AI and Privilege: are your AI chatbot conversations protected?

General purpose AI tools such as ChatGPT and Claude are increasingly being used by consumers for quick, easy and accessible "legal advice". Similarly, legal specific tools such as Harvey and Legora are seeing rapid adoption both amongst corporate counsel and private practice lawyers. But a recent US court ruling has reignited a critical question: what happens to legal privilege when confidential material is shared with an AI chatbot?

While the decision was made under US law, it raises issues that Australian organisations and advisers should be thinking about, particularly around waiver of legal privilege and the limits of protection for AI assisted work.

The US ruling in brief

Earlier this year, a New York court ruled that legal privilege could not be claimed over a defendant's AI-prepared legal documents. The US District Judge found that neither the doctrines of attorney-client privilege nor 'protected work product' would apply.

Further, the Judge stated that the defendant's sharing of information with the AI platform was inconsistent with maintaining confidentiality and therefore, would amount to a waiver of privilege in respect of any otherwise privileged documents and information shared with the AI platform.

Application of the decision to Australian law

Legal professional privilege is essentially a limited right to refuse to produce documents or information that disclose the legal advice and recommendations a person or entity has received from their lawyer. Privileged documents therefore do not need to be produced in response to compulsory court processes such as subpoenas or discovery.

Legal practice in Australia is closely regulated and only a person who has completed a law degree and holds a current practising certificate is able to provide legal advice. While many AI tools can provide superficially convincing responses to queries about the law, they are not lawyers and often lack the nuance and judgment needed to provide correct and appropriate advice about the application of the law to the circumstances.

Advice and recommendations provided by an AI tool are therefore not protected by legal professional privilege.

Where things can get more complicated is when an AI tool is used in relation to actual or anticipated litigation. This is because in Australia, there is a concept known as litigation privilege. Under the Evidence Act, a person cannot be compelled to disclose confidential documents or communications made for the dominant purpose of the client being provided with "professional legal services" relating to actual or anticipated litigation to which the client is or may be a party. If the AI tool is being used by a lawyer to assist them in efficiently providing professional legal services to the client, the prompts and outputs used will likely be privileged. Even if the client is the one interacting with the tool, it is possible that litigation privilege may still apply if they are doing so in order to complete tasks that their lawyer has directed them to undertake or which are being undertaken for the purpose of seeking the (human) lawyer's advice.

If, however, there is no lawyer in the loop and the communications are occurring solely between client and AI, it unlikely that the "professional legal services" element will be met, meaning the communications will not be privileged.

Beware of inadvertent waiver

Even if privilege initially applies, sharing privileged legal advice with a third party, or referring to it in public, can amount of a 'waiver' of that privilege. In US case referred to above, the Judge found that sharing information and advice with the AI tool was enough to waive privilege over that material.

In the context of waiver, the privacy settings and policies applicable to the AI tool are critical. If the AI platform being used is an enterprise grade tool that keeps all data confidential and confined to the particular user's tenancy, uploading privileged material to that tool is unlikely to waive privilege. By way of contrast, if, as was the case in the recent US decision, the user was making use of a free, publicly available AI tool that expressly states that user data and prompts are used by it to train and refine the large language model (LLM) that powers the platform, the risk of inadvertent waiver is real.

In Australia, the question to be answered when determining whether privilege has been waived is whether the party's actions were inconsistent with maintaining the confidentiality and privilege of the relevant documents and advice.

On the one hand, uploading documents to an AI model or other cloud based software program is not significantly different to photocopying or scanning them on a public or third-party device - it is theoretically possible that third parties will be able to access them, but there is also a social or moral expectation and understanding that they will not do so. The difference with public / free AI models however is that the price of use is the giving of permission to the model to make use of your data and prompts to improve its outputs not just for you, but also for other users. In that way, it's not much of a stretch to conclude that choosing to upload data to a public AI, rather than one in which you are assured the data will remain confidential, is an act that is inconsistent with maintaining privilege over that material.

What does this mean for organisations?

The personalised responses that can be generated by an ever-growing number of AI tools and LLMs can lead people to treat those platforms as a trusted advisor, and to prompt them with information and documents that they would never dream of publishing online.

Before inputting any confidential information into an AI tool, users should consider the relevant platform's privacy policy and terms of use and assess whether the platform gives any assurance that the confidentiality of such information will be maintained. Further, while a time may come when legal advice provided without human involvement can attract privilege, that is not currently the law. For now, if you want to ensure the advice and disclosures are protected by legal professional privilege, only a human lawyer with a practising certificate can provide that protection.