USPTO Tightens ODP Access and Draws a Harder Line on AI-Drafted Patents
As of June 18, 2026, the USPTO has moved the Open Data Portal (ODP) fully into an account-based access model. On paper, that looks like a security and traffic-management change aimed at unregistered bulk access. In practice, it is more than that. The agency is starting to separate ordinary public use from industrial-scale extraction of patent data, and it is doing so through identity, registration, and controllable access.
The more consequential signal for patent teams is that AI-assisted drafting is getting harder to treat as a routine productivity layer. Once generative AI is used in specifications, claims, or IDS preparation in any meaningful way, applicants and counsel need to be prepared to show that the core inventive conception remained human, and that AI-generated language or references were reviewed closely enough to support filing certifications and disclosure obligations. The real shift is not whether AI can be used. It is that vague use is becoming much riskier.
ODP restrictions are really about data governance, not just login friction
The USPTO is not taking public data off the table. What it is doing is pushing anonymous, high-volume, low-accountability access to the margins. That matters for anyone relying on foundational patent datasets for model training, competitive monitoring, portfolio analytics, or search enhancement. The operational impact will not stop at creating an account. Teams will need to think about access continuity, rate discipline, credential management, internal controls, and how much of their workflow depends on pulling large volumes of data without interruption.
That also changes the economics of who adapts well. Teams with a real compliance function can usually absorb account-based access, auditing, and traffic controls. The strain falls harder on actors that depended on fragmented identities, outsourced scraping, or the assumption that public availability meant effectively unrestricted extraction. In that sense, the ODP move is not a narrow technical tweak. It is an attempt to reintroduce identity governance into the use of federal IP data.
AI-assisted drafting keeps circling back to one question: who actually conceived the invention?
For most sophisticated applicants, using large language models in patent drafting is no longer unusual. The harder question is where the substantive judgment still sits. If a specification is polished by AI, if claims are expanded by AI, or if alternative embodiments are proposed by AI, can the inventors and counsel still explain, clearly and consistently, what the inventive concept is, why the claim boundary was chosen, and which technical contribution came from a human mind rather than a machine output?
That question matters earlier than many teams assume. A filing can look perfectly orderly on day one and still become fragile later if nobody preserved a reliable record of human technical judgment. The better the AI output looks, the easier it is to forget to ask who made the decisive calls. That is where inventorship disputes, written-description problems, and credibility issues start to build. AI can help produce text quickly. It cannot take over human conception, and it cannot absorb the legal consequences when the record becomes unclear.
IDS practice may become one of the most exposed pressure points
AI is attractive in IDS work for obvious reasons: it can gather references fast, cluster them fast, and help populate forms fast. But speed is exactly why this area can go sideways. The practical danger is often not deliberate concealment. It is the submission of references that are nonexistent, irrelevant, redundant, or only superficially plausible because the output sounded convincing. Once that happens, the filing risk is no longer just about poor drafting quality. It starts to touch reasonable inquiry, candor, and the integrity of representations made to the Office.
Many teams still talk about IDS automation as a labor-saving tool. That is too narrow. The safer objective is not to automate bulk submission. It is to automate suspicion: flag the weak references, isolate the dubious ones, force a second review, and make sure that what goes into the record has actually been checked by someone who is prepared to sign. The firms and legal departments that build that review layer will be in a stronger position. The ones that treat AI-generated lists as ready-to-file work product are creating avoidable exposure.
The next adjustment is procedural, not rhetorical
The immediate response should not be to insert a few generic lines into applications saying the invention was made by humans. That kind of language has limited value if the underlying workflow is thin. What matters is whether teams can preserve a usable record: what came from the inventors, what was rewritten by counsel, what was merely language polishing by AI, what prior-art references were independently verified, and where human review actually happened. Those details feel burdensome when everything is moving well. They become critical when inventorship, enforceability, or prosecution conduct later comes under scrutiny.
The same is true for tool governance. Who is allowed to input technical disclosures into external models, which drafting tasks may use generative AI, what outputs require a second-level review, how AI-supported search is separated from IDS decision-making, and how ODP access will be maintained under a compliant account structure should no longer be left to informal habit. The USPTO’s direction is becoming clearer: the entry point for public data is narrowing, and AI-assisted patent work is moving from “generally acceptable” to “acceptable only if you can explain it, verify it, and stand behind it.” That is a material shift for applicants, counsel, and any business building IP processes around generative AI.
