AI and NDAs: 7 Clauses to Strengthen a Conventional Confidentiality Agreement
Most businesses do not need a separate “AI NDA.” They need a conventional NDA that clearly addresses how confidential information may be processed, protected, and retained when AI is involved.
Lewis Ho
A Non-Disclosure Agreement (NDA) establishes the ground rules for sharing confidential information. Two parties disclose information for an agreed purpose and agree to limit its use, restrict access, and protect it. AI does not change that basic structure but creates questions that conventional wording may leave unanswered.
Can a recipient use AI to summarize a confidential presentation? Does permission to evaluate technical documents also permit model training? Must an AI-generated summary be deleted when discussions end?
1. Definition of Confidential Information: Cover Information in New Forms
The definition identifies the information protected by the agreement. It may cover documents, presentations, discussions, technical materials, and other information disclosed during the relationship. Confidential information can change form during AI processing. A presentation may become a summary, a document may become a prompt, or technical information may enter a retrieval database.
What to strengthen for AI
The definition should protect those materials to the extent they contain or reveal confidential information. It should not automatically classify every AI output as confidential.
Sample addition
Confidential Information includes copies, extracts, summaries, prompts, outputs, and other derived materials to the extent they contain or reveal Confidential Information. Processing or transforming information through an AI system does not, by itself, remove its confidential status.
The NDA’s usual exclusions should remain, including for information that is lawfully public or independently developed. Both parties must protect confidential content even when it appears in a different format.
This provision does not decide who owns AI-generated work. Ownership and licensing belong in the relevant development or services agreement.
2. Permitted Purpose: Separate Project Use from Training and Reuse
The permitted-purpose clause limits why the receiving party may use the information, for example, to evaluate a proposed collaboration.
What to strengthen for AI
Using AI to support an evaluation is different from using the disclosed information to improve a model, build a reusable dataset, or support unrelated work. The NDA should make that distinction explicit.
Sample addition
The Receiving Party shall use Confidential Information solely for the Permitted Purpose. Unless separately authorized in writing, it shall not use, or permit the use of, Confidential Information to train, fine-tune, or improve an AI model or service, create a reusable training dataset, or support purposes outside the Permitted Purpose.
Address de-identification directly
Removing names does not necessarily remove commercial sensitivity. An anonymized document may still reveal pricing, product strategy, or a proprietary process.
Rather than merely flagging this risk, the NDA can close the gap:
De-identification, aggregation, or other transformation does not authorize use outside the Permitted Purpose where the resulting material continues to contain or reveal Confidential Information.
Neither party can treat permission to analyze information as permission to repurpose it.
If model training is part of the intended project, the parties should agree separately on the resulting model, access rights, permitted reuse, retention, and ownership or licensing.
3. Disclosure to Representatives: Address AI Providers Without Requiring an Exhaustive List
A representatives clause allows disclosure to specified people or organizations, such as employees, advisers, and contractors, usually on a need-to-know basis and subject to confidentiality obligations. An AI provider may receive or process confidential information without fitting neatly into the existing definition of “representatives.”
But requiring both parties to list every AI provider in an introductory NDA may be impractical. Providers, integrations, and processing arrangements can change.
What to strengthen for AI
A more workable approach is to establish conditions for permitted processing, with specific approval where the risk warrants it.
Sample addition
The Receiving Party may use third-party AI services to process Confidential Information for the Permitted Purpose only where the processing is subject to enforceable confidentiality protections, restrictions on model training and unrelated reuse, and access and retention arrangements consistent with this Agreement. Any additional approval requirements agreed by the parties shall also apply.
For especially sensitive disclosures, the parties can instead require approval of a specified environment before use.
Keep responsibility separate from approval
The NDA should also clarify whether the receiving party remains responsible for provider conduct:
Approval of an AI service does not, by itself, release the Receiving Party from responsibility for its handling of Confidential Information under this Agreement.
That allocation should be reviewed alongside the NDA’s liability provisions.
While removing the need to require a complete provider inventory, each party must still verify that its chosen service supports the obligations it is accepting. An “enterprise” label or a statement that customers own their data is not enough on its own.
4. Standard of Care: Make Protection Apply to AI Workflows
A standard-of-care clause usually requires reasonable protection, sometimes measured against how the receiving party protects its own confidential information. The obligation should extend to the ways employees and contractors actually use AI.
What to strengthen for AI
Relevant measures may include restricting personal-account use, limiting access to project materials, selecting appropriate retention settings, and preventing information from entering shared knowledge bases.
Sample addition
The Receiving Party shall apply reasonable technical and organizational measures to protect Confidential Information processed through AI systems, including appropriate access controls, account settings, and internal instructions to prevent unauthorized disclosure, retention, or reuse.
Both parties need workable internal controls, not merely contractual promises.
5. Unauthorized Use or Disclosure: Include AI Processing in the Notification Duty
An incident provision requires the receiving party to notify the disclosing party when confidential information is used or disclosed without authorization. An unauthorized AI upload may require containment even before the parties know whether anyone accessed the information or whether it was used for training.
