The headline for AI inventorship is simple and consequential: in both the United States and Europe, only natural persons qualify as inventors. AI systems, whether foundation models or bespoke engines, are tools, not inventors. For AI developers that means mapping each claim element to a human who conceived it, documenting how the system was used, and framing claims around concrete technical improvements. Done properly, you get durable rights without priority traps or inventorship challenges.
Where the US and the EPO stand
Despite the high-profile DABUS test cases, the core rule has not moved: name humans as inventors. US law limits inventorship to natural persons, and the European Patent Office refused the DABUS applications on the same basis and has stayed there since. Eligibility for AI-related claims in the US runs on the same software law you already know, the two-step analysis in MPEP § 2106. None of this is AI-specific inventorship law. AI is treated as a tool.
| Jurisdiction | Can AI be named as inventor? | What to do |
|---|---|---|
| United States | No. Only natural persons may be inventors; AI remains a non-inventor tool. | Identify and name the human conceivers for each claim, emphasize the technical improvement under § 101, and keep AI names out of your priority chain. |
| Europe (EPO) | No. The EPO refused the DABUS applications on the ground that an inventor must be a natural person. | Name the human conceivers and proceed under ordinary EPO practice. |
Why it matters: misnaming an inventor can put validity in question, and in the US, claiming priority from a foreign or PCT filing that listed an AI as inventor creates avoidable formalities work unless it is corrected. Align your naming across jurisdictions from day one.
United States: naming inventors and eligibility
Inventorship in the US turns on conception: who formed a definite and permanent idea of the claimed invention. A human who conceived the claimed subject matter is the inventor, even where an AI system assisted, and the AI itself cannot be named.
USPTO guidance on AI-assisted inventions has been issued and revised more than once, so check the current version rather than working from a summary. The stable part is the underlying doctrine: traditional conception standards apply, and there is no separate AI test. Federal Circuit precedent limits inventorship to natural persons and requires a contribution to the conception of the claimed subject matter. That case law developed outside the AI context, but the principle carries straight over. Cite the human contributions to conception and treat the AI as an enabling tool, no different in kind from a lab instrument or a compiler.
On eligibility, AI-related inventions are examined like any other software claim. Applicants who show the how and the why of an improved computer performance do better than applicants who claim the outcome.
Practical steps for US filings:
- Map each claim to at least one human conceiver, and keep a contemporaneous record connecting that person's contribution to the claim language.
- Describe the AI's role as a tool: what inputs the humans supplied, how parameters, datasets, prompts or training choices were selected, and how those choices led to the claimed result.
- For § 101, draft claims that implement a specific architecture, pipeline or control mechanism that measurably improves system performance. Explain the mechanism in the specification and use those hooks in prosecution. Our post on software patents, Alice and § 101 works through the drafting in detail.
Europe after DABUS
The EPO position has not shifted since the DABUS matters: only natural persons can be named as inventors, and a filing that lists an AI as inventor is not accepted. Name the human conceivers and proceed under ordinary EPO practice. Align inventorship across your global filings, because inconsistencies complicate priority and formalities.
One practical cross-border point: if an early foreign filing named an AI, for example to make a point in a test case, plan a clean US and EPO path with human inventors before you claim priority from it. Do not rely on AI-as-inventor anywhere you actually intend to hold protection.
A compliance roadmap for AI patent filings
For in-house counsel and founders steering AI research, the fastest way to de-risk is to systematize inventorship and eligibility from the start.
1) Identify the human conceivers, claim by claim.
- Before drafting, list each claim element and the person or people who conceived it. Keep dated notes, design documents or tickets linking the idea to a person.
- Capture why that person's contribution mattered: a new training regimen, a novel control loop, a data-path optimization.
2) Treat the AI as a tool, and document it.
- Record prompts, hyperparameters and dataset choices, and note who chose them and why.
- Store versioned artifacts: model cards, evaluation reports and ablation studies that show the intentional human design choices.
3) Frame claims on concrete technical improvements.
- Articulate measurable benefits such as reduced inference latency, lower memory use, better throughput or improved robustness, and tie each to a claimed mechanism.
4) Scrub priority across jurisdictions.
- Avoid naming an AI in any filing that could sit in your priority chain. If a foreign or PCT filing already did, plan corrections or fresh filings with counsel before US or EPO prosecution.
5) Prosecute under the standard rules.
- Argue eligibility in the US under MPEP § 2106, focused on the technical improvement, and reference the current USPTO inventorship guidance where the examiner raises it.
6) Monitor the eligibility case law.
- Track PTAB and Federal Circuit outcomes touching AI. The direction of travel matters for how much architectural detail you put in the specification.
Drafting strong AI claims: show the how, not just the result
Most eligibility misfires in AI practice come from abstract, result-oriented claiming. A few drafting habits prevent them:
- Specify the architecture. Name the layers, modules or control elements and how they interact, for example a scheduler that gates token generation based on a learned latency model, rather than claiming "optimized inference".
- Tie the claim to a concrete improvement. "Achieving sub-50ms median latency on edge hardware by pruning and quantization coordinated through a feedback controller", not "a faster model".
- Disclose the mechanism in the specification. Include ablations and benchmarks that isolate the claimed technique as the cause of the improvement.
- Anchor the improvement to computer functionality: fewer cache misses, fewer memory swaps, less network chatter. Those are effects examiners recognize as technical.
A useful claim pattern:
- Independent claim: a system or process implementing your core control or data-path innovation.
- Dependent claims: concrete variants, such as parameter ranges, cache layouts, quantization steps, scheduler policies or data-fusion rules.
- Method and computer-readable medium claims: mirror the system with steps and storage, to cover several modes of infringement.
Back it with evidence. Data in the specification showing the improvement, for example a percentage reduction in latency on named hardware from a named mechanism, gives the examiner a clear hook. Even if the numbers move later, the mechanism-first story holds.
Common pitfalls
- Naming an AI as an inventor anywhere in your priority chain, which creates formalities problems and can undermine the US strategy.
- Mixing tests from superseded guidance with joint-inventorship factors that were never designed for AI-assisted work. Work from the current guidance and traditional conception standards.
- Drafting abstract, result-oriented claims that never show how the improvement is achieved.
- Failing to document human conception and the AI's tool-like role. If you cannot connect a person to each claim, your inventorship story is exposed.
- Overlooking cross-border filing risks that complicate US priority. Clean up the naming before you file in the US or at the EPO.
Fees, timing and portfolio planning
There is no AI-specific fee schedule. Standard USPTO fees apply, and they vary with entity status, claim count and route, so work from the current official fee schedule and your counsel's quote rather than from an estimate in an article. Our patent filing service sets out what the drafting and prosecution work costs.
Expect a crowded field. AI-related filings have grown quickly across technology subclasses, which is an argument for filing early where you have a defensible improvement rather than waiting for the product to settle.
Portfolio planning tips:
- Align inventorship and eligibility reviews with sprint demos. Treat every major architecture or pipeline change as a potential disclosure event.
- File provisionals to capture fast iterations, then consolidate into non-provisionals once the data stabilizes.
- Coordinate with your privacy and data-licensing leads, so the training-data story is ready when diligence asks for it.
The bottom line
In the US and in Europe, AI systems are not inventors. For every AI-assisted filing, your job is to name the humans who conceived each claim element, document the AI as a tool, and draft claims that concretely improve computer performance, with the mechanism explained in the specification. Operationalize those habits and you cut priority risk while improving your odds of allowance.
Sources
Need help with your trademark?
Get a free trademark check from our specialists, no obligation.
