intelligence. They contend that copyright owners currently lack adequate tools to determine if their creative works are being used without permission to train AI models, which can then generate new content that competes with their original creations. This bill provides a necessary, streamlined mechanism for transparency, enabling creators to enforce their rights and ensure fair compensation.
Opponents contend
Opponents express concerns that the bill could impose significant burdens on AI developers, potentially stifling innovation and creating a flood of administrative subpoenas. They argue that the broad definition of 'developer' and the administrative nature of subpoena issuance, which lacks initial judicial review, could lead to misuse or harassment. Furthermore, they contend that requiring disclosure of training data could compromise proprietary information and trade secrets essential to AI model development.
The bill is concise and can be read and understood by an informed reader within a reasonable timeframe, allowing for thoughtful evaluation.
Section 2(a), adding 17 U.S.C. 514(a)
Definitions for AI Subpoena Process
This provision defines key terms central to the bill's operation, including 'artificial intelligence model,' 'developer,' 'generative artificial intelligence model,' and 'training material.' Notably, 'developer' includes those who design, produce, or substantially modify models and curate/use training datasets, while excluding noncommercial end users. 'Substantially modify' covers updates that materially change functionality, including retraining or fine-tuning.
Supporters argue
Proponents argue that clear definitions are essential to ensure the subpoena process targets the appropriate entities responsible for training AI models, preventing evasion while protecting noncommercial users. The definition of 'substantially modify' is crucial for capturing ongoing development and updates.
Critics contend
Opponents may argue that the definition of 'developer' is too broad, potentially encompassing entities with limited control over the final AI model or its training data, thereby creating undue burden. Concerns might also be raised about the scope of 'substantially modify' and its implications for continuous AI development.
Tradeoffs
This provision balances the need for clear legal scope to ensure accountability with concerns about overreach and potential burden on a wide range of AI-related activities.
Section 2(a), adding 17 U.S.C. 514(b)
Administrative Subpoena Request Authority
This section grants legal or beneficial copyright owners the right to request a U.S. district court clerk to issue a subpoena to an AI developer. The subpoena would compel disclosure of records identifying the owner's copyrighted works used to train a generative AI model. The request must be based on a 'subjective good faith belief' that the developer used the copyrighted works, and explicitly limits the subpoena to the requester's own works.
Section 2(a), adding 17 U.S.C. 514(e)
Mandatory Subpoena Issuance by Clerk
This subsection dictates that if a proposed subpoena and its accompanying sworn declaration are in proper form and properly executed, the U.S. district court clerk *shall expeditiously issue and sign* the subpoena. This makes the issuance a ministerial act, meaning the clerk does not exercise discretion or evaluate the merits of the 'good faith belief' at the time of issuance.
Section 2(a), adding 17 U.S.C. 514(i)
Rebuttable Presumption for Non-Compliance
This subsection establishes a significant enforcement mechanism: if an AI developer fails to comply with a subpoena issued under this Act, that failure creates a 'rebuttable presumption' that the developer made copies of the copyrighted work in question. This shifts the burden of proof in potential copyright infringement cases, requiring the developer to demonstrate they did not use the work.
Section 2(a), adding 17 U.S.C. 514(j)
Sanctions for Bad Faith Subpoena Requests
This subsection allows a court to impose sanctions on a copyright owner or authorized person if they request a subpoena under this Act in bad faith. It specifies that Rule 11(c) of the Federal Rules of Civil Procedure, which governs sanctions for improper filings, shall apply. This provision is intended to deter frivolous or malicious use of the subpoena process.
The definition of 'developer' explicitly includes persons or agencies that 'engage in or supervise, including as a third party training dataset curator' the curation or use of the training dataset. This expands the scope beyond just the primary AI model creator.
Section 2(a), adding 17 U.S.C. 514(a)(3)(A)(ii)
Why it matters:This is standard legislative practice for defining scope, but its implications for the supply chain of AI training data are significant.
Case for: Supporters argue that including third-party curators is essential to ensure comprehensive accountability, as these entities often play a critical role in assembling the datasets used to train AI models. Excluding them would create a loophole for developers to outsource responsibility.
Case against: Opponents might argue that this broadens the definition of 'developer' too much, potentially subjecting entities with limited direct control over the final AI model to significant legal burdens. This could complicate the AI supply chain and increase costs for data providers.
Estimated impact: Significantly expands the universe of entities potentially subject to subpoenas, impacting data brokers and specialized AI training data companies.
A legal or beneficial owner or authorized person who receives copies or records from a developer under this section 'may not disclose the copies or records to any other person without proper authorization or consent.'
Section 2(a), adding 17 U.S.C. 514(g)
Why it matters:This is a standard confidentiality clause, but its placement might make its importance less obvious.
Case for: Proponents argue this confidentiality clause is vital to protect developers' proprietary information and trade secrets that might be revealed during the subpoena process. It ensures that copyright owners use the information solely for their stated purpose of protecting their rights, not for competitive intelligence or public disclosure.
Case against: Critics might argue that enforcing this confidentiality could be challenging, especially if the information is needed for broader legal proceedings or if the 'proper authorization or consent' is difficult to obtain. It could limit the ability of copyright owners to share information with legal counsel or other affected parties.
Estimated impact: Provides a layer of protection for AI developers' sensitive information, but may create practical challenges for copyright owners in utilizing the disclosed information.