Direct Answer: Eligibility Depends on the Claimed Technical Solution
Yes, a U.S. patent application may claim a deepfake-detection system, but “AI-powered” is not itself a sufficient basis for eligibility. The USPTO evaluates the claimed invention under 35 U.S.C. § 101 and distinguishes patent-eligible technical improvements from claims directed only to abstract ideas, mathematical formulas, or mental processes. A claim reciting specific processing of media data, generation of a technical authenticity signal, and detection of manipulated content is generally stronger than one claiming the result of applying an AI model to audiovisual material. The legal test also depends on how the claims are drafted, not merely on the commercial description in the application. As of September 25, 2026, applicants should check the current USPTO guidance and controlling cases rather than assuming that the subject of deepfakes is automatically eligible or ineligible.
Also worth reading: What is the definitive patent strategy for data center cooling technology in the age of AI? · What does an AI patent disclosure compliance checklist look like in 2026, and how should technology companies implement it? · What are the modern estate planning strategies for 2026 and how does AI patent review technology reshape asset protection?
The best candidate inventions are often improvements in how computer systems identify, classify, or reduce the effect of synthetic media. Examples may include a detector that analyzes temporal inconsistencies in compressed video, a distributed system that assigns confidence scores using device telemetry, or a method that improves the computational efficiency of biometric verification. By contrast, a claim whose sole objective is “classify a video as real or fake” using unspecified neural-network rules may be treated as a mathematical or mental process. Eligibility under § 101 is only the first examination stage: the application must also satisfy requirements such as utility, novelty, nonobviousness, written-description support, and enablement. Accordingly, the correct answer is conditional: deepfake patent eligibility exists, but acceptance depends on a technically specific claim and a properly supported application.
How USPTO Patent-Eligibility Review Works
USPTO examiners generally apply the two-step analytical framework associated with Mayo Collaborative Services v. Prometheus Laboratories and Alice Corp. v. CLS Bank. Step one asks whether the claim recites a judicial exception, such as a mathematical concept, a certain method of organizing human activity, or a mental process. Step two asks whether the claim integrates that exception into a practical application or adds an inventive concept sufficient to transform the abstract idea into a patent-eligible claim. An applicant should not treat these steps as a simple checklist of hardware words, because merely adding a generic computer or server does not automatically supply the missing technical substance. The analysis focuses on the claim as a whole and on the limitations that actually distinguish the proposed solution from prior art.
USPTO subject-matter eligibility guidance concerning AI became publicly available in 2024, including the July 17, 2024 guidance update. That material emphasized careful treatment of AI-related claims and warned against conclusions based only on the presence of words such as “neural network” or “artificial intelligence.” The guidance also addressed the need to consider whether a claim merely uses a computer to perform an abstract process or instead improves computer functionality or another technology. Guidance informs examination practice, but it does not replace the statute, prior art, or judicial decisions. Patent counsel should map each proposed independent claim to a specific technical mechanism, identify the relevant prior art, and document why that mechanism would not have been obvious to a person skilled in the art.
Eligibility review is also separate from novelty and obviousness review. A deepfake detector may be eligible because it uses a particular signal-processing technique, yet that technique could still be anticipated or obvious. Conversely, a technically narrow claim may survive § 101 but fail because an earlier paper disclosed substantially the same arrangement. The USPTO may issue different outcomes to different claims in the same application, so broad system claims may be rejected while narrower claims reciting a distinctive implementation remain viable. An application designed around eligibility alone can therefore waste filing fees without producing an enforceable patent. The practical objective is a coordinated strategy in which eligibility, prior-art distinctions, and disclosure quality support the same technical story.
What Makes a Deepfake Claim Technically Specific?
A strong claim usually explains what data is processed, what technical operation is performed, and how the operation produces a defined technical result. For video, that may involve extracting frame features, comparing temporal motion patterns, evaluating audio-visual synchronization, or testing compression characteristics. For audio, claims may describe spectral anomalies, voice-conversion artifacts, or phase inconsistencies. A detector can also claim a particular architecture, such as a first neural network generating manipulation evidence and a second subsystem comparing that evidence against a trusted reference. These details give the examiner a concrete basis for evaluating whether the claim is more than a result-oriented use of a generic classifier.
