Generative artificial intelligence (Generative AI) is increasingly used by businesses, as well as by private practice and in-house lawyers, to support with the delivery of legal services. In this briefing, we consider the interplay between Generative AI tools and legal professional privilege by reference to four scenarios, and consider how the protections provided by privilege may be preserved.
Generative AI and privilege: why does this matter?
When users enter prompts into a Generative AI tool, the prompt and the resulting outputs - including any emails or other documents based on the prompt/outputs - are “documents” for the purposes of legal proceedings. Given the increasing use of Generative AI to deal with real-world legal issues, some of which may develop into disputes, these documents could be subject to disclosure obligations in legal proceedings.
When using Generative AI tools, it is therefore important to consider if and how documents being created can be protected by legal privilege. As yet, there is no substantive English case law authority on how the existing principles of legal professional privilege apply, although it is likely that this will be addressed by the English courts in the future.
Legal professional privilege a brief recap
When considering how the existing privilege rules may apply, it is important to keep in mind the requirements:
- Legal advice privilege protects confidential oral or written communications between a client and their lawyer (or a lawyer's agent) where the dominant purpose is seeking or providing legal advice. The ‘client’ is narrowly construed to mean only those individuals within a company who are authorised to instruct the lawyers and receive legal advice on the entity’s behalf on the matter in question.
- Litigation privilege applies on a wider basis than legal advice privilege. It protects confidential documents or communications between lawyers or their clients and any third party where the dominant purpose of creating the document or communication is to obtain legal advice, information or evidence for use in the conduct of litigation which is ongoing or in reasonable contemplation. Accordingly, litigation privilege protects a wider range of documents.
Scenarios
Scenario 1: A non-lawyer employee in a business enters a prompt into a Generative AI chatbot (e.g. ChatGPT) to request legal guidance on what the business should do based on a particular set of facts. The employee circulates the output generated by the chatbot by email within the business.
This scenario is increasingly common – employees in a business may try to use Generative AI tools to help them with a legal issue before, or instead of, seeking advice from an in-house legal team or external lawyers. Will the prompt and the output by the Generative AI tool (and emails or other documents reproducing or summarising the prompt or output) be protected by legal advice privilege? Our view is that legal advice privilege is unlikely to apply.
Preliminary point: is a Generative AI chatbot a “lawyer”?
The essence of legal advice privilege is that it covers confidential communications between a client and a lawyer concerning legal advice.
Under English case law on privilege, “lawyer” means a qualified legal professional such as a barrister, solicitor, legal executive or equivalent in other jurisdictions. It can also cover other legal staff, such as a trainee solicitor, to the extent that they are working under the supervision of a qualified lawyer. “Lawyer” is defined in a similar way in UK legislation, for example, the definition of solicitor under the Solicitors Act 1974. The Supreme Court has also held, albeit in an IP law context, that an AI tool is “a machine with no legal personality” (Thaler v Comptroller-General of Patents, Designs and Trade Marks [2023] UKSC 49). When applying this definitional approach, a Generative AI chatbot, even one specifically trained on legal material, would not be recognised as a “lawyer” because the law only recognises natural persons with the relevant professional admissions and regulatory status as a lawyer.
Further, in R (on the application of Prudential Plc) v Special Commissioner of Income Tax, [2013] UKSC 1, the Supreme Court held that legal advice privilege does not extend to legal advice given by anyone other than a lawyer and rejected arguments to extend privilege to advice given by accountants, even where the advice was legal in nature. Its reasons included the fact that it was long settled law that privilege only applied to advice given by lawyers, that extending this would cause uncertainty and also that any broadening of legal privilege would raise policy issues that should be addressed by Parliament.
There may also be arguments around whether interactions between an employee and a Generative AI chatbot are “confidential” as required for privilege to apply, particularly when using an “open model” (as discussed further in Scenario 2 below).
Does legal advice privilege apply to this scenario?
In this scenario, where the employee has sought to use Generative AI to directly provide legal advice (rather than through a lawyer), the advice generated by a Generative AI chatbot is unlikely to be protected by legal advice privilege under the existing law. As such, the prompts, outputs and emails or other documents based on the prompt or outputs may be disclosable in legal proceedings (or in regulatory document requests).
