How the 2025 USPTO guidance and an AI-equipped “skilled person” reshape patent challenges
In brief: In 2026, challenging an AI patent means working with two doctrines that sit under unusual pressure. On inventorship, the USPTO’s November 2025 guidance withdrew the AI-specific framework and returned to ordinary conception principles, so challengers must prove who did not conceive the invention rather than point to AI involvement alone. On obviousness, courts and offices have yet to settle whether AI belongs in the toolkit of the “person of ordinary skill.” The strongest invalidity cases combine these arguments with eligibility (§ 101) and disclosure (§ 112) attacks.
1. Why AI patents are a distinct challenge
Patents on machine-learning models, training methods, AI-enabled products and AI-assisted discoveries have grown in number and commercial importance. They are also unusually difficult to invalidate with the standard playbook. Three features create the difficulty:
- Opaque inventive process. When an AI system proposes a molecule, a design or a model architecture, it can be unclear what the human actually contributed, and the evidence needed to show who conceived what may sit in training logs, prompts and experiment records rather than in a lab notebook.
- Fast-moving baseline. The state of the art in AI changes within months. Establishing what was obvious at an effective filing date means reconstructing a moving target.
- Overlapping doctrines. Weak AI patents are often vulnerable on several grounds at once — eligibility, written description, enablement, obviousness and inventorship — and the best strategy chooses among them deliberately.
This article concentrates on obviousness and inventorship, since those are the two grounds on which AI patents face genuinely new questions in 2026, and it briefly places them alongside the other available grounds. The discussion focuses on United States law, with comparative notes on Europe and the United Kingdom. It is a general overview, not a substitute for advice on a particular patent.
2. The legal baseline for any invalidity challenge
- Presumption of validity. In US litigation, issued patents are presumed valid and invalidity must be proved by clear and convincing evidence (Microsoft Corp. v. i4i Ltd. Partnership, 2011). In inter partes review (IPR) before the Patent Trial and Appeal Board, the standard is a preponderance of the evidence, but the grounds are limited to anticipation and obviousness based on patents and printed publications.
- Forum shapes the available grounds. Inventorship, § 101 and § 112 attacks cannot be brought in an IPR. They belong to district court litigation, to post-grant review for qualifying patents within the nine-month window, or to ex parte reexamination where the issue is raised in a form reexamination allows.
- Effective filing date. Obviousness is assessed as of the effective filing date. For AI patents, collecting evidence about what tools, models and datasets were publicly available on that exact date is a core task.
3. Inventorship challenges to AI patents
3.1 The settled core: only natural persons can be inventors
In Thaler v. Vidal (Fed. Cir. 2022), the Federal Circuit held that the Patent Act’s definition of “inventor” as an “individual” refers to a natural person, so an AI system (the “DABUS” system in that case) cannot be named as an inventor. The Supreme Court declined to review the case in 2023. The UK Supreme Court reached a similar conclusion under the Patents Act 1977 in Thaler v Comptroller-General (2023), and the European Patent Office’s Boards of Appeal reached a comparable view on the European Patent Convention. A patent that names an AI system as sole inventor is therefore not a viable US patent; most such disputes are resolved at the application stage.
3.2 The 2025 USPTO guidance changed the framework, not the human requirement
In February 2024, the USPTO issued guidance applying the Pannu factors (from Pannu v. Iolab Corp., Fed. Cir. 1998, a case about joint inventorship among several people) to a single person using AI, asking whether the person made a “significant contribution”. On 28 November 2025, the USPTO rescinded that guidance. The revised guidance states that the same legal standard for inventorship applies to all inventions, regardless of whether AI systems were used, and that there is no separate or modified standard for AI-assisted inventions. AI systems are treated as instruments used by human inventors, comparable to other research tools. The Pannu factors remain the framework for deciding joint inventorship among multiple natural persons, even where they were assisted by AI. The guidance covers utility, design and plant patents.
For challengers this has several consequences:
- There is no lower “AI threshold” to exploit. Arguing that a human’s contribution to an AI-generated result was merely “insignificant” under an AI-specific test no longer matches the Office’s position. The question is whether a natural person conceived the claimed invention — formed a definite and permanent idea of the complete and operative invention.
- Heavy AI involvement does not by itself defeat inventorship. If a named human conceived the claimed invention, using AI to generate, test or refine it does not disqualify them.
- The challenge shifts to evidence about conception. The strongest arguments show that the named inventors only recognised, selected or reduced to practice an output generated by an AI, or that no human formed the definite idea of the specific claimed solution.
