Measuring the Gap

The demographic numbers across this series are not all the same kind of fact. Patent records directly report inventor names, filing dates, technology classes, and grant outcomes. What they generally do not report is an inventor's self-identified gender, race, or veteran status.

So researchers reconstruct demographics by inferring likely gender from names, linking patents to other datasets, or surveying inventors, each method answering a slightly different question with a different kind of error. That is why figures that seem to describe "the same gap" can count completely different things: the USPTO's 12.8% concerns unique U.S.-based inventor-patentees on grants in 2019, while WIPO's ~18% concerns inventor listings in global PCT applications. Neither is wrong; they are not interchangeable.

This article presents published research and legislative-record data for educational purposes. It is not legal or policy advice. Legislative status last checked 10 July 2026. See How we measure for the statistical methodology.

What is observed, inferred, and missing

A U.S. patent application requires each inventor's legal name, residence, and mailing address. It does not routinely collect self-reported gender, race, or veteran status in the examination record. But that is different from saying the USPTO never infers gender: the agency's own economic researchers run a name-based gender algorithm and release the results through PatentsView (the 12.8% figure comes from exactly this USPTO inference). So the accurate statement is "not collected as self-report during examination," not "never inferred or stored."

Name inference can be useful at population scale, and its accuracy is measurable: PatentsView attributes a gender to roughly 92% of USPTO inventors at its confidence thresholds. But accuracy and coverage vary in knowable ways:

• Unisex, ambiguous, abbreviated, or initials-only names may go unclassified, which changes the study population. Excluding unclassified records is itself a source of selection bias: WIPO found the applications it had to drop tended to have larger teams.
• Coverage and error rates vary by language, location, and era. East Asian given names are often unisex in romanization, so cross-country comparisons are shakier than they look, and misclassification may be concentrated in particular groups rather than distributed as random noise.
• Binary name inference cannot represent nonbinary identity, and it cannot identify whether someone is transgender. These are three distinct limits: initials block classification; a binary classifier cannot record a nonbinary category; and no name method can determine transgender status.

None of this makes the headline numbers wrong. It means residual error and selection bias are hard to fully characterize, not that error is unquantifiable, since validation can bound much of it.

The IDEA Act: a recurring proposal for direct data collection

The Inventor Diversity for Economic Advancement Act (IDEA Act) is a recurring bipartisan proposal for direct demographic collection in the patent system. Its latest confirmed form was the 118th Congress's S.4713 (House companion H.R.9455), "to require the voluntary collection of demographic information for patent inventors." Read against the bill text, it would:

• provide for voluntary collection of gender, race, and military or veteran status from each inventor residing in the United States who is listed with a patent application (note this is inventors, not "applicants" — in patent law an applicant may be an assignee or other rights holder);
• keep responses confidential and separate from the application, and bar examiners from accessing or considering them;
• exempt individual responses from public disclosure under FOIA;
• require annual public reports, disaggregated by demographics, technology class, and U.S. state; and
• require public data that supports subgroup cross-tabulation, with personally identifying information anonymized or omitted. (The text does not promise row-level "microdata," and it does not list entity size as a cross-tabulation category.)

Its legislative history, stated neutrally:

• IDEA collection provisions appeared in the Senate-passed USICA and House-passed America COMPETES Act (as Section 80102 of the House text), but were not included in the final CHIPS and Science Act.
• S.4713 was introduced July 2024, reported without amendment from the Senate Judiciary Committee on 18 November 2024, and not enacted before the 118th Congress ended.
• As of 10 July 2026, we found no patent-focused IDEA Act introduced in the 119th Congress in the official bill records reviewed. (The similarly-named H.R.5826 in the 119th is an unrelated minority-business entrepreneurship-grant bill.)

Why the USPTO doesn't already have a demographic census

This is not simply bureaucratic neglect; every collection approach involves a real trade-off, which is why the question keeps coming back rather than being settled once:

• Mandatory collection could produce more complete data but raises privacy, burden, and acceptance concerns.
• Voluntary collection better respects individual choice but produces nonresponse and self-selection bias (the USPTO has expressly warned that voluntary survey data could be self-selected).
• Record linkage to administrative data avoids new questions on the application but depends on imperfect matching: an earlier USPTO–Census effort matched only about 64% of U.S.-resident inventors on the available fields.
• Name inference is cheap and can be applied to decades of historical records, but classifies identities indirectly.

The 2024 IDEA Act tried to balance these by making disclosure voluntary, separate from the application, inaccessible to examiners, FOIA-exempt at the individual level, and public only in privacy-protected aggregate form. The result would not be perfect data; it would be better-characterized data with a different set of limitations.

Help available now (a separate question from measurement)

Access programs are distinct from demographic measurement, and three commonly-conflated items are different things:

• Patent Pro Bono Program — a regional network matching qualifying financially under-resourced inventors and small businesses with volunteer patent practitioners who provide legal representation.
• Pro Se Assistance Center — provides education and filing help to applicants proceeding without a registered practitioner; its personnel answer procedural questions but cannot give legal advice.
• The Pairolero et al. RCT (AEJ: Economic Policy, 2025) evaluated enhanced assistance for unrepresented applicants, not the Pro Bono Program as such. Both men and women benefited, with patent-obtainment improving especially for women, largely through successful examiner-amendment negotiations.

At the policy level, the USPTO's Council for Inclusive Innovation (CI2) was reconvened at the White House in December 2024 with ten new members, coordinating a whole-of-government approach to broadening participation in invention. Measurement and assistance are complementary: data can locate disparities; councils and support programs can address specific barriers.

What organizations can measure without waiting for Congress

The national measurement gap does not have to be reproduced inside a company or university, but doing it responsibly is not as simple as adding demographic columns to an invention spreadsheet. A sound approach:

• Uses voluntary self-report (with "prefer not to answer"), never name inference of employees' gender.
• Separates demographic responses from the people deciding whether a concept is reviewed or filed. Demographics must not touch patentability, inventorship, promotion, or file/no-file decisions.
• Reports only aggregate results, with response rates and missing-data rates shown, and suppresses small cells (intersectional results can re-identify individuals even without names).
• Tracks the funnel using cohort-based outcomes (follow the applications filed in a period; don't divide this year's grants by this year's filings).
• Never turns demographic goals into inventorship decisions: inventorship is a claim-specific legal determination, not a diversity designation.

Where ObviouslyNot fits

ObviouslyNot can make one early workflow more observable: it analyzes source code to surface candidate code-linked technical concepts and attaches traceable repository references for human review, showing what was surfaced, who was invited to explain the underlying work, and what was routed to counsel.

That is not the same as measuring inventor demographics, and it has firm limits:

• It does not collect a person's gender, determine legal inventorship, assess patentability, or establish that a concept would otherwise have been lost. ("Almost inventor" is not a legal or research category.)
• Source code is only one record of technical work. Architecture decisions, experiments, hardware, security analysis, and user research may live outside the repository, so a review should invite additional contributors rather than treat commit history as an inventorship record.
• Ignoring author identity does not by itself prove a fair result. Which repositories are scanned, who receives which assignments, and who is invited into the follow-up discussion all still shape the outcome.

The strongest supportable claim is a hypothesis, not a result: systematic, code-based concept discovery may reduce reliance on self-nomination at one early stage of invention review. Whether it broadens participation or changes any demographic outcome would have to be measured separately, with a voluntary, privacy-protected demographic process.

Sources

More from the patent gender gap series