While this issue has not been directly addressed by the English courts to our knowledge, an extension of legal advice privilege to situations where no “lawyer” is involved would be a seismic shift and potentially raise regulatory issues and is likely to be an area looked at through government-level policy making, consultation and eventually changes to the relevant legislative and/or regulatory regimes.
What about if a lawyer is involved?
If Scenario 1 is changed so that the client uses a Generative AI tool to help them write an email to a lawyer with the dominant purpose of seeking legal advice, then legal advice privilege should arguably apply to the final email sent to the lawyer, even though Generative AI has been used to help draft it.
Similarly, if a lawyer uses an AI tool to assist them in providing advice to a client, arguably both the prompts and the output would be covered by legal advice privilege (in addition to the final advice sent) on the basis that they form part of the lawyer’s privileged working papers and potentially reveal the substance of their legal advice (subject to the confidentiality point discussed in Scenario 2 below).
What if litigation was reasonably in prospect at the time?
It is possible that litigation privilege would be available to protect the interactions with the AI chatbot if the employee’s dominant purpose in using it was to obtain advice, information or evidence in relation to the litigation.
Is an AI application a “third party” for the purposes of litigation privilege? It seems unlikely that an AI application would be seen as a “third party”; in the context of privilege law, “third party” generally means a factual witness, expert witness or other professional involved in the litigation (and “party” generally means a legal or natural person).
However, arguably this may not prevent litigation privilege from applying. In the same way that communications between two employees in the business relating to the litigation may be protected by litigation privilege despite not being a communication to a third party, a court may hold that the interactions with the AI chatbot are protected provided the “dominant purpose” test is met (and the communication is sufficiently confidential as discussed in Scenario 2 below).
Scenario 2: An employee in a business instructs a solicitor to provide legal advice. After they receive the privileged legal advice from the solicitor, they input it into an open Generative AI chatbot (e.g. ChatGPT), asking it to summarise the advice.
Again, this situation is not uncommon. This is different to Scenario 1 as the material provided to the chatbot is pre-existing legal advice or evidences pre-existing legal advice. The question is how the existing claim to privilege over the legal advice is affected and how privilege applies to the output.
The principal risk to maintaining privilege over the advice in this situation is waiver of privilege through loss of confidentiality. This is particularly an issue if an ‘open model’ is used (i.e. the tool is available to the public and trained using publicly available data, such as ChatGPT). It is impossible to guarantee that data inputted in open AI tools will remain confidential.
Recent guidance issued by the Courts and Tribunal Judiciary of England and Wales (link here), as referenced in The Commissioners for HM Revenue and Customs v Marc Gunnarsson, [2025] UKUT 00247 (TCC), expressly warns against entering "any information into a public AI chatbot that is not already in the public domain” or information that is “private or confidential”, as “any information that you input into a public AI chatbot should be seen as being published to all the world" and mentions risks that open AI models can mishandle confidential information. Similar points are made in R (on the application Ayinde) v London Borough of Haringey, [2025] EWHC 1383 (Admin) which emphasises lawyers’ professional duty to check the accuracy of AI-generated output.
If the Generative AI tool is a ‘closed model’ (i.e. operates within a secure network), then in principle the interactions between the user and the closed AI tool, and the output generated, could remain confidential provided they are stored securely within that network, and access to any materials containing legal advice is restricted to those who need it. However, this has not been confirmed in any court decision or regulatory guidance, and there would need to be appropriate confidentiality obligations in place between the user’s organisation and the provider.
Subject to the above confidentiality issue, the summary of the legal advice prepared by the AI chatbot should be covered by legal advice privilege on the basis that it will evidence the earlier legal advice.
A further issue that commonly arises is that the recipient of legal advice in a business will want to share it with others in the business who need to know. The employee may want to share the AI-generated summary in a similar way. Case law has established that legal advice can be shared without losing privilege if the sharing is limited, confidential, and necessary for the conduct of the business. Appropriate practical measures should be taken, such as marking the document as confidential and privileged, limiting circulation, and ensuring recipients understand the need to maintain confidentiality. Similar principles would apply to sharing the summary of the legal advice generated by AI within the business.
