Shadow AI in Law Firms: Risks and How to Manage It

Shadow AI Defined

Shadow AI includes any work-related AI use outside the firm’s approved approach, including personal accounts and unapproved features inside familiar software. The National AI Centre’s guidance on unapproved AI use uses that approval boundary to define it.

Examples in a legal practice include a solicitor pasting a client chronology into a personal chatbot account. A paralegal could upload discovery documents to an unapproved summariser. An assistant might enable an AI meeting recorder without the firm assessing where recordings go.

ChatGPT can be shadow AI when someone uses it outside the firm’s approval. The same product can be authorised in a managed workspace for specified tasks. Approval must cover the product, account, information and use case.

Microsoft and LinkedIn’s 2024 Work Trend Index reported that 78% of surveyed AI users brought their own AI tools to work. The wider survey covered 31,000 knowledge workers across 31 markets. That is evidence of workplace use across industries, not a measured rate of unapproved AI use in Australian law firms.

Why It Matters for Legal Work

Unapproved AI use can disclose client material, complicate privilege protection and introduce errors into legal work. Each risk needs a separate assessment because a privacy setting does not establish permission to use the material.

Confidentiality and Privilege

A provider can receive confidential information when a staff member pastes a document into its tool. Matter names and contextual details can identify a client even after obvious identifiers are removed.

Solicitors’ confidentiality duties continue when they use AI. The Law Society of NSW’s responsible AI guide connects those duties with Rule 9 of the Solicitors’ Conduct Rules. An employee’s subscription payment does not establish the firm’s authority to disclose a client’s information.

Privilege needs its own legal assessment. Disclosure through an AI provider can put confidential legal communications at risk, depending on the circumstances and safeguards. The Law Society of South Australia’s AI guidance identifies possible waiver and restrictions on documents obtained through court processes. Avoid assuming either that every upload waives privilege or that every private account protects it.

Privacy and Accuracy

For entities covered by the Privacy Act, AI use can engage obligations governing personal information, including overseas disclosure. Client authority to handle a document does not automatically permit every secondary AI use.

The Office of the Australian Information Commissioner recommends keeping personal information, particularly sensitive information, out of publicly available generative AI tools as best practice. Its commercial AI guidance calls for assessment of providers’ terms and data handling. Model-training controls address one issue. Retention and third-party access still need assessment.

AI can invent authorities or omit conditions from a summary. If its use remains hidden, the supervising lawyer may never know which passages need checking. The NSW guide requires practitioners to assess accuracy and completeness, and warns against asking AI to verify its own citations. Compare legal propositions with authoritative sources and factual summaries with the underlying documents.

Finding Shadow AI Use

Start with staff conversations and a confidential inventory, then use authorised technical records to fill the gaps. Explain that the immediate task is to protect client information and provide workable tools.

Ask people which tasks take too long and where they already use AI. Include support staff and contractors, who may use tools outside the legal team’s usual systems. Accept an initial anonymous response for discovery, alongside a named reporting route for any actual client-data exposure.

Collect the product and account type, its work purpose and the categories of information entered. Ask whether it connects to files or records meetings. Keep client documents out of the discovery questionnaire itself.

Compare the responses with subscriptions, expense claims and installed browser extensions. IT can examine existing, lawfully collected access logs for AI services. A domain visit shows access to a service. It does not prove that a person uploaded confidential material, and it may miss AI inside another application.

Explain any monitoring before using it and confirm the relevant legal requirements. The OAIC’s workplace monitoring guidance points to applicable Australian, state and territory surveillance laws. Keep access to the findings limited to the people assessing the use and any incident.

Classify each use by the information involved and the proposed task. A public-source drafting task needs a different response from uploading privileged advice. Treat evidence of an actual disclosure through the firm’s incident process, even while the wider inventory continues.

Bringing It Under Control

Give staff an approved way to complete legitimate tasks and a clear route for exceptions or mistakes. Assign responsibility to a named partner or manager, with IT and privacy support where available.

1. Approve specific workflows. Record permitted tasks, account types and information categories for each tool. Assess provider access, overseas processing and deletion terms before allowing client information.
2. Publish the firm’s rules. Use an AI policy template to document approval and supervision requirements. Give staff examples of permitted inputs and prohibited disclosures instead of relying on a product blacklist.
3. Train with fictional matter material. Demonstrate a safe drafting task and an unsafe document upload. Require staff to identify which information leaves the firm and how they must check the output.
4. Provide a reporting route. Name the contact and an alternative if that person is unavailable. Ask staff to report accidental uploads promptly, including the account used and affected information.
5. Make the approved option usable. Provide access and instructions for the tasks staff need. Give requests for another tool an owner and a response timeframe.
6. Review use when conditions change. Reassess new features and integrations, alongside changes to provider terms. Sample approved work for the required review records and close any gaps.

For an accidental upload, stop further sharing and notify the responsible person immediately. Preserve relevant incident evidence securely before arranging deletion or access removal. The OAIC’s data breach response guidance covers containment and assessment, including whether notification is required. Deleting a conversation does not settle confidentiality, privilege or notification questions.

Test the reporting route with a fictional incident. Staff must know who receives it and the responsible person must be able to identify the affected workflow. A useful management result is a recorded decision for each discovered use, with approval conditions or a clear stop and replacement.