AI for IP Law: Applications, Limits, and Legal Requirements

AI for intellectual property law does four jobs well today: semantic prior art and trademark searching, first-draft patent and trademark documents, portfolio and deadline monitoring, and review of IP licensing and assignment contracts. It does not change who holds the intellectual property rights in the result. The USPTO requires that a natural person significantly contribute to any AI-assisted invention, and the Copyright Act requires a human author for a copyrighted work, a reading the Supreme Court left standing in March 2026. The practical work for an intellectual property law practice is therefore narrower than the marketing suggests: pick the stages where AI earns its place, meet the confidentiality and supervision duties that attach to using it, and know the points where it still falls short, whether that practice sits inside an in-house IP team or an outside law firm handling the filings.

[cta-1]

Key Takeaways

1. AI for intellectual property law earns its place at four stages, prior art and trademark search, first drafts, portfolio monitoring, and IP contract review, where it widens recall and produces the first draft while the attorney keeps the opinion, the claim scope, and the abandonment call.
2. Ownership rules are unchanged. A natural person must significantly contribute to an AI-assisted invention under the USPTO’s 2024 inventorship guidance, and the Copyright Act requires a human author.
3. A pending patent application is confidential by statute, so a tool’s data retention terms decide whether it can touch draft claims at all.

AI Use Cases in Intellectual Property, Stage by Stage

Additional AI Use Cases in IP Practice

1. Patent landscape analysis. Mapping a technology area’s filing density and white space before a client commits research spend to it.
2. Infringement detection. Flagging products or filings that may read on a client’s claims across a monitored market.
3. Design-around analysis. Comparing a competitor’s revised product or filing against a client’s claim set to flag whether a workaround still falls inside the claimed scope.
4. Predictive analytics for IP litigation. Modeling outcome patterns across a venue or judge’s prior IP rulings to inform a litigation-risk assessment, not to predict a specific case’s result.
5. Claim charting. Mapping claim elements against prior art or an accused product, element by element, for an infringement or validity analysis.

Infringement monitoring and competitor watch read the same filing feeds, which is why infringement detection often sits inside the broader AI tools for legal compliance that regulated industries deploy to track enforcement exposure beyond IP specifically.

AI and Prior Art Search: What Changes and What Does Not

AI prior art search for patent attorneys replaces Boolean keyword strings with semantic retrieval: the tool reads a claim or an invention disclosure and returns conceptually similar documents, including art that shares none of the same words. That widens recall, which is the hard part of a novelty search under 35 U.S.C. 102 and 103. USPTO examiners now run the same kind of search on the other side of the desk, through an AI similarity search built into the Patents End-to-End search tool that the agency uses to surface comparable art during examination. That search cross-references both patent and non-patent literature rather than patents alone, a capability documented in the Office’s own implementation notes. The opinion does not change. Precision falls as recall rises, so a semantic result set still needs a human read for relevance, claim-by-claim mapping, and the duty-of-candor call under 37 CFR 1.56.

What a Semantic Result Set Still Needs

1. Non-patent literature coverage. Academic papers, standards documents, and product manuals often sit outside a patent-only index.
2. Foreign-language art. Translation quality varies by tool and by language pair, and a missed Japanese or Chinese reference is a common gap.
3. Publication-date windows. A tool’s index has a cutoff; confirm it against the critical date before relying on a “no close art found” result.

Knowing what an index actually covers before trusting a null result is a legal research discipline rather than a prior art specific one, and it is what decides whether “no close art found” means anything.

AI for Trademark Clearance and Watch

AI trademark search and clearance screening now runs on three signals at once: phonetic similarity, semantic meaning, and visual design. The USPTO added a beta image search and a Class ACT classification tool to its own trademark search system: image search opened in beta on April 2, 2026 and lets a practitioner upload a logo and retrieve marks with similar design elements, and Class ACT, announced in March 2026, assigns international classes, design search codes, and pseudo marks to unclassified applications, work the Office describes as reducing pre-processing time. The WIPO Global Brand Database offers comparable image search across national registers. None of this produces a clearance opinion. Likelihood of confusion is a multi-factor legal judgment, and the USPTO states that image search supplements a traditional clearance search rather than replacing it.

