1. Home
  2. AI News
AI News

AI-Assisted Patents and the Price of Attribution Uncertainty

Attribution uncertainty is a distinct IP risk in AI-assisted inventions. What the USPTO guidance and Fortress Iron mean for valuing patents, diligence and…

Abdelrahman Samir
Certified trainer in AI and marketing for legal professionals · PhD researcher in the creative economy
9 min read

اقرأ المقال بالعربية

Cover: AI-Assisted Patents and the Price of Attribution Uncertainty
ShareLinkedInWhatsAppX
In shortAttribution uncertainty should be treated as a distinct intellectual property risk and priced into the valuation of AI-assisted inventions, argues Paul Atamewan in IPWatchdog. Where human contributions go undocumented, inventorship errors can threaten patent validity and may not be correctable, so buyers, licensees, lenders and insurers should diligence the record rather than the patent face.

What was published

Writing for IPWatchdog on 16 November 2026, Paul Atamewan argues that attribution uncertainty should be treated as a distinct source of intellectual property risk and priced into the valuation of AI-assisted inventions. His reasoning is narrow and practical: where an invention emerges from a workflow in which human contributions were not documented, an inventorship error can threaten patent validity and, in some circumstances, may not be correctable.

The legal anchor is the U.S. Patent and Trademark Office’s November 2025 revised guidance, which, according to the article, confirmed that the same inventorship standard applies whether or not artificial intelligence was used. AI systems are tools, not inventors (citing Thaler v. Vidal), and the guidance withdrew the Pannu joint-inventorship analysis for inventions developed by a single natural person with AI assistance. The inquiry remains human conception, and an inventor must be able to describe the invention with particularity. Critically, the guidance prescribes no recordkeeping method.

He then pairs this with the Federal Circuit’s April 2026 decision in Fortress Iron, LP v. Digger Specialties, Inc. Although the case did not involve AI, the court affirmed summary judgment invalidating two patents for omission of a coinventor and affirmed denial of correction. Under 35 U.S.C. § 256(b), judicial correction requires notice and an opportunity to be heard for all parties concerned; the omitted inventor qualified even without an economic interest, and because that person could not be located, correction was unavailable.

What is genuinely new here

Most commentary on AI and patents circles the same question: can a machine be an inventor? That question is settled for practical purposes. The more commercially damaging question is different, and the article moves it into the valuation room: can the owner prove which natural persons contributed to conception, and can that proof be produced when challenged?

Three features distinguish attribution uncertainty from ordinary validity risk, as Atamewan sets them out. It may depend on private records absent from the patent file. Correction may be unavailable if a necessary person cannot be located and heard. And it is shaped by institutional practices — record retention, inventor interviews, contractor management — more than by technology or claim language. A single-inventor patent with clean assignments looks identical whether it came from a well-documented workflow or one in which three people and a language model each contributed.

The severity is easy to underestimate. In Ethicon, Inc. v. U.S. Surgical Corp., the Federal Circuit held that a person who contributed to two of the patent’s 55 claims was a joint inventor with an undivided interest in the entire patent; he licensed that interest to the accused infringer, and because all co-owners generally must join an infringement action, the suit was dismissed. There is no proportional dilution.

The article’s own contribution is an attribution-confidence score (ACS) across four dimensions: completeness (C), contemporaneity (T), identifiability (I) and independent corroboration (R), each normalised between zero and one, with suggested weights of 0.25, 0.35, 0.25 and 0.15. Contemporaneity carries the most weight because reconstructed evidence is less persuasive. He is explicit that these weights are analytical judgments, not empirically calibrated measures, and that the score is a framework for structured judgment, not a calibrated probability.

He then folds it into an expected-loss model: challenge probability × probability of defect given challenge (first approximated as 1 − ACS, tempered by the clear-and-convincing standard) × probability the defect is uncurable × economic harm. Two portfolios each valued at $50 million before attribution risk, with a 15% chance of challenge and a 90% loss if realised, produce roughly $203,000 expected loss for a well-documented portfolio (ACS 0.85, 20% uncurable) against about $2.13 million for a poorly documented one (ACS 0.30, 45% uncurable). These figures are illustrative, not predictions, and the author flags challenge probability as the weakest input: at 40% for an actively asserted portfolio, the gap grows to roughly $5 million.

No competitor product or rival framework is being displaced here. The contrast is with conventional diligence, which examines assignments, security interests and warranties. Those documents show that the named inventors transferred their rights, not that the correct inventors were named.

Where the framework bites in practice

The three uses below follow the article’s own logic, with examples added.

1. Transaction and financing diligence. Atamewan recommends requesting AI-use policies and adoption dates, retention practices, inventor-identification procedures, contributor assignment agreements, and conception records for material assets. Where records are weak, parties can negotiate price, tailored representations or indemnities, or decline to proceed. A realistic example: a fund acquiring a software portfolio discovers that sole-inventor filings come from teams that otherwise file jointly, and that the patents were filed after AI tools were adopted but before recordkeeping policies existed. That pattern justifies a price adjustment or a specific indemnity rather than a generic warranty.

2. Internal portfolio triage. The article offers concrete diligence tests. Does the organisation have a policy governing AI-assisted research, and when did it take effect? Are prompts, outputs, model versions and selection decisions retained? Is inventorship assessed through substantive interviews with contributors and counsel rather than assigned by default? Do agreements cover contractors, visiting researchers and collaborators? Are there frequent inventor departures, late corrections under 37 C.F.R. § 1.48, or irregularities in declarations? A firm can run these as a scored review before a raise or licensing round, rather than discovering the gap during litigation.

