AI-assisted inventions: Is your document retention policy keeping up?

Artificial intelligence tools — from consumer-facing platforms to enterprise-deployed systems— are now deeply embedded in scientific, engineering, and product development workflows. The rise of generative AI and machine learning tools has transformed R&D across industries, with AI systems increasingly being used to generate novel solutions. Increasingly, these tools are generating outputs that if substantial human contribution is present would constitute patentable inventions.

The USPTO is paying attention — including, most recently, human inventorship requirements for ornamental designs, one of the latest topics at the heart of design patent protection. As recently as March 2026, the USPTO explored the role of Generative AI in design patents, noting that Generative AI is playing an increasingly significant role in designers’ creative processes — and that this may present barriers to design patent protection, since only natural humans can be inventors. That observation is not merely academic. It signals a recognition that the legal framework has not yet caught up to the way designers actually work in the age of artificial intelligence.

That fundamental shift brings with it the need to document the human contribution to prove patent eligibility and in many ways document retention may become even more crucial to prove or defend against attacks to patent validity. Most companies have existing document retention policies designed to capture and preserve evidence of innovation — lab notebooks, experimental results, invention disclosures, and patent prosecution files. However, many of these policies were not designed with AI in mind. The gap between documentation policies and the ability to capture how innovation actually happens is no longer a theoretical concern. It is an active and growing threat to the enforceability of patents that companies have invested significant resources to obtain.

This article examines: (1) how AI tools — both public and enterprise — may be silently circumventing your document retention policies;  (2) the risks of using public AI tools, including inadvertent public disclosure and unintentional IP infringement; and (3) best practices for updating policies, training inventors, and meeting your duty of candor to the USPTO.

The legal framework: What the law currently requires

Under current U.S. law, only a natural person can be named as an inventor. AI systems, including generative AI and other computational models, are instruments used by human inventors — analogous to laboratory equipment, computer software, research databases, or any other tool that assists in the inventive process. Any application listing an AI system as an inventor or joint inventor will be rejected under 35 U.S.C. §§ 101 and 115.

This principle was firmly established in Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022), where Dr. Steven Thaler alleged that his DABUS AI system generated two inventions. Patent applications naming DABUS as inventor were refused by the USPTO, and that refusal was affirmed by the Federal Circuit. The holding was unambiguous: an “inventor” must be a human, and an application for patent which fails to list any human as an inventor should be denied — making proof of human conception critical. In the Patent Act, “individuals” – and, thus, “inventors” – are unambiguously natural persons.

While the Federal Circuit has long held that “conception is the touchstone of inventorship,” Burroughs Wellcome Co. v. Barr Labs., Inc., 40 F.3d 1223, 1228 (Fed. Cir. 1994), the USPTO has since issued successive rounds of guidance on how this principle applies when AI assists in the inventive process. The current operative framework issued November 28, 2025, acknowledges that AI systems may provide services and generate ideas, but they remain tools used by the human inventor who conceived the claimed invention. Thus, when a natural person uses AI to develop an invention, the key question is whether that person made a substantial contribution to the conception of the invention under the traditional standard of conception. While the substantial contribution inquiry is binding on USPTO examiners, the guidance itself is not binding on courts. This almost certainly sets the stage for future evaluation of “edge cases,” where courts and the USPTO will need to determine whether a human who recognizes, selects, or refines an AI-generated output has contributed enough to conception to have “invented” the subject matter — representing the most significant open issue arising from the updated guidance. As AI-assisted innovation accelerates, practitioners should expect further clarification — likely through case law — on conception and patent inventorship in the AI context.

The document retention gap: How AI is undermining your IP protection without you knowing it

Traditional retention frameworks were not built for AI

Most companies with active R&D and design programs maintain document retention policies that capture lab notebooks, experimental results, invention disclosure forms, prior art searches, and patent prosecution correspondence. These records serve a critical legal function: they create a contemporaneous evidentiary record of human conception that can be used to establish and defend inventorship, support validity in post-grant proceedings, and demonstrate compliance with the duty of candor to the USPTO. These frameworks were designed for a world where inventors and designers worked with physical and conceptual materials and documented their work in files that could be saved and organized. AI-assisted design and invention works very differently — and the gap between existing policy frameworks and current innovation practice is widening every day.

The silent gap: What AI interactions are not being saved

When inventors use AI tools — even enterprise or private systems — the prompts, queries, intermediate outputs, and iterative interactions that led to an invention may not be saved or retained at all. This lack of documentation might seem advantageous (i.e., making it difficult for a challenger to prove an AI was involved). However, as discussed below, the duty of candor is still an obligation for inventors, practitioners, and anyone associated with the filing and prosecution of the patent application in making representations before the USPTO. As further discussed below, digital forensics will likely make any perceived advantage from the lack of documentation moot. This creates a critical evidentiary void: if inventorship is later challenged, the company may be unable to demonstrate what the human inventor actually contributed versus what the AI generated — potentially fatal to patent validity. For design patent purposes, this gap is especially dangerous. The entire inventive act in design patent law is the conception of the ornamental appearance. If no record exists of the designer’s iterative decisions — which AI-generated options were rejected, which were refined, what specific visual choices the designer made — then the “conception” by a natural person could be called into question.

Companies should maintain version control for AI-generated materials to track how human inventors conceived of the invention, and refined, selected, and integrated AI outputs — including, for example, the selection of inputs provided to AI tools and any configuration or training of the AI tools. This principle applies with equal — and arguably greater — force to design workflows. Patent applicants must clearly demonstrate how a natural person conceived the invention, and detailed records of human input — especially where AI-generated outputs are involved — will be critical to withstand scrutiny.