The claim should explain the technical relationship between its components. Saying that a model identifies “features indicative of a deepfake” leaves important questions about which features, how they are obtained, and how they affect the output. Stating that the system extracts inter-frame residual signals, combines them with an audio synchronization measure, and alters a detection threshold according to measured compression artifacts is more concrete, although the exact wording must still be supported by the specification. Functional language is not automatically forbidden, but claims that state only a desired outcome may receive an eligibility objection or a written-description rejection. Patent eligibility improves when the application ties the algorithmic operation to a particular technical problem rather than merely announcing that AI solves deception.
The application should also address what the output actually does. A plain label of “authentic” or “synthetic” may be characterized as an abstract result, while an output used to control a media-authentication pipeline, quarantine a transmission, select a lower trust level for a financial transaction, or trigger cryptographic verification may support a different analysis. That distinction is fact-dependent and should not be overstated as an automatic rule. The USPTO and courts have permitted claims directed to technological improvements in computer operation, but they have rejected claims whose improvement is merely a business or human objective implemented with generic computing. A credible filing therefore identifies the technical bottleneck, explains why existing methods fail, and shows that the proposed method changes the behavior or performance of the claimed system.
AI Authorship, Human Inventorship, and Ownership
AI-generated material can be part of a patent application, but the United States requires human inventorship for the claimed invention. In Thaler v. Vidal, the Federal Circuit held that an “inventor” must be a natural person, and the USPTO requires the application to identify individuals who contributed to conception. An AI system cannot be listed as an inventor merely because it proposed an architecture, generated code, or produced a model configuration. Counsel should preserve records showing which humans directed the research, selected the inventive concepts, evaluated alternatives, and contributed to the claimed subject matter. Simply having a person review an AI output does not necessarily make that person an inventor if the person contributed no conception, but detailed human technical contribution may support an appropriate inventorship determination.
Ownership is a separate issue. A company, university, laboratory, or platform may own the application under assignment rules even though an employee or contractor is the named inventor. Employment agreements, contractor agreements, collaboration agreements, and university policies can allocate rights to inventions, data, models, and generated outputs, but those documents may not match the statutory inventorship record. The filing team should also resolve whether third-party code, datasets, biographies, voice samples, or media used in training are subject to contractual restrictions. Patent ownership does not automatically clear copyright, contract, privacy, publicity, or trade-secret claims. A valid patent grant would not authorize use of another person’s face, voice, or copyrighted footage, and a license to analyze material is not necessarily a license to store or reuse it for model training.
The written description and enablement requirements deserve special attention in an AI patent. A specification should define the system sufficiently to support the breadth of the claims, explain relevant alternatives, and provide enough detail for a skilled person to make and use the invention. Generalized statements that a conventional network was trained on “large amounts of data” may be inadequate when training or operation requires a particular data relationship. At the same time, the specification should not misrepresent experimental results or imply technical performance that was never measured. Inventorship, ownership, and disclosure should be addressed before a public release, conference presentation, paper, or commercial launch, because some disclosures can affect available patent rights and foreign filing options.
Patent Versus Trade Secret, Copyright, and Defensive Publication
A patent and a trade secret answer different commercial questions. A patent gives a limited right to exclude others after issuance, but it requires public disclosure, costs money to obtain and maintain, and becomes subject to validity challenges. A trade secret can potentially last indefinitely without publication, but it offers no independent exclusion right against someone who develops the same technology independently or lawfully learns it from another source. Deepfake models, training recipes, detection thresholds, and operational data may be easier to protect as trade secrets than as patents. If a product can operate without revealing its most valuable method, secrecy may be preferable; if competitors can reverse-engineer a deployed detector, patent protection may be more appropriate.