Scenario 3: You instruct a specialist in a technical field as an expert witness in litigation. They inform you that they intend to use Generative AI to help them prepare aspects of their report.
If Generative AI is used to assist in the preparation of materials for litigation (e.g. an expert report), specific safeguards will need to be considered. We focus on privilege in this example, but for further comments on the use of AI in litigation generally, please refer to our previous articles: here and here.
The work product of an expert witness in litigation will be covered by litigation privilege. However, are the expert’s interactions with a Generative AI tool in their preparatory work (which may not form part of the final report) covered by litigation privilege? Subject to satisfying the confidentiality requirement as discussed below, arguably the prompts submitted and outputs from the AI tool which are for the dominant purpose of producing evidence for use in the litigation will be covered by litigation privilege. They can be viewed as analogous to emails and other communications between the expert and their team who are assisting with the analysis of information and preparation of material for the report, which will be protected by litigation privilege.
However, if an expert witness uses a Generative AI tool to assist in preparing research or drafting a report, there is again the risk that this involves inputting privileged and confidential information into an AI tool. As we considered in Scenario 2, this may result in a loss of confidentiality and therefore privilege. Parties should consider including specific safeguards in expert engagement letters or confidentiality agreements to address the use of Generative AI tools (e.g. stating that the expert may only use closed Generative AI tools and that the tool is being used to provide litigation support).
Scenario 4: An employee identifies that actions taken by the business may amount to a serious breach of an important customer contract. A call is held between the employee and their line-manager to discuss it. The employee uses an AI tool to prepare a transcript of the call.
We are seeing increasing use of AI transcription tools and clients should be mindful about switching these on automatically in relation to legal discussions without considering the context of the discussion and the risks surrounding any document that is produced.
In this scenario, the risk is that the transcript generated by AI may not be protected by privilege; this will depend on whether the underlying discussion is privileged. There is no lawyer on the call in this scenario, and it may be too early for litigation privilege to apply if litigation is not in reasonable contemplation. If the customer ultimately brings a claim and the transcript is not privileged, it could be disclosable. A transcript of an initial internal discussion of a potential legal issue may contain initial impressions and reactions not necessarily based on a considered analysis and the full facts, and which could be unhelpful to the defence of the claim.
If a lawyer is included in the call and the discussion, or parts of it, relate to the provision of legal advice, or if the call discusses legal advice previously given, then the transcript (or parts of it) may be covered by legal advice privilege. The client should take care that any privileged material is clearly identified in the final transcript.
Further, if privilege does apply to the transcript, there is a risk of waiver due to a loss of confidentiality if the transcript is too widely accessible. AI-generated transcripts are often accessible to all participants on a call.
Clients should therefore exercise caution in automatically turning to AI tools to prepare transcripts of calls where legal issues are discussed. For sensitive legal issues, it may be advisable that notes of discussions are prepared by lawyers to protect privilege over legal advice given and to limit dissemination.
Conclusion
Legal professional privilege rules have evolved incrementally and adapted to fit new circumstances, such as new forms of electronic communication. The advent of Generative AI has raised a question: will the law of privilege deal with this development on an incremental basis, or does it instead represent a fundamental change in the dynamic in which legal advice is delivered and litigation is conducted? At this stage, only time will tell, including where any standalone “AI privilege” (as recently called for by Sam Altman at OpenAI) will be recognised.
As things stand, there is limited guidance and certainty on how privilege applies to interactions with Generative AI tools. Clients should give careful thought to managing privilege protection over materials created by Generative AI to ensure that taking advantage of the opportunities offered by the technology does not result in unintended consequences. It is important that lawyers and clients put in place appropriate safeguards, including clear policies and supervision around the use of Generative AI in legal advice and adherence to confidentiality.
The authors would like to thank Tara Cavan for their assistance in preparing this briefing.