What Are the EPO’s Actual Controls for AI-Assisted Patent Drafting?
The EPO does not maintain a separate approval system for drafting with ChatGPT, Claude, Gemini, or another generative-AI tool. Instead, its controls arise from the European Patent Convention, the Implementing Regulations, the EPC Guidelines, and ordinary prosecution standards, including inventorship, inventive step, industrial applicability, sufficiency, clarity, and entitlement. Generative AI can therefore be used to propose claim language, restructure a specification, summarize prior art, or identify ambiguities, but a named human applicant must still own the application and satisfy the legal requirements for an invention. The central question is not whether a word was typed by a machine; it is whether the application identifies a human inventive contribution, discloses the invention adequately, and presents patentable subject matter. For patent teams, that means AI use belongs within a documented, supervised drafting process rather than an unsupervised production line. The following is the practical position as of 25 September 2026, subject to confirmation against the EPO’s current Guidelines and fee schedule.
Also worth reading: How Does the EPO AI Drafting Workflow Change Patent Practice in 2026? · How Is AI Changing EPO Patent Drafting in 2026, and Where Does It Still Fall Short? · What Are the Key Performance Indicators for AI-Assisted Patent Drafting in 2026?
The distinction between AI-assisted drafting and AI-invented subject matter matters. If an engineer develops a proposed invention and uses AI to prepare a first draft for counsel to verify and revise, the engineering work may supply the human inventive contribution. If counsel simply asks a general-purpose model to invent a solution and files the result without a technically supported, human-developed concept, problems arise before examination reaches the claims. The EPO has not adopted a blanket rule that all AI-assisted applications are ineligible. Its examination practice instead asks what the natural-person inventors actually contributed and whether the claimed features solve a technical problem under the applicable rules. This approach resembles, but is not identical to, patent-office positions elsewhere, so international filing teams should not apply one inventorship answer universally.
Which Filing Requirements Survive When AI Participates in Drafting?
The EPC’s disclosure requirements do not disappear because most of the prose was generated quickly. The application must enable a skilled person to carry out the invention across the claimed scope, and the claims must be clear enough to define the requested protection. AI can introduce unsupported assertions, invented citations, contradictory terminology, or features that appear in the claims but not in the description. Consequently, every technical statement, dimension, process step, parameter range, and cited document needs verification against an inventor record, test data, laboratory notebook, or authoritative source. Patent offices receive machine-readable filings capable of transmitting extensive text, but capacity is not accuracy. A large, polished specification produced in minutes may still fail because its technical foundations are wrong.
A second requirement concerns the relationship between the filed text and the valid filing date. Later optimization, experimental data, and added disclosure cannot always be backdated to the original priority date. Teams using AI iteratively should preserve dated versions and distinguish between ordinary editing of a disclosed embodiment and the addition of new matter. The EPO generally evaluates whether the original application contained the feature needed to obtain the priority date for an intermediate generalization. AI-generated suggestions are not inherently new matter merely because they were generated before a later filing, but newly added technical material without an earlier basis can be. This is particularly important when AI inserts a range, a preferred embodiment, or a technical effect that the inventors had not disclosed and could not reasonably derive from the original disclosure.
Inventorship must also be allocated to the natural persons who contributed to the inventive concept. Merely approving a draft does not automatically make every reviewer an inventor, and naming people only because they are project participants is equally unsafe. The EPO requires applicants for European patents to be entitled to the patent and treats the inventors as natural persons. An AI system is not named as a co-inventor under this regime, and the DABUS dispute illustrates that a system described as acting autonomously is not itself an EPC inventor. AI-generated proposals should therefore be treated as source material for human evaluation, not as a person from whom legal rights can be claimed.
Why Can Weak AI-Assisted Drafting Surface Years Later?
The central risk is deferred verification. A claim may read professionally while quietly narrowing the invention, broadening a functional phrase, or combining incompatible elements. That error may not become obvious during filing, and it may become damaging only when an infringer requests a declaration of non-infringement, validity is challenged, or the patent is asserted in a transaction. In the United States, an inventor statement is subject to a reasonable belief standard, and later decisions or successful prosecution can expose statements that were inaccurate when made. In Europe, a written declaration or other inventor-related statement is not made irrelevant by good-faith reliance on an AI-generated summary. Later litigation can test whether the underlying facts, not merely the final wording, were correct.
