AI tools for patent generation are marketed most heavily for the task they do worst, and undersold for the ones they do well.
The task they do worst is drafting claims. A claim is a legal instrument whose value depends on what an examiner will allow and what a competitor can design around. Both are predictions about future human behaviour.
The tasks they do well are all volume tasks — searching, classifying, charting, triaging, valuing. These are cases where processing ten thousand documents beats reading fifty, and that is a real advantage.
Where AI genuinely helps
| Task | Why AI helps | Human still needed for |
|---|---|---|
| Prior art search | Semantic matching finds documents using different vocabulary | Assessing what the references actually teach |
| Classification | CPC assignment across large sets | Edge cases and new technologies |
| Claim charting | First-pass mapping of elements to features | Verifying every cited source |
| Portfolio triage | Scoring hundreds of patents consistently | The keep/sell/lapse decision |
| Valuation modelling | Comparables, citations, term, classification | Judging whether the model fits |
| Translation | Reading foreign prior art at speed | Legal significance of specific wording |
| Monitoring | Watching new filings continuously | Deciding what matters |
Semantic search is the strongest single use. Prior art is written in the vocabulary of its own field, so a search for "flow stabiliser" misses a "damping baffle" describing the same thing. Semantic matching crosses that gap in a way keyword search cannot.
Charting is the biggest time saving. Building a claim chart by hand takes hours per patent; a reviewed first draft takes a fraction of that. See patent claim chart.
Triage at portfolio scale is where the economics are clearest. Scoring two hundred patents consistently on term, classification and citation profile is a task nobody does well by hand.
Why drafting resists automation
Every other task in patent work compares documents that exist. Drafting creates one that does not.
| Task | Nature |
|---|---|
| Search | Retrieval against a corpus |
| Classification | Scoring against a taxonomy |
| Charting | Matching text to text |
| Triage | Scoring against known attributes |
| Drafting a claim | Predicting two future human responses |
A claim has to survive an examiner and resist a competitor, and both are adversaries who have not acted yet. Writing one well means anticipating which limitation an examiner will insist on, and which word a competitor's engineer will work around.
That is why AI drafts read fluently and claim badly. Fluency is a property of the text; claim quality is a property of how other people respond to it years later.
The categories of tool
"AI patent tool" covers several distinct products with different maturity.
| Category | What it does | Maturity |
|---|---|---|
| Semantic search | Finds prior art by meaning, not keywords | Strong |
| Classification | Assigns CPC codes | Strong |
| Claim charting | Maps elements to product features | Good, needs verification |
| Portfolio analytics | Scores and triages at scale | Good |
| Valuation | Estimates value ranges from comparables | Varies wildly by dataset |
| Monitoring | Watches new filings and litigation | Good |
| Drafting | Generates specifications and claims | Weakest |
| Office action response | Suggests arguments and amendments | Emerging |
Valuation quality depends entirely on the underlying data, and this is where claims should be read carefully. A tool trained on modelled value estimates is not the same as one with transaction records, and the two are frequently described in the same language.
Ask what the data actually is. Stock-market-derived value estimates, court awards, and recorded sale prices are three different things, and only the last is a transaction.
Drafting is last on maturity for a structural reason. Everything above it is retrieval or scoring against existing documents. Drafting requires predicting how an examiner and a competitor will respond to text that does not exist yet.
Where they fail
| Failure | What it looks like |
|---|---|
| Fabricated citations | Plausible patent numbers that do not exist, or exist and say something else |
| False clean search | A confident "no relevant art found" that missed the key reference |
| Result claims | Claims reciting what is achieved rather than how — fatal under §101 |
| Missing alternatives | A specification covering one embodiment, supporting only narrow claims |
| Confidentiality exposure | Unfiled invention details entering a tool that retains inputs |
| Stale data | Recent filings absent from the index |
The result-claim failure is the most consequential and least visible. A draft claiming "a system that predicts availability" reads fluently and fails subject-matter eligibility, and it takes someone who knows the Alice framework to see why. See what can be patented.
A confident negative is more dangerous than a confident positive. A fabricated citation gets caught when someone opens it. A missed reference gets caught by an examiner two years later, or by a challenger eight years later.
Verify every citation before relying on it. Open the document, read the passage cited, confirm it says what the tool claims. This is not optional and it is where most of the residual human time should go.
Worked example: a search compared
A mechanical damping mechanism. Same invention, three approaches.
| Keyword search | AI semantic search | Professional search | |
|---|---|---|---|
| Time | 3 hours | 20 minutes | 2 weeks |
| Cost | Free | Low | $1,500–$3,000 |
| Patents surfaced | 40 | 180 | 95 |
| Relevant after review | 6 | 11 | 14 |
| Key reference found | No | Yes | Yes |
| Non-patent literature | Poor | Moderate | Strong |
| Foreign-language art | None | Yes, translated | Yes |
The AI search found the key reference the keyword search missed, because that reference described the same mechanism as a "compliant restraint" rather than a damper. No amount of keyword iteration would have surfaced it without guessing that phrase.
It also returned 180 documents to triage. Recall improved and precision did not, so someone still had to read.