What to strengthen for AI
Both parties need a clear escalation route. An initial notice can state what is known, what remains uncertain, and what action has been taken.
Sample addition
The Receiving Party shall notify the Disclosing Party without undue delay after becoming aware of unauthorized AI processing, use, or disclosure of Confidential Information. It shall take reasonable steps to contain the incident and cooperate in addressing its consequences.
More detailed deadlines and investigation procedures can be added for higher-risk engagements.
6. Return or Destruction: Explain What Must Be Deleted and What May Remain
A return-or-destruction clause requires the receiving party to return or delete confidential information when requested or when the relationship ends. The obligation should cover relevant AI conversation histories, uploaded files, and derived materials. It should also distinguish active copies from limited retention exceptions.
What to strengthen for AI
A clear clause should answer three questions:
What must the receiving party delete?
What must it require its providers to delete?
Which copies, if any, may remain?
Sample addition
Upon the Disclosing Party’s request, the Receiving Party shall delete Confidential Information from its active systems, including AI conversation histories and derived materials containing Confidential Information, and instruct its permitted service providers to delete such information under their applicable deletion arrangements.
Copies may be retained only where required by law or held in routine backups that cannot reasonably be selectively deleted. Retained copies shall remain subject to this Agreement, shall not be used for any other purpose, and shall be deleted in accordance with the applicable legal requirement or backup retention cycle.
Both parties should understand material provider-retention limitations before sharing information. And they should separately agree on the deletion timeframe and whether any additional retention exceptions are acceptable.
A no-training commitment is not the same as a no-retention commitment. Similarly, an instruction to a provider to delete information is not proof that deletion has been completed.
The clause should not excuse an earlier breach. If information was uploaded without authorization, difficulty removing it does not make the upload compliant. Nor should either party certify deletion from model parameters or provider infrastructure without a reliable basis.
7. Duration and Interaction with Later Agreements: Preserve the Protection
The NDA specifies how long confidentiality obligations last and how the agreement relates to other contracts. AI-derived materials should remain protected for the same applicable period as the confidential information they contain or reveal.
Later agreements also need careful review. A services contract may introduce broader reuse rights, different liability limits, or wording that supersedes the NDA.
Before signing, check:
Which agreement prevails if provisions conflict.
Whether later liability caps apply to confidentiality breaches.
Whether new data-use rights undermine earlier restrictions.
How long protection lasts.
Whether trade secrets require different treatment.
Whether rights to seek urgent relief are preserved.
The parties should change their confidentiality arrangements deliberately, not accidentally through a later contract.
What Belongs Outside the NDA?
An NDA sets the boundaries for using and protecting confidential information. It does not replace the agreements needed to govern a pilot, development project, or production deployment. As the work expands, additional terms should address processing, ownership, security, and accountability. The trigger is the proposed activity and its risk.
Take personal data as an example. Before processing begins, the parties should assess whether a data processing agreement or dedicated contractual provisions are needed. These should define permitted processing, security measures, retention, third-party provider involvement, overseas access and transfers, incident reporting, and assistance with applicable data protection obligations.
The requirements will depend on the jurisdictions involved. In Hong Kong, the PCPD’s Artificial Intelligence: Model Personal Data Protection Framework provides guidance on AI governance, supplier management, security, and personal data protection. In Singapore, the PDPA establishes the applicable data protection obligations, while the PDPC’s guidelines on personal data in AI recommendation and decision systems provide guidance for those uses.
Where processing spans jurisdictions, the parties should separately assess applicable privacy and cross-border requirements. The NDA’s governing-law clause does not override mandatory laws that apply to the processing.
The Bottom Line
An effective AI-ready NDA should still look and work like an NDA. It should be clear enough to prevent casual misuse and practical enough for both parties to honor.
By strengthening the familiar provisions: definition, permitted use, authorized recipients, reasonable care, incident notification, and return or destruction as well as checking how protection carries into later agreements, the agreement becomes easier to understand, easier to negotiate, and more realistic for both parties to follow.
Lexguard AI provides consultancy support on AI governance and operational readiness. This article and its sample clauses are for general information, not legal advice. Seek qualified legal advice when assessing contractual wording, enforceability, and suitability for your circumstances.
Is a standard NDA enough for AI use?
It may be sufficient if its disclosure and use restrictions clearly cover the proposed activity. For material AI processing, an express AI-use schedule reduces ambiguity. Development and personal data processing usually require additional contractual terms.
Can both parties use one mutual NDA?
Yes, when both disclose information. Each party should act as the disclosing party or recipient for the relevant information. Mutual confidentiality does not require identical technical permissions if the parties use different systems.
Does signing an NDA protect information entered into a public AI tool?
Not by itself. The NDA may bind the person uploading it, but it does not automatically bind the platform or override its terms. Prevention depends on authorization, provider commitments, and technical controls.