03 · What You Need to Know
How AI Systems Can Affect the Confidentiality of Research Inventions
Why does confidentiality matter before a patent application?
Many research inventions begin as confidential technical information. Before a patent application is filed, researchers may need to restrict access to details that explain how the invention works and what distinguishes it from existing technology.
Patent systems generally require inventions to satisfy novelty and other patentability requirements. Information made available to the public before the relevant filing date may become prior art, although the applicable rules and exceptions differ across jurisdictions.
For example, the European Patent Office applies an absolute novelty framework subject to limited exceptions. The United States provides certain statutory exceptions for qualifying disclosures, including a limited grace-period framework. Researchers should not assume that a disclosure tolerated in one jurisdiction will be harmless in another.
Confidentiality may also protect commercially valuable know-how that is not disclosed in a patent application. Trade secret protection generally depends on maintaining secrecy through appropriate measures, among other legal requirements.
These considerations make unpublished technical information different from ordinary research notes intended for public dissemination.
Does entering an invention into AI count as public disclosure?
Not necessarily. A submission to an AI service and a disclosure that makes an invention available to the public are not automatically equivalent.
Whether information has become publicly available depends on the circumstances and applicable law. Relevant factors may include who can access the information, whether confidentiality obligations apply, whether the content is exposed to other users, and whether the disclosure makes the technical invention sufficiently accessible.
Disclosure to an AI service
Technical information is transmitted to a provider or its processing infrastructure. Its confidentiality depends on the service arrangements and actual handling of the information.
Public disclosure
Information becomes available to the public under circumstances relevant to patent law, potentially affecting novelty and other legal requirements.
A confidential submission to an appropriately controlled service may not constitute public disclosure. Conversely, information exposed through publicly accessible conversations, shared links, insecure integrations, or other uncontrolled channels could create substantially different risks.
The distinction should not be interpreted as permission to submit confidential inventions indiscriminately. Even when public disclosure has not occurred, a researcher may still breach an institutional or contractual confidentiality obligation.
Can AI providers retain unpublished invention details?
Depending on the service and its configuration, providers may retain submitted content for specified periods, process it for service operation, permit certain authorized personnel to access it, or use it for purposes described in their terms and policies.
Data handling can differ between consumer accounts, enterprise arrangements, application programming interfaces, and institutionally managed deployments.
Researchers should examine several separate questions:
- Whether submitted prompts and attachments are retained and for how long.
- Whether content may be used for model improvement or training.
- Whether authorized personnel or subprocessors can access the information.
- Whether deletion controls apply to stored conversations and related records.
- Whether third-party integrations receive or process submitted material.
- Whether contractual confidentiality protections cover the intended research use.
A provider's statement that data is not used for model training addresses one potential use of the information. It does not, by itself, establish that the data is never stored, reviewed, transmitted to subprocessors, or otherwise processed.
Likewise, deleting a conversation from the visible interface should not be assumed to remove every retained copy immediately.
Does disabling model training make an AI system safe for confidential inventions?
Disabling training may reduce a particular category of risk, but it is not a complete confidentiality assessment.
Information may still be processed for generating responses, security monitoring, abuse prevention, troubleshooting, or other purposes permitted by the service agreement. The precise arrangements depend on the provider and product.
Researchers should distinguish model-training controls from data retention, access permissions, confidentiality commitments, and the technical security of the deployment.
An institutionally approved AI environment with appropriate contractual safeguards may offer a more suitable arrangement than an unrestricted public chatbot. Even then, approval should be checked against the sensitivity of the particular invention.
Can an AI system reveal an invention to another user?
Researchers sometimes assume that submitting an invention means the system will later reproduce the same technical information for unrelated users. This is not an inevitable consequence of using generative AI.
Nevertheless, unauthorized disclosure may arise through several mechanisms, including misconfigured sharing permissions, insecure integrations, compromised accounts, software vulnerabilities, or unintended exposure through connected tools.
Research on language models has also demonstrated that some models can reproduce portions of memorized training material under particular conditions. Such findings establish a potential information-exposure mechanism, not that every submitted prompt is memorized or retrievable.
The probability and severity of disclosure depend on the specific system and deployment. Researchers should therefore avoid treating either guaranteed secrecy or inevitable leakage as established facts.
How does AI use interact with trade secret protection?
