Are AI chat histories discoverable? The New Zealand position and early overseas signals
The familiar warning in litigation is never put anything in writing that you would not want read by a judge. As artificial intelligence becomes part of everyday legal practice, that warning may also apply to information entered into AI platforms.
Lawyers and clients are increasingly uploading commercially sensitive, confidential and legally privileged information into generative AI tools. This raises two important questions: are AI prompts, outputs and chat histories discoverable, and can using an AI platform compromise confidentiality or legal privilege?
Although these issues remain largely untested in New Zealand, courts are unlikely to develop a separate set of discovery rules for AI. Instead, AI-generated material is likely to be treated in much the same way as any other electronically stored information.
The New Zealand position
In New Zealand civil proceedings, parties are generally required to discover documents within their possession or control that are relevant to the issues in dispute, unless a recognised basis exists to withhold them, such as legal privilege.
The definition of a "document" under the High Court and District Court Rules is deliberately broad. It includes information stored electronically and information derived from that information. There is therefore no obvious basis for excluding material simply because it was created or stored within a generative AI platform rather than an email system, messaging application or document management system.
AI prompts, outputs, uploaded documents and chat histories may therefore be discoverable where they:
exist in recorded form;
are within a party's possession or control;
are relevant to the issues in dispute; and
are not protected by privilege or another recognised ground for withholding production.
Once litigation is reasonably contemplated, parties also have an obligation to preserve potentially discoverable material. Depending on the circumstances, that may require relevant AI conversations and associated documents to be retained or exported before they are deleted.
Control and proportionality
Whether AI material is within a party's "control" will depend on the platform being used. Information that remains accessible through a user's account or can readily be exported is likely to be considered within that user's control. Material that has been deleted, or provider logs and metadata that users cannot access, may be treated differently.
Discovery obligations are also subject to proportionality. Courts are unlikely to require parties to undertake expensive or technically complex searches where the likely benefit is limited. If identifying a small amount of relevant material would require extensive review of lengthy or poorly structured AI chat histories, a court may instead limit the scope of discovery or require the parties to agree a practical search methodology.
Privilege and confidentiality
The use of AI raises separate questions about legal privilege.
A prompt or AI-generated response is not privileged simply because it relates to legal issues, was created by a lawyer or was generated using a law firm account. Equally, the use of AI does not automatically prevent privilege from applying.
Legal advice privilege may protect confidential communications between lawyer and client made for the purpose of obtaining or providing legal advice. Litigation privilege may also apply where material is created for the dominant purpose of existing or anticipated litigation. For example, prompts developed by lawyers to test legal arguments or evidential theories may attract litigation privilege, whereas prompts used for general research or administrative purposes may not.
A more difficult issue is whether privilege is lost when confidential material is entered into an AI platform. Under the Evidence Act 2006, waiver depends on whether the privilege holder has acted inconsistently with maintaining confidentiality. The assessment is contextual and may include:
whether the AI platform is public or private;
the provider's confidentiality obligations;
who can access the information;
whether the provider may use the data for training or other purposes; and
the platform's retention and deletion settings.
One of the greatest practical risks arises where clients discuss disputes with publicly available AI chatbots before speaking with their lawyer. Those conversations may be relevant to later litigation, may not attract legal privilege in the first place, and may, in some circumstances, result in privilege being waived.
Early overseas guidance
Although New Zealand courts have not yet considered these issues directly, decisions from overseas provide some indication of how courts may approach them.
United Kingdom
In UK and R (Munir) v Secretary of State for the Home Department, the Upper Tribunal criticised a solicitor who uploaded confidential client material to ChatGPT for summarisation and drafting assistance.
The Tribunal concluded that uploading confidential documents into what it described as an "open-source" AI tool placed the information into the public domain, breached client confidentiality and waived legal privilege. While those comments arose in the context of professional conduct rather than discovery, they provide a clear warning about entering confidential client information into publicly available AI platforms.
Importantly, the Tribunal distinguished closed-source AI tools, which it considered less likely to place information in the public domain.
United States
Recent US decisions suggest that discoverability will often depend on how AI was used and whether a party relies on AI-generated material in the litigation.
Where lawyers developed prompts as part of litigation strategy or investigation, courts have recognised that those prompts may attract work product protection. However, where parties relied on AI outputs in pleadings or put AI-generated investigations in issue, courts required disclosure of the relevant prompts, outputs and associated settings to enable the opposing party to test the evidence.
By contrast, where a litigant independently used a consumer AI platform without legal direction, courts have been less willing to recognise privilege or work product protection.
Although these decisions arise under different legal frameworks, they illustrate a common principle: the purpose for which AI is used, and how the resulting material is relied upon, will often determine whether it must be disclosed.
Practical implications
Until New Zealand courts provide further guidance, parties should proceed on the basis that:
AI prompts, outputs, uploads and chat histories may be discoverable.
Relevant AI material should be preserved once litigation is reasonably contemplated.
Confidential or privileged information should not be entered into consumer AI platforms without understanding the platform's confidentiality, retention and data-use practices.
Privilege should not be assumed simply because AI is used by a lawyer or for legal work.
Reliance on AI-generated material in pleadings or evidence may make the underlying prompts, settings and outputs discoverable.
Where retrieving AI data would be disproportionate, parties should seek agreement or directions from the court rather than ignoring the issue.
Key takeaway
There are currently no special discovery rules for AI-generated material in New Zealand. Existing discovery principles continue to apply. If AI prompts, outputs, chat histories or uploaded documents are relevant, within a party's control and not protected by privilege, they may well be discoverable.
Contributed by:
Richard Idoine, Claudia Cooper and Harriet Kennelly
