Checking patentability inside Claude: how founders run a real prior-art search in an AI chat instead of trusting hallucinated patent numbers

The $8,000 Lesson in Prompt Wording

In early 2023, a Boston-based computer-vision startup filed a provisional patent application for what their attorney summarized as "AI-driven anomaly detection in video surveillance feeds." The founders had, by their own account, already "searched" the idea — they had asked Claude whether the concept was patentable and received a confident response citing four specific patent numbers as prior art. The founders reviewed those four numbers in Google Patents. Two did not exist. One described an unrelated medical-imaging process. The fourth was a real patent, but its claims covered hardware signal processing, not algorithmic anomaly scoring. Reassured, they paid $8,000 to draft and file a non-provisional. Eleven months later, a USPTO examiner cited US9,111,148 — a real, discoverable 2015 patent on machine-learning video anomaly classifiers — and rejected their independent claims outright.

The examiner did not find a patent Claude missed. The examiner found a patent Claude was never capable of finding, because Claude does not search patent databases. The founders had confused a language model with a retrieval engine, and the confusion cost them a filing fee, prosecution time, and a claim surface they can no longer recapture under 35 U.S.C. §102(a)(1)'s one-year bar.

This is the specific failure mode this article addresses. Claude is genuinely useful for prior-art search — but only when a founder understands exactly where its reliable surface ends and its hallucination risk begins.

The Hallucination-Surface Boundary

Claude's confabulation risk is not uniform across all prior-art tasks. It concentrates at a precise point: the retrieval boundary — the moment a prompt asks Claude to name existing documents. Ask Claude "what patents cover real-time video anomaly detection?" and it will produce plausible-sounding patent numbers drawn from its training distribution, not from a live USPTO query. It has no mechanism to distinguish a real patent from a statistically coherent fabrication of one.

This is the Hallucination-Surface Boundary: retrieval tasks (naming, citing, confirming the existence of specific documents) fall entirely outside Claude's reliable surface. But structural reasoning tasks — decomposing an invention into claim elements, extracting the technical problem each element solves, generating Boolean search hypotheses for real databases — require no retrieval at all. They require only the kind of structured analytical reasoning that Claude does well, and their outputs can be independently verified in minutes by a non-attorney founder.

The practical implication: stop asking Claude what prior art exists. Start asking Claude what prior art would look like, then go find it yourself.

Step One: Claim-Element Decomposition Before You Touch a Database

The most expensive prior-art searches founders run are unfocused ones — keyword searches on the product name, or searches on the broadest possible description of the technology. A real prior-art search targets claim elements, not product categories.

Claude's first legitimate job is to decompose your invention into those elements. A prompt that works:

"Here is a description of my invention: [paste technical description]. Identify the five to seven independent functional elements that together constitute the inventive concept. For each element, tell me: (a) what specific technical problem it solves, (b) what the closest non-inventive prior approach to that problem would be, and (c) what technical characteristic distinguishes this element from that prior approach."

For the video-anomaly startup, a decomposition Claude would reliably produce might identify: (1) frame-level feature extraction using a pretrained backbone, (2) a temporal aggregation method that compresses multi-frame sequences into a fixed-length embedding, (3) a statistical scoring layer that flags frames by distance from a learned normal distribution, and (4) an alert-threshold calibration step that adapts to scene-specific baseline patterns. Each element is now a distinct search target. The founder is no longer searching "AI video surveillance" — they are searching temporal embedding aggregation methods and adaptive threshold calibration in video classification, which are the precise claim surfaces a USPTO examiner will scrutinize.

Step Two: Generate Boolean Hypotheses, Not Patent Numbers

Once claim elements are decomposed, ask Claude to generate search queries — not document names. A prompt that exploits Claude's reliable surface:

"For each claim element above, generate three Boolean search queries I can run in Google Patents Advanced Search. Use CPC classification codes where relevant, combine them with technical keyword strings, and tell me which fields (title, abstract, claims) to apply each query to."

Claude will produce outputs like: CPC/G06V20/52 AND "temporal embedding" AND "anomaly score" applied to claims text, or CPC/G06N20/00 AND "adaptive threshold" AND "baseline calibration" AND "video" applied to abstracts. These are hypotheses about where prior art might cluster in patent classification space. They are not hallucinated citations. You can copy each query directly into Google Patents Advanced Search, USPTO Full-Text Search, or Lens.org — all free, all real-time, all returning actual document numbers — and verify whether the hypothesis surfaces anything relevant within ten minutes.

This is the workflow the Boston founders needed: Claude as hypothesis engine, real databases as retrieval engines. Neither alone is sufficient. Together, they replicate roughly 70% of what a preliminary patentability search from a technical specialist would surface, at the cost of a few hours rather than several thousand dollars.

Step Three: Feed Real Results Back Into Claude for Claim Mapping

Once you have actual patent numbers and their claim text — sourced from USPTO or Google Patents, not from Claude — the session's second phase begins. Paste the abstract and independent claims of each relevant patent back into Claude:

"Here is the full text of US9,111,148, claims 1 through 4. Does any claim read on element 2 (temporal aggregation to fixed-length embedding) or element 3 (statistical distance scoring) of my invention as I described it? Walk through the claim-mapping element by element."

At this stage, Claude is performing claim-construction analysis on documents you have verified exist. Its output is no longer a retrieval claim — it is a reasoning output about a known text, which is within its reliable surface. The analysis will not substitute for formal freedom-to-operate counsel under Enfish LLC v. Microsoft Corp., 822 F.3d 1327 (Fed. Cir. 2016) claim-construction standards. But it will tell a founder whether a deeper attorney review of that specific patent is warranted, which is a material input to how they allocate their legal budget before any application is drafted.

