Inventorship records, assignment language, claim drafting, and AI vendor terms can affect the value of an AI company's patents to a buyer or investor.

In 1991, Roche Molecular Systems acquired the assets of Cetus Corporation related to polymerase chain reaction (PCR), the technique for amplifying DNA. Two years earlier, Mark Holodniy, a Stanford research fellow, had begun visiting Cetus to learn PCR and had signed a visitor's confidentiality agreement (VCA) with Cetus. The Supreme Court later described Roche's purchase as including "all rights Cetus had obtained through agreements like the VCA signed by Holodniy." Holodniy's work at Stanford and Cetus produced a laboratory test for measuring HIV levels in blood, and Stanford later obtained patents on the test and sued Roche for infringement.

The case turned on the wording of the invention agreements Holodniy had signed with each institution. His Stanford agreement stated that he "agree[d] to assign" his inventions to the university. The Cetus agreement, by contrast, stated that he "will assign and do[es] hereby assign" them to Cetus. In Board of Trustees of the Leland Stanford Junior University v. Roche Molecular Systems, Inc., 583 F.3d 832, 842 (Fed. Cir. 2009), the U.S. Court of Appeals for the Federal Circuit held that the Cetus language "effected a present assignment of Holodniy's future inventions to Cetus," while Stanford's language was a promise to assign in the future. Roche held Holodniy's interest through the Cetus acquisition, so Stanford "cannot establish ownership of Holodniy's interest and lacks standing to assert its claims of infringement against Roche." In other words, since Stanford did not own the patents, the university had no legal right to bring the suit.

The Supreme Court affirmed in 2011, 563 U.S. 776, holding that the Bayh-Dole Act, which governs inventions made with federal research funding, leaves ownership with the inventor rather than transferring ownership automatically to the university. The Court's opinion restated the premise underlying the dispute: "Since 1790, the patent law has operated on the premise that rights in an invention belong to the inventor." The Supreme Court expressly left the assignment-language question undecided in a footnote, so the Federal Circuit's distinction between a present transfer and a promise of a future transfer still governs patent assignments.

The Stanford case predates generative artificial intelligence (AI) by two decades, but AI-assisted development exacerbates the conditions behind the ownership conflict: more contributors, working across more organizations, under more inconsistent invention agreements. AI products are often built by founders, employees, contractors, university collaborators, and vendor engineers, using AI tools operated under terms AI vendors wrote. Each contributor is a potential inventor, a potential owner, and a potential recipient of confidential information. The August 11 Pioneering IP article concluded with the observation that reliance on trade secrets and speed of execution can be a weaker strategy for a company whose value will be assessed in due diligence conducted by a buyer or investor before a deal closes. This article examines how acquirers and investors evaluate an AI company's patent portfolio in due diligence.

Acquirers and Investors Put a Premium on Enforceable Patent Rights

Patent counts correlate with exits, meaning acquisitions or public offerings, in at least one angel investor group's data, as an October 6 post in this series discussed. But while a patent count measures filing activity, an acquirer or investor ultimately values the rights granted by the collection of patents. Diligence tests four properties of each patent asset: whether the company owns the asset, whether the patent would survive a validity challenge, whether the claims cover competitors' products, and whether infringement could be detected and proven.

Patent due diligence typically covers the documented chain of title, pending applications, issued claims, prosecution histories in the patent office, employment and contractor agreements, license obligations, prior art, and any pending or threatened lawsuits or patent office proceedings involving the patents. For AI companies, the product has often changed substantially since the first patent filings. A portfolio can include claims directed to an architecture the company has since abandoned, claims competitors can avoid with a design-around, and improvements described in a patent specification but never claimed. A buyer or investor will ask whether the issued claims cover the current product, and whether a pending application still allows new claims that do. Finding mismatches between the claims and the current product (or the market) before a deal begins leaves time to pursue new claims while an application in the family is still pending.