3.3 How an inventorship attack is built
Inventorship errors are generally attacked in district court. Under the America Invents Act, a patent naming the wrong inventors is invalid unless the error can be corrected, and 35 U.S.C. § 256 allows correction of named inventors in an issued patent by the court or the Director without the former requirement of proving deceptive intent. The practical result is that a bare inventorship defect is often curable, so invalidity is realistic mainly where (a) correction is impossible because no human conceived the claim, (b) correction would be unavailable or inequitable on the facts, or (c) the defect is part of a broader pattern of problems, such as improper assignment or ownership.
An effective challenge usually builds along the following lines:
- Claim-by-claim conception analysis. Inventorship is determined claim by claim. Identify the distinguishing limitation of each claim and ask who conceived it.
- Discovery of AI-related records. Seek prompts, chat logs, model versions and configurations, experiment trackers, version-control history, internal presentations and invention disclosure forms. These show where the key idea first appeared and whether it came from a human, an AI output, or a third party.
- Corroboration. Inventor testimony about conception requires corroboration by contemporaneous evidence. Where the patentee’s documentary record shows that the claimed solution first appeared in an AI output, and the human’s role was selection or routine implementation, the testimony may be undermined.
- Third-party contributors. Look for engineers, data scientists, collaborators, and university or vendor personnel who contributed to the claimed idea but are not named. In AI projects, contributions often come from people who prepared data or trained models.
- Pleading and burden. Because the burden is clear and convincing evidence, and named inventors are presumed correct, the evidence should be developed early and concretely.
3.4 Complications that favour the patentee
- A human who frames the problem, specifies constraints, designs the training approach, selects among outputs with insight into why one works, and refines the result will often meet the conception standard.
- Recognition and appreciation of an AI output can in some circumstances count as part of conception, although the law on mere recognition of another’s idea is demanding.
- Remedies such as correction under § 256 mean that a technical inventorship problem may not end the case.
- Patentees increasingly keep careful records of human contributions in response to the 2025 guidance, making inventorship harder to attack as time passes.
3.5 International angles
Other jurisdictions treat the question differently. Many offices require a named inventor to be a natural person, and some allow correction after filing while others do not. For multinational families, a defect in one jurisdiction may not exist in another, and the USPTO will not accept a priority claim from a foreign application that names an AI system as the sole inventor. A challenger should map inventorship across the family, and use inconsistencies between jurisdictions (for example, differing inventor lists for the same invention) as discovery leads.
4. Obviousness challenges to AI patents
4.1 The framework
Under 35 U.S.C. § 103 and Graham v. John Deere Co., obviousness is decided by determining the scope and content of the prior art, the differences between the prior art and the claims, the level of ordinary skill, and any objective indicia of non-obviousness. KSR International Co. v. Teleflex Inc. (2007) emphasised that the person of ordinary skill is a person of ordinary creativity, not an automaton, and that predictable use of known elements according to their established functions may be obvious.
AI complicates several steps in this framework.
4.2 The level of ordinary skill and the “AI-equipped” skilled person
The central unresolved question is whether the hypothetical person of ordinary skill should be treated as having access to AI tools. The USPTO invited comments on this in April 2024 (89 Fed. Reg. 34217), asking how the availability of AI to a skilled person affects obviousness. Commentators split. Some argue that AI is just another research tool, like computers or databases, and should not change the standard. Others argue that if AI is part of the ordinary toolkit of the field, then results that a skilled person could readily generate with commonly available AI tools should be treated as within ordinary skill, raising the bar for patentability. Academic and professional commentary in 2026 continues to explore this, including whether a person who is assumed to know all prior art should also be assumed to use AI that can search and combine it at scale.
A challenger can make the following arguments, depending on the facts:
- AI as a routine tool in the field. If, at the effective filing date, using machine-learning methods to optimise parameters, screen candidates or design architectures was standard in the field, then applying such methods to a known problem with a reasonable expectation of success is arguably obvious under KSR.
- Known tool, known problem, predictable result. Claims that merely apply a known model type to a new data set, or replace one known model with another to obtain predictable improvement, are the most exposed.
- Design incentives and market pressure. Where multiple groups were independently applying the same class of tools to the same problem, this is evidence that the solution was within reach.