What Image Search Does Not Replace

1. Design search codes. The USPTO’s own advice is to run image search alongside, not instead of, a traditional design-code search, which catalogs a mark’s visual elements by category and can surface conceptually similar marks that an image match misses entirely.
2. The classification judgment call. Class ACT assigns international classes and flags design search codes on an unclassified application, but it does not decide a close call on a mark that plausibly spans two classes, and it does not substitute for a clearance opinion once the marks in a crowded class are identified.

What AI Patent Management Tools Track for Law Firms

AI patent management tools for law firms track four things a docket spreadsheet cannot: annuity and response deadlines across jurisdictions, the status of every family member against its national filing requirements, newly published art that reads on a pending claim set, and competitor filing activity in the same technology class. Tools that write back to the docketing system, rather than sitting beside it, avoid a second record of the same deadline.

What to Confirm Before Relying on the Alerts

1. How current the jurisdictional filing-requirement data is. National office rules change, and a stale data feed produces a confident but wrong deadline.
2. Whether annuity deadlines are confirmed against the national office of record, rather than estimated from a general formula that does not account for a specific jurisdiction’s grace periods.
3. Whether a missed or ambiguous status triggers a human review, rather than a silent skip that only surfaces once a deadline has passed.
4. Whether the tool distinguishes an abandoned family member from an active one in its alerts, rather than flagging a dead filing as though it still carries a live deadline.

A tool that surfaces competitor filing activity is a monitoring aid for strategy, not a substitute for a formal freedom-to-operate opinion when a launch decision is on the line.

Patent Protection for AI-Generated Inventions

AI-generated inventions can be patented when a natural person significantly contributed to the invention, and the USPTO’s Inventorship Guidance for AI-Assisted Inventions, published at 89 FR 10043 on February 13, 2024, states that such inventions are not categorically unpatentable. Significance is measured by the Pannu factors, and a failure on any one of the three precludes naming that person as an inventor. Merely recognizing a problem or setting a general research goal does not rise to the level of conception, though the guidance accepts that constructing a prompt in view of a specific problem, to elicit a particular solution, can be a significant contribution. And owning or overseeing the system, what the case law calls intellectual domination, does not make anyone an inventor.

The Pannu Factors, Applied to AI-Assisted Work

1. Contribute in a significant manner to the conception or reduction to practice of the invention.
2. Make a contribution to the claimed invention that is not insignificant in quality, measured against the dimension of the full invention.
3. Do more than merely explain to the real inventors well-known concepts or the current state of the art.

Ownership disputes often start in the employment context, where AI use in employment practice complicates who contributed what, and an employee’s AI-assisted contribution still has to clear these three factors before an assignment clause has anything to transfer.

Patent Eligibility and Obviousness for AI Inventions

Patent eligibility and obviousness for AI inventions turn on 35 U.S.C. 101 and 103 rather than on the inventorship question, and each statute raises a different risk for an AI-assisted application.

Two Statutes, Two Different Questions

1. Section 101, eligibility. Software-implemented AI claims commonly draw an abstract-idea rejection unless the claims recite a specific technical improvement rather than a result.
2. Section 103, obviousness. The person of ordinary skill is assumed to have the ordinary tools of the field, so as AI search and generation become ordinary, the pool of what a skilled artisan would have found obvious is likely to widen.

Neither doctrine has settled for AI, and the USPTO treats both through its 2024 guidance update on subject matter eligibility rather than a separate AI track.

[cta-2]

Copyright Protection for AI-Generated Works

Copyright protection for AI-generated works requires a human author, a rule that traces to the text of the Copyright Act itself rather than to any AI-specific statute. The D.C. Circuit held in Thaler v. Perlmutter, decided March 18, 2025, that the best reading of the Copyright Act of 1976 is that human authorship is required for registration, rejecting the argument that a computer system could itself be named author of record. The full opinion lays out why the statute’s use of terms like “widow,” “children,” and “domicile” presumes a human actor throughout. The Supreme Court denied certiorari on March 2, 2026, leaving the D.C. Circuit’s holding undisturbed. Because refusals to register are reviewed in the D.C. Circuit, the human-authorship requirement governs registration practice in effect, although the denial of certiorari is not itself a ruling on the merits. The rule is narrower than it sounds: it bars registering the machine as the author, not registering a work made with the assistance of AI.