| Feature | U.S. patent application | Trade-secret program | Copyright or defensive publication |
|---|---|---|---|
| Main benefit | Potential right to exclude competitors for a limited term, generally up to 20 years from filing | No fixed expiration while secrecy is maintained | Copyright protects qualifying expression; publication can block later patent claims but creates no exclusion right |
| Disclosure | Requires enabling public disclosure | Generally requires no public disclosure | Publication is public; copyright does not protect the underlying technical idea |
| Best fit | Novel, measurable technical improvements that are difficult to keep secret | Model weights, recipes, thresholds, internal data, or rapidly changing know-how | User-interface artwork, source code, documentation, or a public priority disclosure before third-party patenting |
| Main drawback | Cost, prosecution uncertainty, and validity challenges | Leakage, reverse engineering, and independent development risk | No patent remedy, limited technical protection, and little control over later commercial use |
| Typical timing | File before a public disclosure; decide whether to preserve a priority filing | Reduce access and document protection measures | Publish only after considering foreign filing deadlines and the intended effect on competitors |
Common Mistakes That Weaken Deepfake Patent Applications
One common mistake is treating a detector’s accuracy percentage as proof of patent eligibility or patentability. A reported 98 percent accuracy, 2 percent false-positive rate, or 10-fold speed improvement may help establish utility, technical effect, or nonobviousness, but the figures do not answer whether the claim recites a judicial exception. The same percentage also may be misleading unless the test set is representative, the baseline is defined, and the comparison uses comparable conditions. Applicants should report dataset composition, evaluation protocol, hardware configuration, and error categories where appropriate. Unsupported performance language can create credibility problems during examination or litigation and is particularly risky when a claim covers a broad range of devices and data.
Another error is filing a claim that says only “use AI to detect deepfakes.” Such a claim is vulnerable because it leaves the core detection mechanism undefined, invites prior-art challenges, and may be characterized as a generic use of a mathematical classifier. A mistake also occurs when counsel adds a conventional computer processor to a claim without connecting the processor to a distinctive technical operation. Applicants should avoid overclaiming a business objective, such as reducing fraud, without explaining the computer-level improvement that supposedly produces that result. Finally, teams sometimes fail to coordinate patent filings with product deployment, model training, and public demonstrations. A rushed nonprovisional filing can lose the benefit of an earlier provisional filing, while an overly broad provisional may fail to support later claims that introduce new matter.
The answer to a § 101 objection is not necessarily to abandon the subject matter. It may be possible to narrow the independent claim, recast a method as a system or stored program, identify a particular signal-processing relationship, or emphasize a measurable improvement in computer operation. Yet amendment is not always available at the same stage, and adding language during prosecution can create new written-description or prior-art issues. Before filing, a search should cover patents, papers, conference materials, product documentation, open-source projects, and known attacks against the detector. A technically authentic invention can still fail if a competitor or researcher had already developed the same mechanism. Good patent review evaluates the actual prior art rather than relying on the fact that deepfake detection is a relatively new commercial category.
Practical Steps, Timing, and Cost Considerations
A practical first step is to document the inventive contribution before filing. The team should identify the specific failure of existing detectors, the new mechanism addressing it, and the measurable technical difference. A prior-art search should then determine whether the improvement is genuinely novel and whether narrow alternatives may be necessary. Drafting should proceed from a concrete implementation, with several claim levels covering a system, a computer-implemented method, and possibly a non-transitory computer-readable storage medium where supported. The specification should include drawings, definitions, alternative embodiments, enabling detail, and carefully qualified performance results. A focused review by a patent practitioner experienced with AI and media processing is preferable to assuming that general software drafting is sufficient.
Timing matters because U.S. applicants generally must file before the relevant subject matter becomes public. A provisional application can establish an early filing date and generally must be supported by a nonprovisional or PCT filing within 12 months to preserve the claimed priority. A PCT application commonly provides a route for seeking protection in multiple jurisdictions, but national-phase rights and fees are governed by country-specific rules and deadlines, including a commonly applicable 30-month point from priority for many PCT filings. The public-use bar, one-year grace period in limited circumstances, prior disclosures by collaborators, and foreign law can make “we will file later” risky. Filing decisions should be made before a paper, demo, sale, customer disclosure, repository release, or standards submission.
Official USPTO fees are substantially lower than the total cost of a professionally prepared application, and the USPTO’s fee schedule and small-entity or micro-entity designations can materially change the government charge. Private search and drafting costs vary widely: a preliminary patentability search may cost roughly $3,000 to $15,000, attorney-drafted technical claims often fall around $10,000 to $30,000 or more, and a full foreign or PCT strategy can require a budget well above $50,000 over several years. A low-cost automated filing service may reduce drafting expense but may not reliably perform prior-art analysis, inventorship review, or claim-level eligibility analysis. The best value is obtained by budgeting for search, drafting, prosecution, foreign filings, and maintenance as separate decisions. An attorney-client consultation is also appropriate where the invention combines undisclosed model technology, sensitive media, third-party data, or a near-term product launch.