The patentee will answer that hindsight is the great danger. Reconstructing a path from prior art to the invention using a specific AI tool, specific data and specific settings, when the actual inventors needed insight to choose them, is textbook hindsight reasoning. A patentee may also argue that the existence of a powerful tool does not make every output it produces obvious: a result that the tool could in principle generate is not necessarily a result that a skilled person would have had reason to pursue or a reasonable expectation of reaching.
4.3 Reasonable expectation of success and the “black box” problem
Obviousness requires a motivation to combine or modify and a reasonable expectation of success. In AI, outcomes can be unpredictable: a model that works well on one data set may fail on another, and small architectural changes can produce large performance differences. This helps patentees where the claim depends on a specific, unexpected result. It hurts them where the specification itself describes the technique as routine or says that known methods work. Challengers should examine admissions in the patent’s own background and the prosecution history, since statements that a technique is well known or standard can be used against the claims.
4.4 AI-generated and AI-scale prior art
AI changes what the prior art looks like. Large volumes of machine-generated text, code, designs and candidate compounds can be published online, and some of these may qualify as printed publications if they were publicly accessible. For challengers, this raises opportunities and difficulties:
- Opportunities. Public model outputs, open datasets, open-source repositories, technical blogs, model cards, preprints and benchmark leaderboards often disclose the very techniques that were later claimed. Systematic searching of repositories and archived web pages can reveal prior art that a conventional patent search would miss.
- Date proof. Version histories, commit timestamps and web archives are evidence of public accessibility, but must be authenticated, and later-modified pages require care.
- Enabling disclosure. For anticipation, prior art must enable the claimed subject matter. For obviousness, a reference need not be enabling in the same way, but its teachings must be assessed in light of what the skilled person could do.
- Open questions about AI-only outputs. Whether and when an output generated by AI without human involvement counts as prior art in the same way as human publications is debated, particularly in design patent contexts, as the USPTO’s own March 2026 bulletin on generative AI and design patents acknowledges.
4.5 Objective indicia and nexus
Patentees rely on commercial success, industry praise, copying and long-felt need to show non-obviousness. In AI products, commercial success may be due to factors other than the claimed feature: compute resources, proprietary data, brand, distribution, user interface or marketing. A challenger can argue that there is no nexus between the claimed invention and the success. Likewise, rapid copying within a field where everyone uses the same open-source frameworks is weak evidence of non-obviousness.
4.6 Practical evidence for obviousness
- A technical expert who can show what AI practitioners were doing at the effective filing date, using contemporaneous papers, code repositories, product documentation and conference proceedings.
- A timeline of publicly available models, frameworks and datasets, to show what was accessible and how widely used.
- Reproduction of the prior-art approach where possible, to show that a skilled person would have reasonably expected success.
- Analysis of the patent’s own disclosure: does it identify which part of the claimed pipeline is new, or does it describe each component as conventional?
- Evidence of simultaneous or near-simultaneous independent development.
5. Other grounds that often work better than either
Because inventorship is curable and obviousness is contested, many successful challenges to AI patents run on other grounds, sometimes in combination.
| Ground | Why it is a strong candidate for AI patents | Points to watch |
| Subject-matter eligibility (§ 101) | Claims that recite generic machine learning applied to a field or data type may be directed to abstract ideas. The Federal Circuit held in Recentive Analytics v. Fox Corp. (2025) that merely applying generic machine-learning techniques to a new data environment did not make claims eligible, and later decisions in 2026 have continued to stress that claims need a specific technical improvement. | USPTO decisions such as Ex parte Desjardins signal receptiveness to claims framed as improvements to how a model operates; check the current guidance and case law. |
| Enablement and written description (§ 112) | Broad functional claims, such as those covering any model that achieves a result, are vulnerable under Amgen Inc. v. Sanofi (2023), which requires enablement of the full scope of the claim. | Training data, hyperparameters and model details may be missing from the specification. Check whether the patent discloses enough to reproduce the result. |
| Anticipation (§ 102) | Open-source code and published papers can disclose claimed methods exactly. | Must show that each limitation was disclosed in a single reference in an enabling way. |
| Indefiniteness (§ 112(b)) | Terms like “trained to optimise”, “sufficiently accurate” or “neural network model” without structure can be unclear. | Claim construction outcomes drive this ground. |
| Ownership and standing | Problems with employee, contractor or vendor assignments can affect enforceability even where the patent is valid. | Not an invalidity ground, but a powerful defence. |
6. Procedure and strategy
6.1 Choosing the forum
- IPR: best for obviousness and anticipation on printed publications, with a lower standard of proof. Inventorship and eligibility arguments are unavailable. Open-source code and online documents may qualify as printed publications if public accessibility is shown.