3. Workflow design for research teams. The article describes how AI-assisted research fragments contribution: one person frames the problem, another develops prompts, another selects among generated candidates, refines an output or validates the result. Teams can capture that path as it happens — contributors named, prompts and model versions logged, selection rationales recorded — so the conception record is contemporaneous and attributable. This is a process discipline, not a technology purchase.

Who and what it affects

The article names the affected parties directly: buyers, licensees, lenders, insurers and technology-transfer offices. Each prices the same asset for a different reason and each is exposed to the same blind spot.

For buyers and licensees, the exposure is enforceability rather than novelty. A patent is worth less if its owner cannot reliably establish who invented the claimed subject matter — and the article’s point is that this holds even when the patent’s face looks clean. For lenders and insurers, the expected-loss framing converts a qualitative worry into a sensitivity table. For technology-transfer offices, the output is policy: retention rules, inventor interviews and contractor management become valuation infrastructure.

The methodological point deserves emphasis because it is where professionals will get this wrong. Under the income approach, the adjustment is better reflected in expected cash flows than in the discount rate, because the risk is discrete. Under the market approach it informs comparability adjustments, although comparable transactions may carry the same unpriced risk. The cost approach is less useful because the cost of recreating an invention does not show whether its owner can enforce the patent.

The limits are stated plainly. Diligence cannot prove that no omitted inventor exists or that a record is complete. It can distinguish a documented process from one in which inventorship was decided with little supporting evidence, and that distinction is relevant to price and deal terms. The author also notes that the model treats the four terms as independent, although poor documentation likely travels with poor contractor management, and that it does not apply to defects already known, which should be modelled directly. No reliable base rate yet shows how often AI-assisted patents face inventorship challenges.

What to watch

Recordkeeping as the real control. The guidance prescribes no method, which places responsibility on organisations to ensure their evidence supports the inventorship they declare. The practical risk is not that teams use AI, but that they use it without leaving a trace. Inventor recollection alone may not suffice: conception requires corroboration, assessed under a rule of reason, and records assembled after a dispute arises are worth little.

Correction is conditional, not automatic. Fortress Iron does not mean every inventorship error is incurable, but the correction mechanism has conditions. Contributor locatability and workforce stability therefore become risk inputs, not administrative details — which is why the article ties P(uncurable | defect) to them.

Numbers used as evidence. The $50 million examples and the weighting scheme are illustrative and analytical. Anyone putting an ACS into an investment memo should present the assumptions and a sensitivity analysis rather than market statistics — the author says as much. Frameworks of this kind travel badly when they are quoted without that caveat.

Professional scope. This is a valuation and diligence discipline. The legal characterisation of inventorship on a specific asset remains the work of qualified patent counsel, and the article’s own tests depend on interviews with contributors and counsel. For teams building AI-assisted research processes more broadly, the parallel discipline is verification: see our note on AI hallucinated citations and verifying legal research before court, and for the client-facing dimension of AI adoption, AI for lawyers.

How and when to use it

The article is not a launch announcement and offers no product or trial. It is a framework published on 16 November 2026, available to read now, and usable immediately as a diligence checklist and a valuation overlay.

A sensible sequence for a firm or portfolio owner is short. First, establish whether an AI-assisted research policy exists and when it took effect. Second, check whether prompts, outputs, model versions and selection decisions are retained for material assets. Third, test whether inventorship was assessed through substantive contributor interviews rather than defaulted. Fourth, score completeness, contemporaneity, identifiability and corroboration, and run the expected-loss arithmetic with documented assumptions and sensitivity ranges. Fifth, if records are weak before a transaction, negotiate price, representations or indemnities rather than relying on warranties that cannot be tested.

The important caveat is the one the author supplies: the score organises evidence and structures judgment. It does not prove the record is complete, and no base rate yet exists. Used that way, it is a defensible way to price documentation quality — and documentation quality is something a team can actually change.

Sources

Frequently asked questions

Does the USPTO treat AI-assisted inventions differently from other inventions?

No. According to the November 2025 revised guidance discussed in the article, the same inventorship standard applies whether or not artificial intelligence was used. AI systems are tools, not inventors. The inquiry stays focused on human conception, and an inventor must be able to describe the invention with particularity. The guidance withdrew the Pannu joint-inventorship analysis for inventions developed by a single natural person with AI assistance, and prescribes no recordkeeping method.

Why can an inventorship error be uncorrectable?

Under 35 U.S.C. § 256(b), judicial correction requires notice and an opportunity to be heard for all parties concerned. In Fortress Iron, the court affirmed denial of correction because the omitted coinventor could not be located. The decision does not mean every error is incurable, but the conditions matter — and in AI-assisted research, where records fragment across people, tools, sessions and employers, meeting them is harder.

What is an attribution-confidence score?

It is the framework proposed in the article: a score across completeness, contemporaneity, identifiability and independent corroboration, each normalised between zero and one, with suggested weights of 0.25, 0.35, 0.25 and 0.15. Contemporaneity weighs most because reconstructed evidence is less persuasive. The author stresses it structures judgment rather than producing a calibrated probability, and that no reliable base rate yet exists.

ShareLinkedInWhatsAppX

Want to apply this to your own work or your team? Tell me where you stand, and I will reply personally.

Get in touch
Abdelrahman Samir
About the authorAbdelrahman Samir

I help legal professionals and business owners put AI to work in their marketing and daily practice, and I research the creative economy for my PhD. I write here about what I apply and test myself.