What an AI patentability review actually determines
An AI patentability review asks whether a proposed invention is potentially patentable, not whether an examiner will definitely grant it. For AI-related inventions, the review ordinarily examines patent-eligible subject matter, novelty, inventive step or non-obviousness, utility, enablement, written-description or inventive-step requirements, and any applicable software or business-method exclusions. The answer also depends on the jurisdiction: the United States, European Patent Office, United Kingdom, and other offices apply related technical tests through different rules, procedures, and case law. As of 28 September 2026, applicants should therefore evaluate the same technical proposal under the law of each intended office rather than assuming that acceptance by one office ensures acceptance elsewhere. A patent search is only one component of the review, and a favorable search result does not establish that an AI claim is eligible or sufficiently inventive. The practical product is usually a reasoned risk assessment, a set of corrections, and a filing strategy.
Also worth reading: Is an AI Invention Patentable in 2026, and How Should You Navigate an AI Patentability Review? · How Do AI Patent Review Tools Evaluate Patentability in 2026? · Do I Need an AI Inventorship Disclosure Review Before Filing a Patent Application?
The assessment should separate three questions that are often wrongly combined. First, does the invention exist and differ from the prior art? Second, may an office patent that difference at all? Third, will the application satisfy formal, disclosure, and procedural requirements? A system trained on a conventional neural network may be novel in its application but still encounter an eligibility objection under the U.S. software-related cases. Conversely, an eligible medical, manufacturing, or network-security invention may be patentable, although a poor search could still destroy novelty. The strongest review therefore tests both the legal theory and the evidence supporting it. Its conclusion should be probabilistic because prosecution is an adversarial process in which the applicant does not know the examiner’s complete record.
How AI inventions are tested for patent eligibility
In the United States, an examiner may apply the two-step analytical framework derived from Mayo Collaborative Services v. Prometheus Laboratories and Alice Corporation v. CLS Bank International. Under the first step, the claims must not recite a judicial exception such as an abstract idea, law, mathematics, or certain business practices. Under the second step, the claims are considered as a whole to determine whether they contain an inventive concept sufficient to transform the exception into a patent-eligible application. AI claims based only on producing a prediction, generating content, or using mathematical relationships may be vulnerable because those functions can be characterized as mental processes. Claims that integrate the AI into a specific technical process and define concrete technical improvement may fare better, but the claimed improvement must be plausible, supported by the specification, and not merely asserted with a result such as “improves efficiency.”
European practice is similarly demanding but is structured around technical character, technical contribution, and excluded subject matter or programs for computers. The EPO normally expects a computer-implemented invention to produce a further technical effect beyond normal computer operation. Applying a model to classify images, optimize a machine, diagnose a fault, or reduce network resource consumption may support technical character depending on the disclosure. Running an LLM to draft marketing text or organize a generic workflow may not suffice merely because software is involved. The EPO does not use the Alice test as such, and U.S. eligibility cannot safely be exported to Europe. Because AI functionality is often difficult to separate from generic computing, claim drafting should express the technical process, its inputs, its operation, and its measurable technical result at an appropriate level of abstraction.
Why prior art and inventive step deserve special attention
AI patentability frequently turns on a narrow prior-art search involving published papers, model documentation, source-code repositories, conference presentations, theses, standards, product manuals, and patent applications. Searching only by a founder’s product name is inadequate because early versions may be described under a research label, a dataset name, a model architecture, or a technical problem. Search concepts should include synonyms, acronyms, mathematical techniques, data types, training methods, inference methods, deployment structures, and the intended field of use. Patent databases are also incomplete, and public use or sale can create prior art outside them. A credible review should identify the closest references, map each relevant limitation against the proposed claims, and explain whether a difference would have been obvious to a person skilled in the field.
Novelty and inventive step must be kept separate. Novelty asks whether one prior-art disclosure contains every limitation of a claim, while inventive step or obviousness asks whether the differences would have been an obvious solution to a skilled person. A review should therefore look for combinations of known references rather than treating a single matching paper as decisive. Common arguments include a new training objective, a new architecture, a particular data combination, a latency optimization, a new control loop, or an unexpected operational result. The specification should explain why that difference was not predictable, include comparative examples, and connect the stated benefit to an identified technological problem. Saying that the result is “better” without identifying the metric, baseline, operating conditions, or mechanism can leave both inventive step and enablement exposed.
A practical preparation process for AI applications
Preparation should begin by defining the invention rather than the desired claim form. Prepare a technical disclosure that identifies the problem, the people and components involved, the relevant prior systems, the smallest system boundary, the sequence of operations, the training and inference distinctions, and the technical effect. Include diagrams, equations where useful, model and data descriptions, alternative embodiments, examples with actual or realistic parameters, and test results. For a generative-AI invention, disclose useful details about the model family, prompt construction, retrieval, context handling, output validation, safeguards, memory, latency, compute demand, and deployment. If training is part of the inventive concept, explain the data selection, preprocessing, loss or reward design, optimization, and reproducibility. If only inference is relevant, distinguish conventional model operation from the claimed improvement.
A professional review should then produce at least two or three claim strategies. One may claim a system, another a method, and another a narrower technical operation or controlled apparatus, but their scope should be coordinated rather than copied mechanically. Claims should include meaningful functional and structural relationships, avoid purely result-based wording, and cover embodiments actually supported by the application. A broad independent claim may secure an early priority position but invite objections, while a narrow claim can be easier to allow but may be commercially weak or vulnerable to design-around. Before filing, counsel should verify inventorship, confidential subject matter, ownership agreements, disclosure timing, foreign-filing deadlines, and whether any public disclosure could affect available rights. The application should be drafted so that it remains useful if the commercial implementation changes before grant.