- Post-grant review: broader grounds, but only within nine months of grant and for first-to-file patents.
- Ex parte reexamination: useful for substantial new questions of patentability on patents and printed publications, with no estoppel for the requester in the same way as IPR, but the requester has limited participation.
- District court: the only forum for full inventorship and ownership discovery and for combining all grounds, but with the highest burden of proof and the highest cost.
6.2 Sequencing arguments
A typical effective approach is to file an IPR (or ex parte reexamination) on prior-art-based obviousness, while preserving eligibility, § 112 and inventorship arguments for district court, and to use discovery to test inventorship. Be aware of estoppel consequences: grounds raised or reasonably could have been raised in an IPR that reaches a final written decision cannot later be asserted in court on the basis of patents and printed publications.
6.3 Evidence preservation and discovery
AI development leaves unusual records. Early preservation demands should cover training logs, experiment-tracking platforms, repository histories, model checkpoints, prompt and conversation histories with AI tools, data provenance records and internal communications. Courts have increasingly treated these as discoverable, but proportionality and confidentiality disputes should be expected, especially where trade secrets in models and datasets are involved.
7. Comparative notes: Europe and the UK
- Inventive step in Europe. The EPO uses the problem-solution approach: identify the closest prior art, define the objective technical problem, and ask whether the claimed solution would have been obvious to the skilled person. Where AI helps to identify the solution, the question is whether the skilled person, faced with the same problem, would have taken that route. Commentators observe that the path from prior art to invention is usually reconstructed after the fact, and that an AI-generated recommendation is evidence of a possible route, not proof it was obvious.
- The skilled person. UK and European commentary in 2026 asks whether routine access to AI should change the skilled person’s ordinary technical skill, and notes the risk that arguing the invention was obvious to a person using a specific AI tool in a specific way is an ex post facto analysis.
- Inventorship. The EPO and UK require a human inventor; challenges typically relate to entitlement or formal requirements rather than a detailed conception analysis.
- Technical character. In Europe, claims to AI and machine-learning methods must show a technical character, and only technical features contribute to inventive step. This feeds directly into the inventive step analysis and is a distinct route to attack AI claims.
8. Defending AI patents: what challengers should anticipate
A realistic challenge anticipates the patentee’s responses.
- Inventor declarations and well-organised conception records showing the human decisions that shaped the claimed solution.
- Expert evidence that the choice of architecture, data pre-processing, loss function or training regime reflected non-obvious design decisions.
- Evidence of unexpected results, such as performance far above the prior art on standard benchmarks.
- Arguments that the challenger’s AI-based obviousness story depends on hindsight and on selecting tools and parameters that only the invention reveals.
- Requests to correct inventorship under § 256 rather than invalidate.
9. Checklist for challengers
- Identify the effective filing date of each claim and fix the relevant technology snapshot: public models, frameworks and datasets as of that date.
- Map each claim limitation to the prior art, noting which limitations are conventional AI components and which are truly specific.
- Review the specification and prosecution history for admissions that techniques were routine or well known.
- Collect open-source, preprint and repository prior art; secure authenticated date evidence.
- Evaluate § 101 and § 112 alongside § 103: do the claims recite a specific technical improvement, and does the specification enable the full scope?
- Investigate inventorship: identify named inventors, their roles, and unnamed contributors; seek records showing where each distinguishing idea first appeared.
- Choose the forum and sequence the arguments with estoppel in mind.
- Retain a technical expert who can speak to what practitioners actually did at the effective filing date.
- Prepare for objective-indicia evidence and develop nexus arguments.
- Reassess regularly, as USPTO guidance and Federal Circuit decisions on AI are changing quickly.
10. Conclusion
In 2026, invalidating an AI patent is less about proving that a machine did the inventing and more about ordinary patent law applied to unusual facts. After the USPTO’s withdrawal of its AI-specific inventorship framework, the inventorship question is the traditional one — who conceived the claimed invention — and the evidence needed to answer it lies in logs, repositories and records that must be pursued early in discovery. On obviousness, the law is open on how far AI should expand the skilled person’s toolkit, which creates both an opportunity for challengers and a risk of hindsight accusations. The most dependable strategies pair these arguments with eligibility and disclosure attacks, build a rigorous record of what the AI field actually knew on the effective filing date, and choose the forum that fits the grounds. Because the guidance, case law and commentary are all moving, anyone planning a challenge should verify the current state of the law at the time of filing..
