IP News
Audit the Channel, Not the Model: What AI Use Really Puts at Risk in Patent Practice
•
9 min read
•324 views
//
Using AI tools during patent preparation does not automatically create prior art or destroy novelty. The real risk is a confidentiality failure driven by the channel—the specific product tier, provider terms, and retention settings—through which the information travels. For practitioners, the immediate casualty of a wrong channel choice is typically trade secret protection, not patentability, though foreign filing rights can be forfeited even when a U.S. grace period applies.
On July 15, 2026, IPWatchdog published an analysis by Lana Akopyan dismantling the claim that AI use inherently destroys patent rights. The article argues that warnings about prompts becoming prior art or corrupting inventorship misidentify the problem. The risk “almost never comes from the AI model’s involvement,” Akopyan writes. “It comes from whether the information stayed confidential, and that turns on the channel it traveled through: the product and tier, the provider’s terms in force, the retention and training settings, who may see the data, where the output goes, the jurisdiction, and now the specific model.”
The analysis separates two distinct failures: a confidentiality failure, where the wrong channel surrenders protection, and a judgment failure, where unexamined AI output weakens the asset itself. It then isolates the overstated risk (loss of novelty under Section 102) from the real ones—trade secret loss and privilege waiver—and addresses inventorship as a separate issue not driven by channel selection.
Three routine scenarios now generate concern across patent practices: an inventor pastes an unfiled disclosure into a chatbot to clean wording; an associate runs a draft specification through an AI tool to pressure-test claim support; a client forwards an AI-generated analysis of their own case. Each has produced warnings that the prompt becomes prior art, novelty is lost, or inventorship is corrupted.
Akopyan frames the legal question under Section 102(a)(1) as one of public accessibility: whether interested persons exercising reasonable diligence could locate the material, citing In re Hall and GoPro v. Contour (898 F.3d 1170, Fed. Cir. 2018). Disclosure to someone under an obligation of confidentiality does not meet that standard, per Cordis v. Boston Scientific, though a confidentiality label alone will not save material disseminated broadly, per Weber v. Provisur.
A prompt submitted through a business, enterprise, or API channel with real confidentiality, no-training, and retention limits is not indexed, searchable, or locatable. Retention by a provider is not publication. Two contrary arguments fail: the technical argument that a model splices input fragments into training data and serves them to another user skips the public-accessibility requirement; the legal argument that permissive provider terms make input public conflates disclosure to a bound party with disclosure to the public. The on-sale bar under Helsinn v. Teva is distinguished—a prompt is not a sale or offer for sale.
The analysis recalibrates where practitioners should direct their caution. On a controlled channel, novelty is intact. On a consumer or individual tier whose terms permit training, broad retention, human review, or disclosure without meaningful confidentiality commitments, the immediate damage is usually not novelty but loss of secrecy. Training on input does not, by itself, make it publicly accessible; the world still cannot retrieve the specific prompt. Novelty is threatened only where the input becomes publicly retrievable, or where foreign rights are in play.
That last point carries direct consequences for Korea-based filers. The U.S. grace period under Section 102(b)(1)(A) can absorb an inadvertent domestic disclosure, but it does not rescue absolute-novelty jurisdictions. A Korean applicant who uses an uncontrolled consumer AI channel before filing loses the ability to secure valid rights in Korea, China, Japan, and Europe—even if U.S. novelty survives. The loss is immediate and jurisdictional, not theoretical.
Concrete action: Before any inventor or drafter submits technical content to an AI tool, verify the provider’s terms for the specific tier in use. Confirm that the channel imposes no-training obligations, zero retention, and no human review. Document that verification in the file. For any tool that does not meet that standard, treat the session as a disclosure that forfeits absolute-novelty foreign rights and adjust the filing strategy accordingly—file before use, or restrict use to jurisdictions with a grace period and accept the foreign loss.
Expect continued confusion in the market as AI providers market enterprise tiers with varying confidentiality postures and as courts have not yet tested whether a provider’s internal retention constitutes public accessibility. The key variable to watch is whether any plaintiff successfully establishes that AI training data is “publicly accessible” under Section 102—a finding that would upend current assumptions. No such holding exists as of the article’s date.
What to do now: For any pending application where an inventor or agent used a consumer-grade AI tool before filing, commission an immediate foreign-filing rights audit. Identify every jurisdiction where absolute novelty applies and assess whether the disclosure can be argued to have remained confidential under that jurisdiction’s standard. Where the answer is no, decide now whether to abandon those foreign filings or proceed with a documented risk acceptance. For new matters, adopt a written AI-channel policy that names approved enterprise tools by tier and prohibits all others for pre-filing technical content.