Inventorship Defects Become Ownership Headaches

Under its February 2024 guidance, the U.S. Patent and Trademark Office (USPTO) required a natural person to make a "significant contribution" to every claim of an AI-assisted invention, applying the joint-inventorship factors of Pannu v. Iolab Corp., 155 F.3d 1344 (Fed. Cir. 1998). The requirement applied even when only one person worked with the AI. A claim the AI largely generated, with the human only posing the problem or recognizing a useful output, could be rejected for lacking a human inventor. The USPTO's Revised Inventorship Guidance for AI-Assisted Inventions, 90 Fed. Reg. 54636 (Nov. 28, 2025), withdrew the February 2024 test and now treats AI as a tool, like laboratory equipment or software. A person is the inventor if the person conceived the claimed invention, even with the assistance of AI, and the Pannu factors now apply only to deciding which of several human beings are joint inventors.

Under the revised guidance, inventorship turns on conception, the formation in the mind of a complete and workable invention, and an inventor must be a human being. The Federal Circuit rejected an attempt to name an AI system as an inventor in Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022). The guidance also states that "[g]enerally, the USPTO presumes those inventors named on the application data sheet or oath/declaration are the actual inventor or joint inventors," meaning the inventors identified in the application's filing papers. Disclosure and documentation of AI use are left to the applicant, so records of human conception matter for an evidentiary reason: an accused infringer in litigation can challenge whether the named inventors are correct by comparing them against the existing record. An acquirer or investor should run the same test during due diligence.

Errors in naming inventors can create ownership defects that reduce the price a buyer or investor will pay. An omitted coinventor without a signed assignment holds an undivided ownership share in the whole patent. Under the Patent Act's co-ownership provision, 35 U.S.C. § 262, each joint owner can make, use, or sell the invention "without the consent of and without accounting to the other owners," and each co-owner can also license others without the other owners' consent. In Ethicon, Inc. v. United States Surgical Corp., 135 F.3d 1456 (Fed. Cir. 1998), the accused infringer located an unnamed coinventor during the case, took a license from him, and moved to add him to the patent. The infringement suit was dismissed, because an infringement suit proceeds only if every co-owner joins as a plaintiff, and the new co-owner had already licensed the defendant.

Correcting inventorship does not always preserve a patent's validity. Section 256 of the Patent Act, 35 U.S.C. § 256, allows a court or the USPTO to correct the list of named inventors. But a court can order correction only "on notice and hearing of all parties concerned," and an omitted inventor who cannot be found cannot be given that notice. In Fortress Iron, LP v. Digger Specialties, Inc., 171 F.4th 1310 (Fed. Cir. 2026), the patent owner could not satisfy that requirement. Fortress conceded that two engineers at an overseas supplier were coinventors. Fortress located one but failed to locate the other in time to give the required notice. The Federal Circuit held both patents invalid. As was reported in a September 22 post in this series, the missing engineer was found after the opinion issued, and Fortress's petition for certiorari to the Supreme Court, No. 26-362, remains pending with a response due October 19, 2026.

AI-assisted development widens the pool of contributors who can become omitted inventors: outside developers, vendor engineers, and contractors who helped design a system, signed no assignment, and may later be difficult to locate. Diligence on an AI portfolio therefore tests three facts for each important family of patents: who conceived each claim, as shown by a dated record of the human contribution and of how AI tools were used; whether every contributor signed a present assignment; and whether every contributor can still be reached if needed.

Conveyance Language Controls Changes of Ownership

Under the Patent Act's assignment provision, 35 U.S.C. § 261, patents and applications "shall be assignable in law by an instrument in writing." Section 261 also gives legal effect to recording with the USPTO: an unrecorded assignment "shall be void as against any subsequent purchaser or mortgagee for a valuable consideration, without notice, unless it is recorded in the Patent and Trademark Office within three months from its date or prior to the date of such subsequent purchase or mortgage." In other words, a later buyer who pays for a patent without knowing of an earlier, unrecorded transfer, and who records promptly, can take priority over the earlier assignee.

Recording an assignment gives public notice of the transfer but does not establish that the transfer is valid. Section 301 of the Manual of Patent Examining Procedure (MPEP), the USPTO's handbook for examiners, describes recording as "merely a ministerial act," meaning the Office files the document without reviewing the document's contents, and the USPTO's regulation at 37 C.F.R. § 3.54 provides that recording "is not a determination by the Office of the validity of the document or the effect that document has on the title." The USPTO's assignment records are the starting point for diligence, but ownership is confirmed only by reviewing the underlying agreements themselves.

