Taking a Closer Look: What the Copyright Office’s 2025 AI Reports Mean for 2026

By: Matt Sumner and MacKinzie Neal

In January and May 2025, the U.S. Copyright Office (the “Office”) released the final two installments of its comprehensive study on copyright and artificial intelligence. These reports, addressing the copyrightability of AI-generated works and the fair use implications of AI training, provide the most comprehensive regulatory guidance to date on these contested issues. With many AI training lawsuits pending in federal courts (as discussed in further detail here), these reports will likely serve as a significant reference point for courts and practitioners throughout 2026.

The Human Authorship Boundary

In its January 2025 Report, the Office reaffirmed the “human authorship” requirement for copyrightability, a position consistent with its earlier refusal to register works solely generated by AI models.  The Report reiterates the Office’s fundamental positions on copyrightability: human authorship remains a bedrock requirement, and that works entirely generated by AI are not copyrightable. Notably, the Report provides guidance regarding the role of “prompts” that will be valuable for corporate legal departments managing AI-assisted creative workflows.

The Report clarifies that most prompts do not constitute “authorship” because users do not exercise sufficient control over the specific expressive elements of the output. As the Office explained, “prompts may reflect a user’s mental conception or idea, but they do not control the way that idea is expressed.” The Office analogized prompt use to the joint authorship context: just as a person who solely describes to an author what a commissioned work should look like is not a joint author, someone entering a prompt into an AI model cannot be considered an author of the AI-generated output.

The Office further rejected the contention that iterative prompt refinement to achieve a desired output constitutes a sufficient basis for copyright protection. The Office characterized such arguments as impermissible “sweat of the brow” claims, a theory the Supreme Court rejected in Feist Publications, Inc. v. Rural Telephone Service Co., which fails to address the threshold requirement of originality.

For businesses, this creates a potential “public domain gap.” Works in the public domain are, by definition, not under copyright protection. So, if a marketing department relies on raw AI outputs for a public campaign, that content may lack copyright protection (i.e., it might sit in the public domain) and could be freely copied by competitors. However, businesses leveraging AI tools to make content can strengthen copyright protection by demonstrating meaningful human contributions that involve: (1) significant creative modification or selection and arrangement of the AI-generated material in a sufficiently creative way such that the resulting work as a whole constitutes an original work of authorship, (2) AI use limited to assistive roles that facilitate the human creative process (e.g., removing unwanted elements from an image or using AI for ideation), or (3) ‘expressive inputs’ where a copyrightable work is used as an input and remains perceptible in the output, in which case the human can claim authorship of the perceptible portion.

The Office emphasized that such determinations will be made on a case-by-case basis. The Report also noted that while some AI systems allow users to set “seed” values for more consistent outputs, this does not satisfy authorship requirements given the lack of guaranteed consistency. The Office acknowledged that future technological developments may warrant revisiting this analysis.

Finally, the Office declined to recommend sui generis protection for AI-generated works, concluding that existing laws are adequate and that additional protection could undermine human creativity and flood the market with low-quality content.

Training and the “Continuum of Transformativeness”

In the absence of definitive judicial precedent, the May 2025 Report, Generative AI Training, addresses a critical question confronting the industry: whether “ingesting” millions of copyrighted works to train a model constitutes fair use. Fair use is an affirmative defense to claims of copyright infringement, requiring analysis of four statutory factors: (1) the purpose and character of the use; (2) the nature of the copyrighted work; (3) the amount and substantiality of the portion used; and (4) the effect on the potential market for the original work. “Transformativeness” has traditionally been regarded as a significant, and often determinative, way to analyze the fair use factors.

Rather than adopting a binary view of whether a specific use meets the transformativeness threshold, the Office articulated a continuum:

• High Transformativeness: On one end of the spectrum, the Office identified uses where the AI model is deployed for research or a “closed system that constrains it to a non-substitutive task,” such as training on a wide set of data for content moderation purposes.

• Low Transformativeness: On the other end are uses where AI models are trained specifically to produce content (e.g., text, music or art) that serves the same or similar aesthetic purpose as the training data. The Office cited the example of a model training on copyrighted images from an animated series to reproduce characters from that series.

  
  

The Pivot to Market Harm

Based on the Office’s guidance, the fourth fair use factor, the effect upon the potential market, is likely to constitute the determinative consideration in AI copyright disputes in 2026. The Office identified three critical areas where AI training risks a finding of market harm:

1. Lost Licensing Revenue: The Report highlights that a voluntary licensing market is already maturing, as evidenced by recent deals between publishers and AI labs. The existence of a viable licensing market makes it harder to argue that unauthorized training is necessary or fair.

   
   

2. Market Dilution: Even if a specific AI output does not infringe a specific work, the aggregate impact of AI-generated content can dilute the value of the broader market for human-authored works.

   
   

3. Lost Sales: Using pirated datasets to create a training library could enable consumers to generate verbatim or near-verbatim copies of their desired work, effectively eliminating a potential sale by the creator to that consumer.

   
   

Practical Guidance for 2026

The Office’s reports, while non-binding, constitute a cornerstone by which legal practitioners, federal agencies and potentially legislators approach these issues. While not all companies face the same considerations as developers of foundational models, there are broadly applicable takeaways. As we move into 2026, companies should consider the following:

• Establish a “Creative Audit Trail”: For critical IP, companies should document specific human interventions, such as editing, structural changes and creative work, utilized to ensure that AI outputs are copyrightable assets. This documentation may prove essential in future registration applications or litigation.

  
  

• Negotiate Training Data Warranties: In licensing agreements with AI providers, customers should consider representations that training data has been lawfully acquired and does not infringe third-party intellectual property rights. AI providers, in turn, should maintain detailed records of training data sources and implement remediation processes in the event of infringement claims.

  
  

• Monitor Evolving Guidance: The May 2025 report was released amid significant leadership changes at the Copyright Office. Companies should monitor whether the new administration issues revised guidance that could alter the regulatory landscape.