Confidential information, NDAs and AI: emerging risks
Artificial intelligence (“AI”) is increasingly being used to review documents, prepare summaries and analyse data. In many cases, confidential information is uploaded to AI platforms without sufficient consideration as to whether the use of AI is permitted by the confidentiality provisions governing the information.
Information uploaded to an AI platform is typically processed using systems operated by a third-party provider. Depending on the platform, that information may be stored, transferred overseas or used to train the underlying model.
Depending on the terms of the relevant non-disclosure agreement (“NDA”) and the AI platform in question, processing confidential information through an AI platform may amount to an unauthorised disclosure, creating contractual, commercial and regulatory risk for both the disclosing and the receiving parties.
Confidentiality provisions in an NDA are intended to ensure that confidential information is used only for an agreed purpose and is not disclosed to unauthorised third parties. The use of AI platforms to process confidential information can create uncertainty as to whether these obligations have been complied with.
What are the risks for the disclosing party?
For the disclosing party, the principal concern is the loss of control over information that was disclosed on the basis that it would remain confidential and be used only for a defined purpose. Once confidential information has been entered into an AI platform, the disclosing party will typically have very limited (if any) visibility over how it is processed, where it is stored and who can access it. Even where the AI provider offers robust security measures, the information may no longer be controlled solely by the receiving party or processed within the framework contemplated by the NDA.
This is particularly acute where the information comprises commercially sensitive business data, such as technical know-how or trade secrets, the value of which depends on confidentiality being strictly maintained.
Wider legal and regulatory considerations also arise. Confidential information frequently includes personal data (for example, of customers, contractors and employees). Processing such data through AI systems without appropriate technical and organisational measures may give rise to data protection issues, including questions around lawful basis, transparency, international transfers, security, and processor arrangements.
How can disclosing parties manage the risk posed by AI?
A blanket prohibition on the use of AI may not be commercially realistic or operationally attractive. Many businesses now permit the use of AI in some form or other and may wish to retain the efficiencies that responsible AI use can provide. A more practical approach may be to address the use of AI expressly within the confidentiality provisions of their NDAs.
Disclosing parties may wish to permit the use of AI subject to specified conditions, including:
- making clear that the use of AI is at the receiving party’s own risk and does not limit its obligations or liability under the NDA (thereby preserving the disclosing party’s position in the event of any security breach);
- restricting the receiving party’s use to “closed” AI environments that do not use inputs or outputs for model training, or approved enterprise AI platforms;
- prohibiting the use of confidential information for AI model training, development, improvement or fine-tuning;
- requiring the receiving party to ensure that any AI platform is subject to technical and organisational security measures that, in line with good industry practice, protect against unauthorised access, use, disclosure, alteration or loss;
- limiting the categories of confidential information that may be processed;
- requiring anonymisation or redaction of confidential information where practicable and, where AI-derived data is permitted, ensuring it does not identify the confidential information, the disclosing party or any of its associated persons;
- ensuring that AI-generated outputs are accessible only to persons who are authorised to receive the confidential information under the NDA; and
- ensuring that any confidential information contained in the output of AI processing is treated as the confidential information of the disclosing party.
Clear and express drafting reduces uncertainty and helps avoid disputes over whether the receiving party’s intended use of a particular AI platform is permitted under the NDA.
What are the risks for the receiving party?
Receiving parties should exercise caution before uploading confidential information to AI platforms.
The fact that an AI platform is secure does not, of itself, mean that its use is permitted under an NDA. The key question is whether the receiving party has the contractual right to provide the information to the AI platform at all, and whether doing so is consistent with the purpose, use and disclosure restrictions in the NDA.
Many NDAs limit the disclosure of confidential information to specified categories of recipients and may not contemplate disclosure to an AI platform provider or its sub-processors. Using an AI platform may not fall within those permitted categories. Accordingly, the receiving party may expose itself to liability even where the integrity or security of the information has not been compromised. The mere act of uploading the information may itself constitute a disclosure. Whether that amounts to a breach of the NDA will depend on its terms; but many existing confidentiality provisions were not drafted with these scenarios in mind.
Receiving parties should also consider the practical difficulties that arise in the event of a dispute. If challenged, they may need to demonstrate which AI tools were used, what information was processed and what safeguards were in place at the relevant time. Businesses with clear AI governance policies, approval processes and audit trails will be better placed to evidence compliance if challenged.
Receiving parties should also review the terms of service of any AI platform under consideration to understand how confidential information will be processed, stored and secured, and whether the provider asserts any rights to use, retain, analyse or derive benefit from the information, prompts or outputs.
What about privilege?
A further risk is the potential waiver or loss of legal professional privilege. If privileged material is uploaded to an AI platform, that may constitute a voluntary disclosure to a third party and may undermine the confidentiality on which privilege depends.
Before uploading any confidential material to an AI platform, parties should assess whether the material is or may be privileged, whether the platform provider or any sub-processors may access or use it, and whether the platform’s terms of service and technical controls are sufficient to preserve confidentiality. Even where the platform terms include confidentiality obligations, factors such as the scope of processing rights, the involvement of sub-processors, and the potential for data to be used in model training could undermine any argument that the disclosure was sufficiently restricted to maintain confidentiality and preserve privilege.
Conclusion
The widespread adoption of AI has exposed a gap in many existing confidentiality agreements, many of which were not drafted with AI-enabled processing in mind. As AI technology continues to evolve, industry practice on the treatment of confidential information in AI contexts will develop with it.
- Businesses should review existing confidentiality arrangements and consider whether express AI provisions governing are now required to ensure that those agreements remain fit for purpose.
- Disclosing parties should consider whether existing confidentiality provisions adequately control how disclosed information could be processed using AI tools.
- Receiving parties should avoid assuming that the use of AI is permitted simply because an NDA is silent on the issue.
- Parties should assess whether any confidential material proposed for AI processing is or may be subject to legal professional privilege, and take steps to preserve that privilege before proceeding.
- Careful drafting, clear internal AI policies, appropriate approval processes and robust technical safeguards can help parties realise the benefits of AI while preserving contractual confidentiality protections.
