5 in ’25: Artificial Intelligence

2025 marked a pivotal shift in the adoption and implementation of artificial intelligence (AI) technology by intellectual property (IP) practitioners, driven by new regulatory guidance, landmark decisions, and the release of transformative AI tools. Here, we highlight several stand-out developments.

1. Human authorship required for copyright protection

The United States Copyright Office (USCO) reaffirmed its position that fully produced AI-generated content is not inherently copyrightable. In January 2025, the USCO released Part II of its report on AI. The report confirmed that human authorship is critical to copyright law, clarifying that AI output may be protected when there is some level of human contribution (e.g., such as arranging or modifying AI artwork outputs), but prompting alone, even if elaborate and unique, is insufficient to grant authorship.

The D.C. Circuit affirmed this in Thaler v. Perlmutter. There, the court considered Dr. Stephen Thaler’s "Creativity Machine," an AI system that autonomously generates artwork. Thaler sought to copyright a piece of artwork made by Creativity Machine, identifying Creativity Machine as the sole author and himself as the owner on the copyright application. The USCO denied the application, and the district court affirmed the denial. The D.C. Circuit reviewed the case and affirmed, holding that an AI machine cannot be the sole author of a work of art. Dr. Thaler filed a petition for a writ of certiorari with the U.S. Supreme Court, which is pending as of the date of this writing.

Following the USCO report and the Thaler decision, the question remains as to what will be "sufficient human creativity" to warrant protection for an AI output.

2. USPTO sets precedent regarding Section 101’s application to AI in Ex Parte Desjardins

The United States Patent and Trademark Office (USPTO) Appeals Review Panel (ARP) issued a decision in Ex parte Desjardins regarding the Section 101 eligibility of certain AI technology. Many have interpreted the result to signal that the USPTO may be taking a more lenient approach to patent eligibility for AI technologies.

Specifically, the applicant sought to patent a method for training machine-learning models in continual learning, a method for learning new tasks without "forgetting" earlier ones. The method aimed to reduce storage demands and improve efficiency. The Patent Trial and Appeal Board (PTAB) rejected the method claim under Section 101, finding that the claims were directed to an abstract idea without any practical application. The ARP — convened by newly confirmed USPTO Director John Squires — disagreed, vacating the Section 101 rejection. Because the claims recited a technical improvement to the operation of a machine-learning model, the ARP found that the claims constitute a practical application of an abstract idea and, therefore, are patent-eligible. On November 4, 2025, the PTAB designated the ARP's decision precedential.

Many view Desjardins as signaling a shift toward broader patent eligibility for machine-learning models, as the ARP directed both patent examiners and PTAB panels to include greater specificity in their Section 101 analyses and assess whether the claims improve the technical functioning of the AI model.¹ Applicants prosecuting applications related to AI and machine-learning should ensure they are emphasizing the improvements made to the technology in the specification.

3. Training generative AI with copyrighted works

In May 2025, the USCO released a pre-publication version of Part III of its report on AI, giving preliminary guidance on the use of copyrighted materials to train generative AI models. The USCO provides a nuanced breakdown of how courts should evaluate the fair use factors for training models.

• Factor 1 — Purpose. Training models on massive datasets can be transformative, “but transformativeness is a matter of degree” that depends on “the functionality of the model and how it is deployed.” AI training is “most transformative when the purpose is to deploy it for research, or in a closed system that constrains it to a non-substitutive task.”
• Factor 2 — Nature. The use of “more creative or expressive works” and previously unpublished works weighs more heavily against fair use than the use of “factual or functional works.”
• Factor 3 — Amount. While copying the entirety of a work may be “functionally necessary” for training, it generally weighs against fair use unless the model includes “guardrails” to prevent the reproduction of the protected expression.
• Factor 4 — Market effect. The ability of AI to flood the market can lead to significant harms for copyright owners, including lost sales, market dilution, and lost licensing opportunities. “Where licensing options exist or are likely to be feasible,” bypassing them weighs against fair use.

Ultimately, the USCO advocates for the growth of existing data licensing frameworks to train AI models. The USCO acknowledges the logistical and organizational challenges that come with licensing for training and recommends the use of collective management organizations and extended collective licensing as a solution.

The report is already impacting existing litigation. Copyright owners have started using the report as supplemental authority in response to fair use defense arguments from AI companies. Upcoming decisions in 2026 will give light to judicial perspectives on this issue.

4. USPTO Automated Search Pilot Program

In October 2025, the USPTO launched its AI pilot program, Artificial Intelligence Search Automated Pilot (ASAP!), which allows applicants to request an AI-powered prior art search before substantive examination. ASAP! eligibility applies to original, noncontinuing, nonprovisional utility applications filed under 35 U.S.C. § 111(a) between October 20, 2025, and April 20, 2026, or until 1,600 applications are accepted across all technology centers.

The AI tool was trained using publicly available patent data, including patents, published applications, patent classifications, document citations, and human-rated similarity. The training data excluded applicant, inventor, and assignee information to avoid potential bias. ASAP! provides applicants with an Assisted Search Results Notice (ASRN). The ASRN contains a top 10 list of potential prior art issues in descending order of relevance and provides the applicant with potential courses of action, such as filing a preliminary amendment or requesting deferral of the application.

The introduction of ASAP! emphasizes the USPTO’s greater initiative to ensure that patents are “born strong,” in Director Squires’ words. From an examiner’s perspective, the AI tool could discover relevant prior art that may otherwise be missed, in turn reducing a patent’s vulnerability in post-grant proceedings or litigation. From an applicant’s perspective, the AI tool gives early insight into potential hurdles in the application process and can inform prosecution strategy.

5. Agentic AI and IP

Last year saw the emergence of agentic AI. Unlike generative AI, which produces outputs in direct response to user inputs, agentic AI uses at-least-semi-autonomous decision-making with minimal human intervention. Many agentic AIs can independently set goals, make decisions, and execute multi-step tasks without constant prompting, making it a helpful tool for complex project management.

AI companies are incorporating agentic AI into new IP monitoring and enforcement tools. For example, agentic AI is being used for real-time infringement monitoring using image recognition and text matching to detect potential trademark and copyright infringement and then draft takedown notices. Agentic AI can also be used for patent portfolio landscaping by analyzing a portfolio’s technology, classification, and potential competitors and then estimating the relative value.

The rise of agentic AI also brings potential risks. For infringement monitoring, attorneys should ensure adequate review of generated materials or proposed actions. Allowing AI agents to make final decisions could lead to liability for errors. In the prosecution context, innovators should be cognizant of how agentic AI is being used to generate potentially patentable solutions. Prosecution carries with it a duty to disclose prior art, but the USPTO has yet to opine on the extent to which information considered by AI agents might be encompassed by those disclosure requirements. As always, companies should focus on human intervention and innovation when identifying potentially patentable solutions.