Generative models are particularly vulnerable to citation errors. A model can fabricate a patent number, publication, passage, holding, scientific reference, or technical standard. Even when the cited document exists, the model may misstate what it teaches or reverse the relationship between a feature and an effect. This failure is especially consequential during prior-art analysis because a missed relevant document can weaken the case for inventive step, while an incorrect citation can damage the applicant’s credibility with an examiner. Automated search and drafting tools are useful for generating queries and candidate documents, but an appropriately qualified reviewer should confirm the publication identity, publication date, relevant passages, family position, and legal status before the application relies on it.
AI can also flatten a technical distinction that was valuable during prosecution. Claim sets often define a narrow relationship between components, data types, operating conditions, and effects. A model may replace that relationship with a broad functional statement because the latter appears conventional and easier to generate. Later, a competitor may adopt the exact architecture that supplied the patentability, leaving the patentee asserting a claim whose words are broad but whose technical teaching is not. The remedy is not to reject AI editing, but to require the human drafter to understand why each limitation belongs in the claim and whether the description supports its full scope.
What Human Controls Should Patent Teams Put in Place?
A workable control process begins with role-based access. Inventors provide the problem, essential technical features, alternatives considered, experiments, and unresolved uncertainties. A technically competent human drafter decides how the disclosure supports priority, inventive step, and the requested claim scope. Counsel reviews legal characterization, inventorship, entitlement, unity, and filing strategy. Automated tools may propose language or search results, but they should not independently declare an application complete. For higher-risk matters, a second reviewer should examine the written description independently of the generated draft so that verification is not limited to checking whether the model formatted the request correctly.
The team should retain an AI-use record that states which tool and model version were used, which tasks were assigned, who reviewed the output, what sources were checked, and which changes were made. This record need not reproduce every prompt in the public application, but it should allow a later professional to reconstruct the drafting decisions. Dates matter because model behavior and legal standards can change, and a record created months later may not establish what actually happened. Confidentiality is a separate control: unpublished patent applications often enter contractual, prior-use, or commercial negotiations, so enterprise data-retention settings should be checked before proprietary invention details are entered into a public or shared model.
Claims require a feature-to-effect check. For each important limitation, the reviewer should identify its source in the disclosure, confirm that it is operationally meaningful, and consider whether skilled people could implement it without inventive work. A similar check should test the description for enabling passages across the broadest surviving scope. Where AI proposes numerical ranges, the team should ask whether the endpoints come from actual testing, a mechanical or scientific principle, or ordinary knowledge, and whether the specification teaches how to work across the range. Fluent explanations are not substitutes for evidence.
How Do EPO, UK, and US Approaches Compare?
No single global “AI patent drafting standard” exists. The comparison below is a practical orientation, not a substitute for jurisdiction-specific legal advice, and the United States position is only loosely comparable to the EPC because US law recognizes an “individual” as a possible inventor, while the European system concerns natural persons.
| Feature | European Patent Office | United Kingdom | United States |
|---|---|---|---|
| Can generative AI be used to assist drafting? | No categorical prohibition is stated; the filed application must meet the EPC and procedural requirements | Use is not itself a general disqualification; drafting must support the claimed human invention and the required statements | Use is not itself a general disqualification; USPTO guidance applies to AI-assisted inventions |
| Who may be an inventor? | Natural persons contributing to the European invention | Natural persons | A “natural person” or, in principle, another “individual” under US law |
| How is inventorship assessed? | The human inventive contribution behind the claimed solution | Actual contribution to the invention, not merely a formal job title | Significant contribution to at least one claim, assessed by more than conception alone |
| Main drafting exposure | Missing human inventive contribution, insufficient disclosure, lack of clarity, lack of inventive step, and incorrect inventorship | Inventorship errors, insufficient disclosure, and failure to disclose relevant inventorship material | Inventorship errors, duty of candour, disclosure duties, enablement, and written-invention-statement risk |
| Practical response | Keep human invention development and claim decisions traceable to qualified personnel | Record human contribution and disclose the correct inventors | Document conception and inventorship claim by claim; correct errors through applicable procedures |
Which Practical Workflow Applies from Disclosure to Filing?