Comparing alternative ways to protect an AI invention
Developers frequently treat patents as the only form of protection, yet trade secrets, copyrights, database rights, contracts, and open-source controls can address different parts of the same commercial system. A patent is most useful when the applicant wants a defined exclusive right that can be asserted against independently developed competitors. Trade secrecy can be stronger for model weights, datasets, optimization parameters, or internal operational know-how that is difficult to reverse-engineer, but protection disappears if the information is disclosed without a confidentiality obligation. Copyright generally does not protect the underlying idea, an abstract model architecture, or a functional method, although it may protect original code, text, images, or other expression. A layered approach is often sensible, but each mechanism has costs and enforcement limits.
| Feature | Patent application | Trade secret | Copyright |
|---|---|---|---|
| Main asset | Claim-defined technical invention | Confidential information with economic value | Original expression fixed in a tangible medium |
| Typical AI target | Technical system, method, apparatus, or control process | Weights, data recipes, parameters, operational know-how | Source code, documentation, interface artwork, generated material where applicable |
| Duration | Generally about 20 years from the earliest effective nonprovisional filing date if maintained | Potentially indefinite while secrecy is preserved | Life plus 70 years for many works in some jurisdictions, subject to local rules and exceptions |
| Detection | Claim interpretation and infringement proof can be difficult | Misappropriation is difficult to detect | Separate originality and copying analysis |
| Principal risk | Cost, validity challenge, eligibility rejection, and design-around | Accidental or compelled disclosure | Protection of expression rather than functionality |
Common mistakes in AI patent reviews
A frequent mistake is treating the use of AI as a legal category. The USPTO and EPO do not grant or deny patent protection because a claim involves machine learning; they apply existing law to the claimed subject matter. Another error is assuming that a dataset makes an invention novel, without showing how its composition or processing differs from prior disclosures. Several applicants also describe a commercial objective, such as personalization, decision support, or automation, but fail to show the technical mechanism by which it is achieved. Generic references to “artificial intelligence,” “a neural network,” or “a trained model” can leave scope broad yet indefensible. Claim terms such as “determine,” “analyze,” or “generate” are legal placeholders unless the specification and claims explain their inputs, steps, and output sufficiently.
Another major mistake is relying on a tool-generated patentability opinion without human verification. AI search and drafting tools can identify terminology, accelerate document review, and expose omissions, but they may invent authorities, misread passages, miss relevant art, or assert eligibility conclusions. Any source citation should be opened and checked, and any factual statement about a model, implementation, or technical effect should be confirmed by an inventor or engineer. Applications should also be reviewed for invented technical features that were never implemented or experimentally supported. Automated tools are useful assistants, not substitutes for practitioner judgment, legal analysis, or technical evidence.
When to act before disclosure or filing
A review should normally occur before the first non-confidential public disclosure, launch, demonstration, sales offer, academic submission, or repository publication. Rights can differ by country, and a limited public-use or grace period in one jurisdiction does not create a comparable safe harbor elsewhere. Patentability can also be lost through earlier public disclosure, although confidentiality may avoid the public-disclosure issue while preserving a trade secret. If product publication is imminent, a provisional or priority filing may help establish an early date, but only where the application adequately supports the invention and has genuine patent value. A rushed application that relies on broad marketing claims can cost more than it protects.
Timing should be balanced against the need to collect evidence. Filing too early may omit preferred architectures, benchmark results, or design alternatives; waiting too long can create prior art and reduce negotiating value. For a university spinout, startup, or enterprise project, the useful trigger is usually completion of a technically specific conception together with enough evidence to describe how the system works. That point may precede a polished product. A staged disclosure plan can preserve secrecy before filing, identify remaining patentable features, and coordinate launch, publication, and international filings. Counsel should also check funded-research obligations, joint-development agreements, government rights, and employee-assignment rules, since ownership problems cannot be repaired by adding technical language to a claim.
What a defensible AI patentability opinion should deliver
A defensible opinion should not say simply “patentable” or “not patentable.” It should identify the relevant jurisdictions, the proposed claim categories, the closest prior art, the legal tests, the risk under each test, and the assumptions requiring confirmation. The opinion should separate fatal issues from weaknesses that can be addressed through amendment, narrower claims, added technical detail, or a different filing strategy. It should also identify facts that materially affect the result, such as whether a feature was public, whether training was part of the inventive process, whether the system produces a measurable technical effect, and whether the embodiment is fully disclosed.
Before a filing decision, the applicant should request a claim chart comparing each proposed independent claim with the closest art and with the governing eligibility authority. The chart should distinguish direct disclosure from a possible obviousness combination. A second review should test whether the specification supports all foreseeable fallback positions and whether proposed alternatives are credible rather than hypothetical. Cost and enforcement should be estimated over the intended life, including maintenance, translation, opposition, invalidity, monitoring, and litigation. The final recommendation should explain not only the probability of initial allowance but also the value and durability of the resulting right. As of 28 September 2026, procedural rules and fee schedules can change, so applicant should confirm current official information immediately before filing and should avoid relying on a search tool’s uncited assertion as a final legal conclusion.