The Stanford case illustrates the importance of the following test for every agreement: whether the language assigns now or promises to assign later. "Agree to assign" leaves ownership with the inventor until a second document is signed. "Hereby assign" transfers future inventions as they arise. Founder agreements signed before incorporation, consulting agreements on a vendor's standard form, university sponsored-research agreements, and offshore development contracts are the types of documents most likely to use the weaker "agree to assign" form, if they say anything about inventions at all. Confirmatory assignments (new documents in which each inventor confirms the transfer) obtained before a sale process begins can cost a fraction of what the same signatures cost once the inventor becomes aware that an important transaction depends on the signature.

Patent Eligibility Risk Affects Many AI Claims, and the Patent Office and the Courts Apply Different Standards

An issued patent can still be held invalid at any stage of litigation for lack of patent eligibility, which requires that a claimed invention be more than an abstract idea. In Recentive Analytics, Inc. v. Fox Corp., 134 F.4th 1205, 1216 (Fed. Cir. 2025), the Federal Circuit held "that patents that do no more than claim the application of generic machine learning to new data environments, without disclosing improvements to the machine learning models to be applied, are patent ineligible." Recentive's claims applied conventional machine learning to scheduling live events and building network maps for television broadcasters. The decision leaves open at least one way a machine learning claim can be eligible: by reciting a specific improvement to how a computer or another technological process operates. The Supreme Court declined to review the decision on December 8, 2025.

GoTV Streaming, LLC v. Netflix, Inc., No. 2024-1669 (Fed. Cir. Feb. 9, 2026) shows that a patent owner can win a jury verdict and still risk the patent being found ineligible on appeal. GoTV's patents claimed server-side tailoring of content to a wireless device's capabilities, a mobile-content invention analogous to some AI cases. A Central District of California jury found infringement of one patent and awarded GoTV $2.5 million. On Netflix's own appeal, the Federal Circuit held all asserted claims ineligible and ordered judgment for Netflix. Judge Taranto's opinion states the rule the court applies when evaluating the eligibility of any claim, including AI-enabled claims: although the specification informs what a claim is directed to, "only features that are claimed, not unclaimed details that appear in the specification, can supply something beyond ineligible matter." The opinion relies in part on US Patent No. 7,679,637 LLC v. Google LLC, 164 F.4th 1373 (Fed. Cir. 2026). The court's opinion on the '637 patent at issue rejected "result-focused functional language" lacking "specificity about how" the result is achieved, thus rendering ineligible claims that recite what a system accomplishes without providing how. The Supreme Court declined to review the '637 case on October 5, 2026, leaving the claims-focused rule in place.

The USPTO, by contrast, has adopted a more permissive eligibility standard for examining applications. An August 4, 2025 USPTO memorandum instructs patent examiners to reject applications on eligibility grounds only where ineligibility is "more likely than not," and Ex parte Desjardins, an appeal decision by a USPTO review panel that is binding on USPTO examiners and the appeal board, treated an improvement in how a machine learning model is trained as an improvement in the model's operation. Patents issued since August 2025 have been examined under the Office's approach and will be litigated under Recentive and GoTV. As the September 10 article in this series explained, eligibility challenges by accused infringers are ordinarily decided in court rather than at the USPTO. A buyer of an issued patent should value each claim under the court's standard applied when the patent is asserted against an accused infringer, and a large portfolio of purely functional AI claims can be worth less than a small set of claims reciting the technical features allowing the model, pipeline, or system to achieve the beneficial result.

Trade Secret Status Depends on Who Had Access and on What Terms

Much of an AI company's value derives from assets typically protected as trade secrets rather than by patents: training methods, curated datasets, model configurations, evaluation suites, and deployment know-how. Under the Defend Trade Secrets Act, 18 U.S.C. § 1839(3), information qualifies as a trade secret only if the owner "has taken reasonable measures to keep such information secret" and the information derives independent economic value from not being generally known.

Whether secrecy measures are reasonable depends heavily on who received the information and on what confidentiality terms. In Turret Labs USA, Inc. v. CargoSprint, LLC, No. 21-952 (2d Cir. Mar. 9, 2022), the U.S. Court of Appeals for the Second Circuit explained that where the secret lies in what users of a system can see, "the reasonableness analysis will often focus on who is given access, and on the importance of confidentiality and nondisclosure agreements to maintaining secrecy." Every disclosure to an AI tool adds a recipient: the tool's vendor. The August 24 article in this series compared vendor terms in detail and found a clear difference between enterprise and consumer offerings. Major enterprise agreements treat customer inputs and outputs as the customer's confidential information. By contrast, the two consumer terms of service reviewed there do not use the word "confidential" at all. Whether and to what extent contractual confidentiality in a vendor relationship preserves trade secret status remains an open question in the case law.