What Carries a Copyright, and What Does Not

1. Protectable. Where a human author selects, arranges, or materially modifies machine output, those human contributions are copyrightable in the resulting work.
2. Not protectable. The untouched machine output itself does not carry a copyright, no matter how much post-generation selection surrounds it elsewhere in the same project.

The line matters most in licensing. A content team that generates a first-draft product description with AI and then substantially rewrites and arranges the result may hold a copyright in the human-authored elements of that rewritten version; the unedited first draft does not carry one, and a license that purports to grant exclusive rights in that raw output is granting rights that do not exist. Licensing counsel reviewing a content-supply agreement should confirm which version of a deliverable, the AI draft or the human-edited final, is the one actually changing hands under the grant.

Challenges in Copyright Law Regarding AI-Generated Content

Several open challenges in copyright law regarding AI-generated content remain unresolved across jurisdictions:

1. Training data and fair use. Whether ingesting copyrighted works to train a model is a fair use is being litigated across multiple circuits with inconsistent early results.
2. Ownership of hybrid works. The line between AI-assisted human authorship, which can be copyrighted, and autonomous AI output, which cannot, is fact-specific and turns on how much the human contributor actually selected or arranged.
3. International divergence. Some jurisdictions outside the United States extend limited protection to AI-generated works under different authorship theories, which complicates a multinational portfolio.
4. Attribution and moral rights. Even where a work is protectable, questions about crediting the tool or disclosing AI involvement remain unsettled in most registration regimes.

The Copyright Office’s Part 2 report on copyrightability is the current primary source for where the agency’s own thinking stands on each of these.

Confidentiality and Supervision Duties When Using AI in IP Work

AI confidentiality duties for patent attorneys start with a fact specific to IP practice: a pending application is confidential by statute under 35 U.S.C. 122, so pasting draft claims into a tool that trains on its inputs can put both the filing and the client’s trade secrets at risk. The duty runs alongside the ethics rules: competence in the specific tool’s limits, supervision policies under Model Rules 5.1 and 5.3, and honest billing for AI-assisted work. The USPTO issued parallel guidance in April 2024 for practitioners appearing before the Office, holding them to the same duties of candor and verification that apply to every other filing.

The ethics rules a practitioner owes a client reach further than that agency-specific notice, and they do not relax just because the tool is new. ABA Formal Opinion 512, issued July 29, 2024, applies the existing Model Rules to generative AI and asks for informed client consent before confidential information enters a self-learning tool. It is the operative standard behind every check below.

What to Check Before a Tool Touches a Pending Application

1. Data retention and training terms. Whether the vendor retains inputs, and whether those inputs train a model shared across customers.
2. Who at the vendor can see inputs. Internal access controls on a self-learning tool matter as much as the training policy itself.
3. Whether outputs are logged for supervision. ABA Formal Opinion 512 applies the Model Rule 5.3 duty to supervise nonlawyer assistance to a firm’s use of generative AI, and that supervision needs a record a supervising attorney can actually review.
4. The written firm policy Model Rule 5.1 expects. A partner-level policy on which tools are approved for which matter types, not an informal norm.

Where AI Falls Short in IP Practice

The limitations of AI in intellectual property law show up in fabricated citations, loss of precision, claim scope drift, and questions that are legal rather than technical.

Where the Failures Show Up

1. Fabricated citations. In IP work this means an invented reference in an office action response or a non-existent case in a validity opinion.
2. Precision loss. Semantic search trades precision for recall, so a result set needs triage rather than acceptance.
3. Claim scope drift. Claim language is adversarial and exact, and a fluent AI draft can narrow or widen scope in ways only a reading for infringement and validity will surface.
4. Legal judgment questions. The inventorship and authorship questions are legal rather than technical: no tool can tell you whether a natural person contributed enough to be named.

What Reduces the Risk

1. Citation fabrication. A verification pass against the cited source before the document leaves the building.
2. Precision loss. A second human filter after the initial semantic retrieval, narrowing results before anyone reads for relevance.
3. Scope drift. A claim-by-claim comparison against the prior draft or the client’s instructions before the application goes out.

The fabricated-citation failure is exactly what AI legal research tools are built to catch before a filing goes out, by checking a generated citation against the actual case or reference before it reaches an office action response. None of this is unique to IP work, but the stakes are higher than in general commercial drafting: a fabricated citation in an office action response or a scope error in an issued claim is not easily corrected after the fact, which is why the human read stays mandatory at every stage listed above, not just the ones where a rule requires it.