Third-party vendors: The contractual gap

Enterprise and third-party AI vendors may not, by default, log or preserve user prompts and inputs, version histories of AI-generated outputs, iterative refinements made by the human inventor, or the AI model version used at the time of generation. Existing vendor agreements and terms of service may not address document retention obligations at all, let alone align them with the company’s internal IP policies or industry standards for innovation recordkeeping. Vendor defaults may actively work against the company’s interests: some systems auto-delete session histories; others use inputs for model training, creating potential public disclosure risks; and few provide the kind of tamper-evident logging that would satisfy evidentiary standards in litigation.

Recommended contractual protections include: requiring vendors to retain all session data for the full term of any patent issuing from the relevant application (up to 20 years from filing for utility patents, and 15 years from grant for design patents), plus any applicable post-expiration challenge periods, to ensure that documentation remains available to defend against validity challenges that may arise throughout the patent’s enforceable life; requiring exportable, tamper-evident logs; prohibiting use of inputs for model training; and including audit rights.

The public AI risk: Disclosure and inadvertent infringement

Many public AI systems — consumer-facing services like ChatGPT, Gemini, and similar tools — learn from user interactions, meaning that prompts and inputs may be retained and used for model training, may resurface as outputs provided to other unrelated users, and may constitute a “public disclosure” under patent law that starts the one-year clock under 35 U.S.C. § 102 — or, in some circumstances, such as for patent filings in countries that have no grace period, immediately bar patentability. For design work, this risk is particularly concrete. If a designer uses a public AI image-generation tool to develop a product’s ornamental appearance and those prompts and outputs are retained and used for training, the visual concepts that could have been protected as design patents may have already been publicly disclosed.

There is a second, often-overlooked danger: inadvertent IP infringement. AI systems generate outputs based on training data that may include patented inventions, proprietary methods, and protected prior art. Inventors who rely on AI-generated solutions without conducting a Freedom to Operate (FTO) analysis run serious risks. The AI-suggested solution may already be patented by a third party, making commercialization an infringement risk; the company may unknowingly pursue patent protection for subject matter that is already prior art, wasting prosecution resources and exposing the application to invalidity; and there is no way to determine the provenance of AI-generated outputs without independent investigation. Best practice: treat AI-generated outputs as you would any third-party contribution — conduct a thorough FTO search and prior art analysis before investing in patent prosecution or committing resources to product development based on AI-suggested solutions.

Best practices: Protecting design and utility IP rights in an AI-assisted environment

The duty of candor

Under 37 C.F.R. § 1.56, applicants, attorneys, and agents owe a duty of candor and good faith to the USPTO, including a duty to disclose all information material to patentability. The USPTO guidance does not introduce a formal requirement to disclose the use of AI systems in the inventive process. However, it highlights that this falls within the existing duty of individuals involved in patent applications to disclose all information material to the invention’s patentability, including proper inventorship. For AI-assisted inventions, this means disclosing evidence if an inventor’s claimed contribution was actually performed by an AI system, as such information raises questions of proper inventorship.

Failure to make appropriate disclosures where material information exists could render the patent unenforceable for inequitable conduct. See Therasense, Inc. v. Becton, Dickinson & Co., 649 F.3d 1276 (Fed. Cir. 2011) (en banc). Inventors and companies should proceed on the assumption that, in any future patent challenge or litigation, opposing counsel will use digital forensic techniques to identify AI tools used during development, recover prompt histories, session logs, and AI-generated outputs, and demonstrate that the claimed human conception was actually AI-generated. If the AI contribution is substantial and undisclosed, forensic evidence could be used to invalidate the patent, support an inequitable conduct defense, or demonstrate that the named human inventor did not actually conceive the invention.

Updating your document retention and AI governance policies

Companies should take the following concrete steps to align their document retention and AI governance policies with the current legal landscape:

• Audit existing document retention policies to identify gaps relating to AI-generated content and AI-assisted workflows.
• Update invention disclosure forms to require inventors to identify: (a) whether AI tools were used; (b) which tools; and (c) the nature and extent of AI involvement.
• Require that all AI-assisted innovation and design activities be conducted using only approved private or enterprise AI systems.
• Negotiate and update vendor agreements to include appropriate data retention, logging, confidentiality, and audit provisions.
• Implement version-controlled AI session logs as part of the standard innovation documentation package.
• Train all R&D and design personnel on the distinction between public and private AI tools and the legal consequences of using the wrong tool.
• Establish a regular review cycle for AI governance policies given the pace of legal and technological change.

For design teams specifically, documentation practices should mirror — and in some respects exceed — those required of utility patent inventors. The designer’s iterative decision-making process is the inventive act. Require inventors to document their design conception, including processes, problem framing, and design choices before running or integrating AI tools, and when AI proposes multiple outputs, document why the inventor selected or modified a particular output. Leaving such decisions undocumented is a risk no company should take.

Ensuring vendor and enterprise system compliance

Companies must also audit whether their AI vendor agreements address: data retention and logging obligations; access and exportability of session logs and prompt histories; data security and confidentiality protections consistent with trade secret requirements under the Defend Trade Secrets Act; and alignment with the company’s internal document retention schedule. This is especially critical for design teams relying on third-party generative AI image tools, where vendor terms rarely contemplate the IP documentation needs of a patent applicant.

Conclusion

The USPTO’s March 2026 guidance is a marker, not a finish line. The legal framework governing AI-assisted inventions is still evolving, and the evidentiary standards that will ultimately define what counts as human conception in an AI-assisted workflow have not yet been set by the courts. What is clear is that the gap between how companies innovate today and the records they will need tomorrow is real, growing, and carries consequences that existing document retention policies were not designed to address. The question for companies that use AI in their R&D and design workflows is not whether this issue will arise, but whether their documentation practices will be adequate when it does.