Trade secret diligence should examine the company's historical records of what information was disclosed to AI tools. An acquirer will ask which AI tools received source code, datasets, or unfiled invention disclosures, whether under enterprise or consumer terms, whether logs show what was shared, and whether controls prevented unfiled inventions from being entered into AI tools without confidentiality protections. A company that cannot document these disclosures gives the acquirer grounds to discount the value of its trade secrets, and secrecy lost through an uncontrolled disclosure generally cannot be restored.

Continuation Families Offer Strategic Value but Carry Legal and Transactional Constraints

A continuation, a later application claiming the benefit of an earlier one under 35 U.S.C. § 120, lets the owner pursue new claims based on the original application's description while using the original filing date as the basis for which public disclosures count as prior art. For a buyer, a pending continuation in a core family can be among the most valuable items in the data room, since the buyer can draft claims aimed at products (including competitors' knock-offs) currently on the market. The benefit extends only to those claims described and enabled by the original application, that is, the original application explains the invention in enough detail for a skilled engineer to make and use the invention under 35 U.S.C. § 112(a). MPEP § 201.07 specifies a continuation's description "must not include any subject matter which would constitute new matter if submitted as an amendment to the parent application," which mirrors the bar on new technical content that 35 U.S.C. § 132(a) places on amendments. In contrast, claims that rely on new technical content, whether in a new standalone application or in a continuation-in-part, receive only the later filing date, so public disclosures made before that date count as prior art against them.

Two constraints on continuation families are frequently the subject of due diligence review in transactions: terminal disclaimers and patent term. A terminal disclaimer is a stipulation a patent owner makes with the patent office, disclaiming any term of a second patent that would extend beyond the term of the first, to overcome an obviousness-type double patenting rejection. The USPTO issues that rejection when two commonly owned patents claim obviously similar inventions. Under 37 C.F.R. § 1.321(c)(3), every terminal disclaimer must state that the patent "shall be enforceable only for and during such period that said patent is commonly owned with the application or patent which formed the basis for the judicially created double patenting." An asset sale that divides a patent family, or a carve-out or reorganization that moves some patents to a buyer and leaves the rest with the seller, can leave terminally disclaimed patents unenforceable.

Patent term adjustment (PTA), the extra term the USPTO grants to compensate for its own examination delays, can also cause related patents in a family to expire on different dates. In In re Cellect, LLC, 81 F.4th 1216 (Fed. Cir. 2023), the Federal Circuit held that obviousness-type double patenting is measured from each patent's expiration date after PTA is added, so the added term can make a patent invalid over a related, earlier-expiring patent claiming an obvious variant. In Allergan USA, Inc. v. MSN Laboratories Private Ltd., 111 F.4th 1358 (Fed. Cir. 2024), the court limited that rule, holding that a first-filed, first-issued patent cannot be invalidated by a later-filed, later-issued family member that expires earlier. As a July 14 post in this series discussed, the USPTO's Patent Trial and Appeal Board has since applied Allergan to additional continuations. A buyer should therefore confirm each patent's actual expiration date and whether any added term is exposed to a double patenting challenge. Continuations filed more than six years after the family's earliest nonprovisional filing date also now carry a USPTO filing surcharge, as covered in the August 11 article.

Acquirers Also Review Enforceability, Third-Party Rights, and Copyright in AI-Generated Code

A patent portfolio can protect differentiating features, support licensing, deter copying, and preserve the ability to claim later developments. An acquirer therefore asks what each patent is used for, such as protecting a product feature or supporting licensing, and whether the patent can actually serve that purpose. A patent on an internal training technique can be difficult to enforce if infringement cannot be detected from outside the infringer's systems. A broad patent can appear valuable yet face substantial eligibility and prior-art challenges. A narrow claim on a commercially important and observable implementation can provide more durable protection.