An unpublished invention may contain information that qualifies for trade secret protection if the applicable legal requirements are satisfied.
Under the framework reflected in the WTO Agreement on Trade-Related Aspects of Intellectual Property Rights, relevant considerations include whether the information is secret, has commercial value because it is secret, and has been subject to reasonable steps to maintain secrecy.
Sharing information with an external AI service does not automatically eliminate trade secret protection. However, disclosure without adequate safeguards may undermine an argument that reasonable protective measures were maintained.
This is particularly relevant when the information includes manufacturing methods, unpublished formulations, proprietary algorithms, or other commercially valuable technical know-how.
Patent protection and trade secret protection should not be treated as interchangeable. A patent ordinarily involves eventual disclosure of the claimed invention in exchange for legally defined exclusive rights, while trade secret protection depends on continued secrecy.
What if the invention belongs to a university or research sponsor?
Researchers may have confidentiality obligations even when they personally developed the technical idea.
Employment agreements, sponsored research contracts, collaboration agreements, and institutional intellectual-property policies may restrict how unpublished inventions can be shared.
These restrictions can apply to external AI services, particularly where the service processes information outside the institution's approved environment.
Before uploading unpublished technical material, researchers should establish who holds the relevant rights to the research invention and which confidentiality obligations govern it.
Institutional approval may also be required even when the AI service appears technically secure. The fact that a platform can process confidential information does not necessarily mean a researcher is authorized to submit it.
Some institutions may impose specific restrictions on AI use for commercially sensitive research, including limitations on approved services and external data transfers.
Can removing names and identifiers protect an invention?
Removing a researcher's name, institution, project title, or funding reference may reduce the amount of identifying information submitted. It does not necessarily protect the invention itself.
The commercially valuable information may be contained in the technical description rather than in personal identifiers.
For example, an anonymized prompt that fully describes a novel chemical process may still reveal the essential inventive concept.
Researchers should distinguish anonymizing identifying details from withholding confidential technical substance. Both may be useful, but they address different risks.
Can researchers obtain useful AI assistance without disclosing the invention?
Often, yes. The necessary level of disclosure depends on the task.
A researcher seeking general advice about experimental design may be able to describe the methodological problem without revealing a novel mechanism. Similarly, questions about documentation structure, statistical methods, or general engineering principles may not require proprietary specifications.
Possible approaches include using generalized technical descriptions, substituting hypothetical parameters, separating nonconfidential questions from sensitive material, or using an approved local or institutionally controlled system.
However, abstraction is not always sufficient. A distinctive combination of technical details may reveal the invention even when individual identifiers are removed.
The decision to enter potentially patentable ideas into public AI systems should therefore depend on the information actually required for the task and the protections available.
04 · A Practical Example
A Researcher Uploads an Unpublished Device Design to an AI Chatbot
Hypothetical Example
An unpublished biomedical sensor
A university research team develops a new biomedical sensor that uses a novel arrangement of materials to improve measurement stability. The researchers have completed preliminary laboratory testing but have not submitted a patent application.
One team member wants an AI chatbot to identify possible design weaknesses. The researcher prepares a prompt containing the sensor's detailed mechanism, material composition, and experimental results.
The AI service is accessible through a personal account, and the researcher has not reviewed its data-handling terms or the university's confidentiality requirements.
Step 1: Identify what would be disclosed
The proposed prompt includes technical details that may reveal the invention's distinctive features. Removing the researcher's name would not eliminate this exposure.
Step 2: Examine the service arrangement
The researcher determines that the account's applicable retention, access, and confidentiality conditions have not been verified. The team cannot yet establish whether the service is appropriate for unpublished inventive material.
Step 3: Choose a lower-disclosure alternative
Rather than uploading the full design, the researcher asks general questions about known sensor-stability problems without revealing the novel material arrangement.
Step 4: Seek institutional guidance
The research team consults its technology transfer office to determine whether the invention should be disclosed internally, whether patent filing is appropriate, and which AI environments may be used.
Step 5: Document the decision
The team records which information was withheld, which service was considered, and what safeguards were adopted for subsequent AI-assisted work.
In this scenario, the researchers can obtain some methodological assistance without submitting the confidential inventive concept.
If the detailed design had already been uploaded, the appropriate response would be to preserve relevant records, identify the service and applicable terms, and seek prompt institutional assessment. The upload alone would not establish that patent rights had been lost.