What Claude Cannot Tell You (And Why That Boundary Matters to Investors)

Three prior-art questions remain outside this workflow entirely, regardless of how the prompt is structured.

• Published patent applications filed in the last 18 months. Under 35 U.S.C. §122(b), patent applications are not published until 18 months after their priority date. A competitor who filed six months ago is invisible to every public search, including Claude, Google Patents, and USPTO Full-Text Search. The only way to surface these is a professional search service with access to pre-publication filing data.
• Non-patent prior art. Academic preprints, conference proceedings, open-source repositories, and commercial product launches all qualify as prior art under §102(a)(1). Claude can help you identify which journals and conferences are most likely to contain relevant disclosures for a given technical domain — a useful scoping exercise — but systematic non-patent prior-art search requires direct database queries in IEEE Xplore, arXiv, and ACM Digital Library.
• Obviousness combinations under §103. An examiner can reject a claim by combining two patents that individually say nothing about your invention. Mapping all plausible obviousness combinations across a prior-art landscape is a judgment call that requires patent prosecution experience. Claude's §103 analysis is directionally useful but not reliable enough to inform filing decisions without attorney review.

Investors conducting IP due diligence will ask whether a startup has identified these three categories explicitly. A founder who can say "we ran a structured Claude-assisted element search, verified results in Google Patents and Lens.org, and then scoped the residual risk in these three categories for professional search" is demonstrating IP process maturity. A founder who says "we searched and nothing came up" raises a diligence flag, especially if they cannot describe the search methodology.

The Prompting Mistakes That Trigger Hallucination

Four prompt structures reliably push Claude across the Hallucination-Surface Boundary and should be avoided entirely:

1. "What patents exist for X?" — Asks for retrieval. Claude will confabulate.
2. "Is my invention patentable?" — Too broad; Claude will give a generic §101 summary without engaging your specific claim surface, and may cite Alice Corp. v. CLS Bank, 573 U.S. 208 (2014) as if it resolves the question when it only frames the test.
3. "Who owns the IP in [technical domain]?" — Landscape ownership questions require live database queries. Claude's training-data snapshot is not current.
4. "Can you confirm that [patent number] does not cover my invention?" — Asking Claude to confirm a specific document's scope on a number you supply is safer than asking it to generate numbers, but if the number was itself AI-generated, you are verifying a hallucination against itself.

Each of these mistakes has a reliable substitute. Replace them with the decomposition and hypothesis prompts described above, and Claude's output becomes something you can act on.

FAQ

If Claude's prior-art search is hypothesis-only, does it actually reduce the risk of examiner rejection — or just give founders false confidence at lower cost?

A Claude-assisted element search reduces the probability that a founder files claims on a claim surface they would have abandoned had they known about existing art — which is the most expensive form of false confidence. It does not eliminate examiner rejection risk, because unpublished applications and obviousness combinations remain invisible. The correct framing for investors: a structured Claude search is a triage layer that identifies whether professional search is urgent before the filing clock starts running, not a replacement for it. Founders who skip this layer and go straight to filing are not saving money; they are converting a $500 search cost into a $3,000–$8,000 prosecution problem.

Does using Claude for prior-art search create a waiver or duty-of-disclosure problem at the USPTO?

Under 37 C.F.R. §1.56, applicants and their attorneys have a duty to disclose information material to patentability — including prior art discovered during a search. If a Claude-assisted search surfaces a real patent (via a hypothesis that you then verified in Google Patents), that patent is material and must be disclosed on an Information Disclosure Statement. The Claude session itself is not a disclosed document; the real patent it helped you find is. Hallucinated patent numbers Claude generates but you never verify are not material to anything, because they are not real documents. The duty applies to what you found, not to how you searched.

At what funding stage does a Claude-assisted prior-art search become legally insufficient, and what triggers the upgrade to professional search?

The threshold is not funding stage — it is claim-surface value and competitive exposure. A founder filing a provisional to lock a priority date on a concept that will be refined over 12 months can run a Claude-assisted triage search responsibly. A founder filing a non-provisional with independent claims they intend to assert, or a startup approaching Series A where IP is a valuation input, needs a professional patentability search before filing. The three categories Claude cannot cover — unpublished applications, non-patent prior art, and obviousness combinations — are exactly the categories that appear in investor IP diligence and in IPR petitions filed by well-resourced competitors.

Can a competitor's use of Claude for prior-art research create a freedom-to-operate risk for my issued patent?

Not directly — freedom-to-operate analysis concerns whether a product infringes an issued patent, not how prior-art research was conducted. The indirect risk is that a competitor using Claude correctly identifies the same prior art an examiner would find, uses it to draft design-around claims before you file, and thereby narrows your claim surface in the same prosecution window. This is a race dynamic, not a legal exposure, and it cuts both ways: a founder who runs a structured Claude search before filing is better positioned to draft claims around identified prior art rather than discovering it during prosecution when amendment options are constrained by the written description requirement under §112.

If Claude generates a plausible-sounding patent number that I verify as real, is that a reliable search result or a coincidence I should distrust?

Treat it as a coincidence until you have verified that the patent's actual claims are relevant to your invention — not just that the patent exists. Claude's training data includes enough patent-number formatting that a fabricated number may occasionally match a real but unrelated document. The verification step is not "does this number exist in Google Patents?" — it is "do the independent claims of this patent read on any element of my invention?" Those are different questions. A real patent with irrelevant claims provides no prior-art information about your invention and no basis for amending your claims. The element-mapping prompt described above is the correct verification step, not a simple existence check.

This article is for informational purposes only and does not constitute legal advice. Consult a registered patent practitioner before making filing decisions.