Using AI without caution can cost you privilege. That is the message from the courts in recent decisions in both the US and the UK, highlighting that the use of artificial intelligence (AI) systems by lawyers and their clients can have significant legal consequences, including inadvertent waiver of privilege.
As summarised below, the use of ‘open’ or consumer-facing AI platforms, such as ChatGPT or Claude, can undermine confidentiality. In some circumstances, this can lead to the loss of legal professional privilege, with the consequence that advice which would otherwise be protected may become disclosable in legal proceedings.
That said, AI tools can still deliver significant value where they are deployed within an appropriate governance framework, with suitable controls around confidentiality, data handling and use cases.
Legal professional privilege
The starting point remains unchanged. Legal professional privilege protects certain communications from disclosure, and in practice turns on two familiar categories:
- Legal advice privilege: protects confidential communications for the purpose of giving or receiving legal advice, usually between a lawyer and client;
- Litigation privilege: protects confidential communications between lawyers, clients and third parties where litigation is in reasonable contemplation or underway, provided they are created for the dominant purpose of obtaining advice or evidence for the litigation.
Privilege belongs to the client and depends on confidentiality being maintained. It may be waived (intentionally or inadvertently), most obviously where material is shared outside the privileged circle or otherwise handled in a way that undermines that confidentiality.
AI in the context of legal advice and litigation
In the case of United States v. Heppner, No. 25 Cr. 503 (S.D.N.Y.), the U.S. District Court for the Southern District of New York ruled that the use of AI systems to generate arguments and strategies relating to the legal case did not, on the facts of the case, result in the outputs from the AI system having the benefit of legal privilege.
The defendant, Mr Bradley Heppner, was arrested for suspected securities and wire fraud on 4 November 2025. Before he was arrested, he had engaged counsel.
After receiving legal advice, Mr Heppner used Claude (operated by Anthropic) to generate thirty‑one documents relating to his defence. These documents outlined defence strategy and potential legal arguments based on the advice from counsel and which he later shared with his lawyers.
After Mr Heppner’s arrest, federal agents searched his mansion and seized the AI-generated documents. Mr Heppner argued that attorney-client privilege applied to the AI-generated documents.
The court did not agree. The materials were not confidential nor were they communications between client and lawyer, having been independently generated using a third‑party AI platform.
The defendant’s use of a public AI service weighed heavily in the court’s reasoning that the documents were not confidential. In particular, it was highlighted that Claude’s terms of use allow Anthropic to use data inputted by the user for ‘model improvement’ and also state that data could be shared with third parties.
Furthermore, the court did not accept that Claude was agent of the legal counsel. It emphasised that the AI platform was not operating under the lawyer’s direction or control.
Meanwhile, in UK v Secretary of State for the Home Department [2026] UKUT 00081, the Upper Tribunal commented on whether legally privileged material would lose its protected status after being input into an AI platform. The case dealt with two separate sets of immigration proceedings which were brought together for the court to consider the conduct of lawyers following false case citations, generated by AI tools, which had been submitted before the court. During the hearing, an immigration advisor also disclosed that he had uploaded client documents to ChatGPT for the purpose of summarising Home Office decisions.
The tribunal reaffirmed a core principle: legal professional privilege depends on confidentiality. If confidential legal advice is shared with a third party in circumstances where it is no longer controlled or protected, privilege may be waived. The tribunal considered that inputting information into an “open-source” AI tool, such as ChatGPT, is to place it on the internet in the public domain.
The tribunal distinguished between the use of specialist and secure AI tools and non-specialist, freely available “open” tools, and used the terminology “open-source” informally to mean publicly available or consumer tools.
Does the type of AI tool used change the risk profile?
Yes, it can do. In both cases mentioned above, it was relevant that the platforms used were publicly available AI systems, where there is a risk that input data may be shared with third parties.
The risk profile may have been different if an AI system operated in a closed environment with no data sharing were used. Indeed, this distinction was emphasised in the UK case.
It is also notable that US case considered the use of the AI system by a client. It underlines that a client’s interaction with an AI system can impact privilege (even where advice from a lawyer was initially protected).
Although the facts of the cases are different, the message should be the same: both clients and lawyers should be cautious when using AI systems in relation to legal advice or preparation for litigation. It should not be assumed that the use of AI systems in such circumstances is risk free, and a considered approach should be adopted.
This is because many widely available AI tools do more than generate answers. They may store prompts, analyse them, or use them to improve future performance. From a legal perspective, that can look very much like sharing confidential advice with a stranger. Even if no human reads the material, the lack of control over what happens to the data may be enough to undermine confidentiality and, therefore, privilege. Once privilege is waived, it is usually gone for good, and the material may have to be disclosed in litigation or regulatory proceedings.
Is the case law settled?
No. This is a developing area, with courts applying established privilege principles to fast‑moving AI use cases. Although Heppner is a US decision with no direct effect in the UK and the Upper Tribunal decision is not binding on higher English courts, both point in the same direction: privilege depends on maintaining confidentiality, and using publicly available (consumer) AI tools may result in a loss of that confidentiality.
There are, however, indications that outcomes may differ on the facts. Two further US decisions, Morgan v V2X Inc (D. Colo. 30 March 2026) and Warner v Gilbarco Inc (E.D. Mich. 10 February 2026), appear to adopt a more lenient approach where self‑represented litigants used public AI tools. In both cases, the court held that privilege (or protection) was not lost, applying the US “work product doctrine”, which can protect certain materials prepared for litigation.
The key takeaway is that privilege risk will be fact‑sensitive and jurisdiction‑specific. Given the current uncertainty, parties should remain cautious when using AI tools in connection with confidential or privileged material and should assume that using consumer AI platforms increases the risk of waiver unless robust safeguards are in place.
What this means for solicitors in practice:
For solicitors and advisors, the common theme emerging from the case law, and echoed in a recent speech on the topic by Sir Colin Birss (Chancellor of the High Court), is the distinction between publicly available AI tools and specialist, secure systems. Careful use of appropriate AI tools, together with the exercise of personal responsibility, can support and enhance legal work. However, consumer AI platforms may present real risks to privilege and confidentiality. Solicitors should be very cautious about uploading legal advice or documents relating to legal matters into such tools. Preference should be given to closed, contractually-protected AI systems, which do not allow data to be shared externally, and terms and conditions and privacy policies should always be reviewed carefully.
Clear internal guidance is essential. Legal staff should be trained to understand privilege risks and to treat publicly available AI tools as external third parties. Consideration should also be given to maintaining an appropriate record of AI use in legal work (for example, the tools used and the purpose), particularly in more complex or sensitive matters.
As the case law develops, those involved in disputes and investigations should anticipate that questions about prompts and outputs may become the subject of disclosure requests or tactical challenges in litigation.
Solicitors should also consider the risk of AI use by clients. Clients may “test” legal arguments with AI tools or upload documents without appreciating the implications and may later face arguments that privilege has been lost. It may therefore be prudent to notify clients expressly of these risks and to address AI use in terms of business or engagement documentation.
Similar considerations apply to third parties handling client material, such as expert witnesses. Solicitors may wish to warn such parties of the risks or require appropriate confidentiality undertakings, against the use of publicly available AI tools.
Overall, AI use does not have to be incompatible with legal privilege, but safe adoption depends on governance, contractual protections and informed use by both lawyers and those handling client information.
Copyright & permissions
.png&w=3840&q=75)
.jpg&w=3840&q=60)