The first stage is a controlled invention conference involving the people who actually developed the solution. They should describe the starting problem, the non-obvious design decision, failed approaches, alternatives, and any test results. The output should be a human-authored invention record, with AI limited to transcription support if the record is checked. This is preferable to asking a model to reconstruct the invention from a short instruction, because the model may fill gaps with plausible but incorrect technical detail. A project manager’s description may not be detailed enough to determine inventorship, particularly where a system combines features developed by different engineers.
The second stage is source-grounded drafting. The drafter can use AI to compare claim versions, generate headings, flag repeated terms, or propose questions for inventors. Every material proposal should be tagged for verification against the invention record. Prior-art searching should use recognized patent databases and specialist tools, followed by review of the actual published documents. The same evidentiary standard applies to scientific literature, standards, and public disclosures: the document must exist, the quoted proposition must be present, and its date must be relevant to the assessment.
The third stage is a pre-filing review. The application should be read for new matter, unsupported absolutes, missing dependencies, unclear antecedents, inconsistent terminology, and claims that extend beyond the worked examples. A separate reviewer should confirm inventorship and the deadline for any required information. Only after those checks should the file be authorized for submission through the EPO system, including confirmation of the applicant’s entitlement and payment of official fees. The internal process does not guarantee grant or validity, but it reduces avoidable defects and makes later advice better supported.
What Does AI-Controlled Drafting Cost?
The EPO’s official fees concern the application process, not the commercial price of drafting. As a working planning range, filing a European patent application commonly costs several thousand euros before translation, attorney fees, prior-art work, and responses during prosecution. A tightly scoped technical matter may be quoted in the low thousands, while a complex portfolio instruction with extensive searching, validation in multiple countries, and negotiation work can run into tens of thousands. Drafting offices may charge a modest premium for human-reviewed AI assistance, but the difference between packages is less important than the verification, qualification, and responsibility included.
Some automated tools are available through low-cost subscriptions, while enterprise arrangements with private models, audit logs, access controls, and security review are priced separately. A pilot using a company’s existing tool may require little more than staff time; a controlled project requires model evaluation, confidentiality review, training, validation against a historical filing set, and ongoing monitoring. That measurement matters because an incorrect application can cost more than the original drafting fee. Budgets should therefore track review hours and rework, not merely the number of generated applications.
Cost pressure is a poor reason to remove human checks. Comparing an unchecked AI draft with a reviewed professional service is like comparing a proof-of-concept with a filing that has been tested for legal and technical defects. Patent counsel should ask the provider who is professionally accountable, whether the data is retained or used for model training, what version generated the text, and whether prior artifacts are preserved. A stated percentage saved on drafting time is not a percentage saved on prosecution, litigation, or invalidation risk. The EPO continues to evaluate applications on the merits of the disclosure and the human contribution, so production speed does not alter the examination task.
When Should a Team Act, and What Are the Common Mistakes?
A team should establish controls before the next filing because the first AI-assisted application sets a precedent for later work. The immediate need is greatest where applications cover commercially important software, medical technology, robotics, machine learning systems, or inventions created by distributed engineering teams. Those fields often involve dense terminology, overlapping prior art, and inventorship questions that a general drafting checklist cannot resolve. A controlled pilot of approximately 10 to 20 recent applications can reveal whether the tool introduces new errors, but the sample should include difficult inventions rather than only simple claim-rewrite exercises. Teams should then document the measurable error rate, reviewer time, and rework rate before expanding use.
The most common mistake is treating fluency as evidence. The second is asking an AI system to invent the solution, then searching for human inventors who can support whatever it generated. A third mistake is assuming that because the model cites a patent, the citation is safe. A fourth is allowing the model to expand preferred embodiments or claim language without a priority-date check. Others include using one inventorship answer across the EPO, United Kingdom, and United States; failing to preserve prompts and revisions; and placing confidential material into a service with unsuitable retention or training terms. These mistakes are preventable, but only if the firm assigns responsibility rather than describing AI use merely as a software preference.
The appropriate conclusion is neither prohibition nor unrestricted automation. Use AI where it produces measurable drafting efficiency, but retain human control over conception, technical truth, claim scope, inventorship, and the final filing. Re-evaluate the policy when the model changes, when an application is challenged, or when the EPO updates its Guidelines. The reliable control is evidence: a dated human record, verified sources, supported disclosure, and a claim set that a qualified person can defend years after the application was filed.