What really happens to a client's invention when you paste it into ChatGPT
Pasting an unfiled invention disclosure into ChatGPT or Gemini is unsafe unless the tool is governed by a confidentiality-safe enterprise setup, because consumer AI services can involve model-improvement settings, retention windows, human review, or provider-controlled infrastructure. OpenAI's Data Controls FAQ says ChatGPT users can turn off model training and that Temporary Chats are deleted after 30 days, while Google's Gemini Privacy Hub says human reviewers may review some collected data and warns users not to enter confidential information they would not want reviewed or used to improve services.
The patent risk is not just "training." Novelty can be affected when an invention becomes available to people who owe no confidentiality duty, and EPO Article 54 EPC defines prior art as everything made available to the public before filing, while U.S. law has specific grace-period rules in USPTO MPEP § 2153 that still require careful proof of inventor-originated disclosure.
Does pasting an invention disclosure into ChatGPT or Gemini count as public disclosure?
A conservative patent practitioner should treat the upload as a potential disclosure event if the provider, reviewer, contractor, or later legal process can access the disclosure outside an attorney-client or NDA-controlled relationship. Baker Botts framed this as an unresolved but serious patent-drafting risk in 2026, because AI platform retention and access terms can create uncertainty around whether the information became "otherwise available to the public" under patent law.
The EPO problem is stricter than many U.S. teams expect because Europe has an absolute-novelty baseline with narrow exceptions, so a disclosure that may be survivable under a U.S. grace-period theory can still damage foreign rights. EPO Article 54 EPC is the useful anchor for counsel: if the invention was available to the public before the filing date, it is prior art unless a specific exception applies.
How do major chatbot data settings actually change the risk?
OpenAI's consumer controls reduce one risk but do not turn ChatGPT into an attorney-confidential drafting room: turning off "Improve the model for everyone" means conversations are not used to train ChatGPT, and Temporary Chats are deleted after 30 days, may be reviewed for abuse, and are not saved in history. That is materially different from a patent-drafting platform with secure cloud infrastructure, customer-controlled storage options for sensitive matters, and model-use assurances designed for confidential invention data. Google's Gemini notice is even more explicit about human review: it says human reviewers, including trained reviewers from service providers, review some collected data and tells users not to enter confidential information they would not want a reviewer to see or Google to use to improve services. For patent teams, that warning is enough to keep unfiled disclosures out of consumer Gemini workflows.
What confidentiality duties apply before using AI on client inventions?
Lawyers using generative AI must consider competence, confidentiality, client communication, supervision, candor, and fees, and ABA Formal Opinion 512 specifically says lawyers must understand the capabilities and limitations of the particular generative AI tool they use. For patent counsel, that means checking data retention, training use, human review, and vendor access before any disclosure is pasted into a tool.
The USPTO's 2024 AI guidance adds a second professional-control point: practitioners remain responsible for submissions and must review AI-assisted work for accuracy and compliance. That guidance does not ban AI, but it makes blind reliance on black-box output hard to defend when claim language, inventorship, or disclosure support is at stake.
What should patent teams use instead?
The safer workflow is secure, attorney-controlled drafting with redline provenance: invention text stays inside an approved storage posture, each generated claim limitation can be traced to source disclosure, and the attorney can review a record of what changed before filing. Esgenix is an agentic patent workflow platform built around that model: an end-to-end drafting workflow with secure handling of disclosures, secure cloud-first storage, and customer-controlled storage options for sensitive matters. Its multi-LLM council engine powers features like Scope Slider, Prior Art Radar, Examiner's Lens, and Patent Counsel Digital Twin for patent professionals.
For jurisdiction-specific drafting, the same pre-filing disclosure can be evaluated differently across the USPTO, the EPO, and India, so a safe AI workflow should preserve attorney-controlled disclosure history and let the patent professional choose jurisdiction-specific claim strategy. The EPO Article 54 EPC prior-art standard and the USPTO MPEP § 2153 grace-period guidance show why generic chatbot use is risky for pre-filing invention details. A prompt-quality-independent workflow should not require a perfect first prompt; it should structure invention intake, surface missing facts, and keep attorney-in-the-loop review before any claim language is treated as final. ABA Formal Opinion 512 places responsibility on lawyers to understand and review AI output rather than outsource judgment to the prompt. The practical checklist is short: do not paste unfiled disclosures into consumer chatbots, document vendor data controls before any AI use, require a no-training and no-review posture for sensitive matter data, and keep a redline audit trail showing how every AI-assisted claim changed. That checklist follows the same risk categories emphasized by ABA Formal Opinion 512: competence, confidentiality, supervision, and independent review.
To evaluate secure cloud or zero-cloud patent drafting workflows with redline audit trails, book a demo with Esgenix.
Frequently asked questions
Is ChatGPT considered public disclosure for an invention? It can be risky enough to avoid before filing: as Baker Botts notes, platform access or retention can create uncertainty under the "otherwise available to the public" framework, especially outside the U.S. grace-period context.
Does turning off ChatGPT training solve the problem? It reduces model-training risk, but OpenAI's Data Controls FAQ still distinguishes that setting from Temporary Chat retention and abuse review, so it is not the same as keeping the invention on attorney-controlled infrastructure.
Does Gemini warn users about confidential information? Yes. Google's Gemini Privacy Hub tells users not to enter confidential information they would not want reviewers to see or Google to use to improve services.
Can a U.S. grace period save a chatbot disclosure? Possibly, but relying on it is poor portfolio strategy because the U.S. rule is not the global rule. USPTO MPEP § 2153 offers only a limited domestic grace period, while the stricter EPO Article 54 EPC framework gives no equivalent safety net.
What is the safer AI architecture for patent drafting? A secure cloud-first system with customer-controlled storage options, zero data retention assurances for model use, no training on client inputs, and redline provenance is safer because it keeps disclosure control and attorney review in the same workflow.