Ownership of a patent and freedom to operate, that is, the ability to sell a product without infringing someone else's patents, are separate questions. MPEP § 301 states that ownership of a patent "does not furnish the owner with the right to make, use, offer for sale, sell, or import the claimed invention," citing as an example "existence of another patent owner with a dominant patent," meaning a patent held by someone else that covers the patented invention. For a company built on foundation AI models and third-party infrastructure, an acquirer reviews model licenses, open-source obligations, data rights, and relevant third-party patents alongside the company's own portfolio.

Code generated with AI tools raises a separate question: whether the company owns a copyright in that code. In Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025), the U.S. Court of Appeals for the District of Columbia Circuit held that "the Copyright Act requires all work to be authored in the first instance by a human being," and the Supreme Court declined to review the decision on March 2, 2026. The Copyright Office's January 2025 report on copyrightability concludes that "prompts alone do not provide sufficient human control to make users of an AI system the authors of the output." A codebase written substantially by AI tools can carry thinner copyright protection than the codebase's size suggests. Protection for AI-generated code thus can depend largely on trade secret law and contractual obligations.

Diligence findings rarely terminate a transaction outright. More often the findings become price adjustments, additional representations, warranties, or indemnities, or documents or requirements the seller must deliver before closing, such as confirmatory assignments. As in a real estate transaction, each negative finding during due diligence can transfer a portion of the sale's value from the seller to the buyer.

A Pre-Transaction Intellectual Property Checklist for Founders

When acquiring or investing in an AI company whose intellectual property accounts for a significant portion of its value, a sophisticated acquirer or investor (or their legal counsel) is likely to ask about at least the following topics, each of which founders can prepare for before a transaction begins:

  • Inventorship: For each important patent family, identify every contributor, including contractors and vendor engineers. Keep dated records of each human contribution and of how AI tools were used. Keep current contact information for every named and potential inventor.
  • Ownership: Confirm a signed present assignment ("hereby assign") from every founder, employee, contractor, and collaborator. Obtain confirmatory assignments where an agreement only promises to assign, and record assignments with the USPTO promptly.
  • Eligibility: Evaluate the claims to confirm sufficient details are recited about how the technical result is achieved.
  • Disclosure support: Confirm that each patent specification describes and enables the full scope of the claims, and if an earlier priority date is relied upon, that the priority document also describes and enables the claimed scope.
  • Continuations: Where possible, keep at least one continuation pending in each core family, align pending claims with the current product or competitors' products, and consider the effect of any terminal disclaimers before structuring any partial asset sale.
  • Trade secrets: Inventory which AI tools received what information and under which terms, keep the logs, and, through policy and technical controls, restrict the entry of unfiled inventions into AI tools that lack sufficient confidentiality protections.
  • Freedom to operate: Investigate third-party patents, foundation model licenses, open-source obligations, and proprietary data rights.
  • Enforcement value: Assess whether infringement of the most valuable claims can be detected and proven from outside the infringer's systems.
  • Copyright in code: Identify code generated substantially by AI tools and document the human contributions to the code.

Some defects, while often cheaper to avoid, can be cured at any time before a deal: missing assignments, unrecorded transfers, undocumented conception, unaligned continuation claims, and AI-use policies. Other defects can become permanent when they occur. A specification that omitted an implementation can support claims to the implementation only by way of a new application with a later filing date, which may be too late if the implementation has already been disclosed to the public. A missing coinventor found after a negative court judgment cannot be added in time to avoid invalidity, as Fortress Iron illustrates. A trade secret disclosed to an AI tool without confidentiality protections may be lost permanently.

Looking Ahead

Several recent and pending developments bear directly on how acquirers will evaluate AI portfolios over the next year. The response to Fortress Iron's Supreme Court petition is due October 19, and if the Court takes the case, the limits of inventorship correction under Section 256 would come before the Supreme Court. The denial of Supreme Court review in the '637 case leaves the Federal Circuit's claims-focused rule in place, while the USPTO continues to issue machine learning claims under the more permissive approach of Desjardins. AI-assisted development may continue to increase the number of contributors and vendors whose signatures, records, and agreement terms a buyer will request.

Acquirers and investors should be asking these and many other questions during diligence. The answers are usually determined much earlier, at each point where an invention is conceived, assigned, claimed, or shared with a vendor. A portfolio developed with these due diligence questions in mind is more likely to retain its value under scrutiny when closing a transaction.