The professional search found three more relevant references, all in non-patent literature — a conference paper and two trade publications. That remains the weakest area for automated search.
The sensible sequence: AI search first because it is fast and cheap and eliminates most inventions, then a professional search before committing to a non-provisional. See prior art.
Inventorship is unchanged
US law requires inventors to be natural persons. Thaler v. Vidal settled that an AI system cannot be named as an inventor.
| Scenario | Inventorship position |
|---|---|
| Human conceives; AI helps search | Human is the inventor |
| Human conceives; AI drafts the text | Human is the inventor |
| Human prompts; AI produces the concept | Unsettled and risky |
| AI alone | No valid inventorship |
The middle case is the genuinely uncertain one. Conception is the definite and permanent idea of the complete invention. Someone who described a problem and received a solution may not have conceived it in that sense.
Document the human contribution contemporaneously. Who identified the problem, who specified the constraints, who recognised the solution as workable, who selected among alternatives. Reconstructing this years later in a deposition is much harder. See inventorship.
Disclosure duties do not change
Under 37 CFR 1.56, everyone substantively involved must disclose known material prior art, and the duty runs until the patent issues.
| Situation | Duty |
|---|---|
| AI search surfaces relevant references | Submit them |
| You did not read them closely | Still submit |
| The tool ranked them low | Still submit if material |
| References found after filing | Still submit |
"The tool found it and I did not look" is not a defence. Failure to disclose material art can render a patent unenforceable for inequitable conduct, which is a far worse outcome than a narrower claim.
Which argues for erring toward submission. An Information Disclosure Statement listing more references costs little; an unenforceable patent costs everything.
Confidentiality before filing
Putting an unfiled invention into a tool that retains inputs is a disclosure risk.
| What you did | US position | Rest of world |
|---|---|---|
| Tool with contractual confidentiality | Not a public disclosure | Not public |
| Public tool retaining and training on inputs | Arguably risky | Risk |
| Anything genuinely published | 12-month grace period running | Rights likely lost |
The US grace period gives twelve months from your own public disclosure. Most other countries give nothing. The asymmetry means a disclosure that leaves US rights intact can destroy European and Japanese rights permanently.
Check the tool's data handling before entering anything unfiled, and prefer tools offering contractual confidentiality for pre-filing work.
What this changes economically
Costs fall where volume dominates and hold where judgement dominates.
| Task | Direction |
|---|---|
| Prior art search | Down sharply |
| Claim charting | Down sharply |
| Portfolio triage and valuation | Down sharply |
| Translation and monitoring | Down |
| Claim drafting | Roughly flat |
| Prosecution strategy | Flat |
| Opinions and adversarial work | Flat |
The practical effect is more decisions made with evidence. Searching used to cost enough that many inventors skipped it; charting cost enough that most portfolios were never mapped against products. Both are now cheap enough to do routinely.
Which matters most at the renewal decision.
| Ipiry Patent Survival Curve v1.0 | Rate |
|---|---|
| Survive the 3.5-year fee (2022 cohort) | 85.8% |
| Survive the 7.5-year fee (2018 cohort) | 64.6% |
| Reach full term (2014 cohort) | 41.4% |
| Abandoned before full term | 58.6% |
Computed from 27,273,654 USPTO maintenance fee records covering 8,262,336 US utility patents — see the patent survival curve.
A large share of those abandonments were made without evidence. Nobody checked whether a competitor practised the claims, because checking cost more than the fee. Cheap charting changes that calculation directly. See patent portfolio management.
Building a workflow
The useful pattern is AI for breadth, humans for judgement, applied stage by stage.
| Stage | AI does | Human does |
|---|---|---|
| Idea screening | Broad semantic search | Reads the close references |
| Filing decision | Surfaces the landscape | Decides whether to file |
| Drafting | First-pass text, alternatives | Claim scope |
| Prosecution | Finds analogous arguments | Strategy and amendments |
| Portfolio review | Scores and triages | Keep, sell, release |
| Enforcement | Identifies candidate products, drafts charts | Verifies evidence, decides to assert |
Every row where a decision is made stays human. Retrieval and scoring scale; judgement about how an examiner or a competitor will behave does not.
The biggest practical gain is doing analysis that was previously skipped — charting a whole portfolio against competitor products, or searching before every filing decision rather than only the important ones.
Using AI tools for patent work: the checklist
- Use AI for the first prior art pass. It is fast, cheap, and finds documents keyword search misses.
- Verify every citation before relying on it. Open the document and read the passage cited.
- Treat a clean search as unproven, not as clearance. Missed art is invisible until an examiner or challenger finds it.
- Commission a professional search before a non-provisional on anything significant, particularly for non-patent literature.
- Never file an AI-drafted claim set unreviewed. Check specifically whether the claims recite a mechanism or a result.
- Confirm the specification describes alternatives, since that determines how broadly you can ever claim.
- Name only natural persons as inventors, and document the human contribution to conception at the time.
- Submit everything material the tools surface. The disclosure duty is unchanged by how the work was produced.
- Check data retention before entering unfiled inventions into any tool.
- Redeploy the saved time into evidence for renewal decisions — charting against real products is where cheap analysis pays best.