Categories of AI IP Tools, and Where Each Fits

AI IP tools fall into four categories, and a practice often draws on more than one of them. The USPTO itself now runs AI in two of these categories, an AI similarity search for patent examiners and the Class ACT classification assistant on the trademark side.

What Each Category Does, and Does Not Do

1. Patent search and analytics platforms, such as Derwent Patent Search (formerly Derwent Innovation), LexisNexis TotalPatent One, and PatSnap’s analytics tools, widen semantic recall across global patent literature, but none of them renders a novelty or patentability opinion; that judgment still sits with counsel.
2. Patent drafting and prosecution assistants turn an invention disclosure into a first-pass specification and claim set, yet every claim still needs a human read for scope, enablement, and written-description sufficiency before it reaches the examiner.
3. Trademark clearance and watch services, including Corsearch’s TrademarkNow platform, screen phonetic, semantic, and visual similarity across registers, but a clearance opinion is a likelihood-of-confusion judgment that no screening tool resolves on its own.
4. Contract review tools for the licensing and assignment layer surface risk in draft license and assignment language without deciding how ownership, indemnity, and field-of-use terms should actually be negotiated.

Matching a specific product to a practice’s docket size and jurisdiction mix is its own exercise, and the AI tools IP lawyers use divide along these same four categories.

Where Spellbook Fits in an IP Practice

Spellbook is an AI platform for the commercial layer of IP work: the licensing, assignment, and development agreements that move rights around once they exist, not the search, drafting, or docketing of the IP itself. The Review feature reads an incoming license or assignment agreement against a firm’s playbook and proposes redlines as native Word tracked changes, surfacing the ownership, indemnity, and field-of-use terms a licensing attorney checks first, which is where the intellectual property clauses in a license carry the most risk. The Draft feature produces a first pass of standard IP-contract clauses grounded in the firm’s own precedent language and Spellbook’s 20M+ real contracts, rather than generic boilerplate. Spellbook applies zero data retention to every input, which matters specifically for IP work: a draft license referencing an unpublished invention or a pending application is confidential, and the tool’s data handling has to match that from the first paste.

What Spellbook Does Not Do

1. Run prior art searches or produce a novelty opinion.
2. Screen or clear trademarks.
3. Track docket deadlines or patent family status.

[cta-3]

Frequently Asked Questions

Can AI file a patent?

No. A patent application must be filed by or on behalf of a human applicant, and USPTO rules require a natural person to be named as inventor on every application, with the significant-contribution test from the 2024 inventorship guidance applied to any AI-assisted claim. AI tools can help draft specifications and claim language, but the filing, the signature, and the inventorship declaration are human acts.

Can AI help with trademark registration?

Yes, for the search and classification steps. AI-assisted clearance screening checks phonetic, semantic, and visual similarity before filing, and classification tools can assign international classes to a new application. The registration itself, including the specimen of use and the legal judgment on likelihood of confusion, still runs through a human filer and, for anyone but the applicant itself, a licensed practitioner.

Can you trademark something created by AI?

A trademark protects a brand identifier used in commerce, not the creative process behind it, so a logo or wordmark generated with AI assistance can be registered as a trademark as long as it functions as a source identifier and clears examination for likelihood of confusion. That is a different question from whether the same image could be copyrighted, where AI-generated output without sufficient human authorship cannot.

Is AI patent drafting allowed before the USPTO?

Yes, with conditions. The USPTO’s April 2024 guidance on AI-based tools in practice before the Office permits their use but holds the practitioner to existing duties of candor, competence, and verification: every AI-assisted claim, specification, and citation still has to be checked before filing, and the practitioner remains responsible for the submission regardless of which tool produced the first draft.

Does using AI on a pending application waive confidentiality?

It can, depending on the tool. A pending application is confidential by statute under 35 U.S.C. 122, and pasting draft claims into a tool that retains or trains on inputs can expose that confidential information outside the attorney-client relationship, which is exactly what ABA Formal Opinion 512 asks practitioners to check before use. The fix is a tool with contractual data retention terms that match the confidentiality duty, confirmed before a pending application ever touches it, not a blanket avoidance of AI tools.

Try Spellbook for IP Licensing and Assignment Work

Try Spellbook free to see how the Review and Draft features handle an IP licensing or assignment agreement against your firm’s own playbook.