It happens more often than most attorneys would like to admit. A paralegal is buried under three hundred pages of medical records, a demand deadline is approaching, and someone suggests a shortcut: paste the records into ChatGPT and ask it to summarize. It feels harmless. The output looks useful. Nobody gets an immediate alert. But underneath that seamless experience, a chain of data-handling events has already been set in motion — events that most consumer AI platforms were never designed to protect against in a legal or healthcare privacy context.
For personal injury and employment law firms handling sensitive client health information, understanding exactly what happens to that data — and why purpose-built legal AI is a fundamentally different category of tool — is not a technical curiosity. It is a professional concern.
What "Consumer AI" Actually Means for Your Client's Data
Consumer AI tools like ChatGPT, Google Gemini, and similar general-purpose large language model (LLM) platforms are designed for broad public use. Their terms of service, data retention policies, and infrastructure were built for a general audience — not for attorneys bound by professional conduct rules, and certainly not for the handling of protected health information (PHI) under federal privacy law.
When you paste a client's medical records into one of these platforms, several things may happen simultaneously that most users never read about in the fine print:
Data may be used to train future models. Many consumer AI platforms explicitly reserve the right to use submitted content to improve their models. Unless a user has opted out — a process that is not always obvious or complete — the text you submit, including your client's name, diagnosis, treatment history, and provider information, may become training data.
No Business Associate Agreement (BAA) is in place. Under the Health Insurance Portability and Accountability Act (HIPAA), any vendor that handles PHI on behalf of a covered entity or business associate must sign a BAA. General-purpose consumer AI platforms do not offer BAAs to individual users. Without one, transmitting PHI to those platforms creates a compliance exposure that cannot be retroactively corrected by deleting the conversation.
You have no visibility into where the data goes. Enterprise-grade legal tools are built with audit logging, access controls, and data residency guarantees. Consumer AI platforms offer none of these for individual users. Once the text leaves your browser, you have no practical mechanism to track, retrieve, or certify what happened to it.
The gap between "this seems fine" and "this created a serious problem" is often invisible until a complaint is filed or a client asks the question you cannot answer.
The HIPAA Exposure Most Attorneys Underestimate
HIPAA's reach extends beyond hospitals and insurance companies. Law firms that receive, maintain, or transmit PHI in connection with representing clients should consult qualified privacy counsel to assess how HIPAA obligations may apply to their specific practice and the tools they use. What is clear is that any third-party tool processing PHI on a firm's behalf should operate under a signed BAA — and that the firm's ability to demonstrate compliant data handling matters.
The practical consequence: if a paralegal pastes a client's emergency room records, surgical notes, or pharmacy history into a consumer chatbot, the firm has potentially transmitted PHI to a third party with no BAA, no audit trail, and no contractual data-handling protections. That is not a technical violation buried in fine print. It is the kind of exposure that surfaces in bar complaints, malpractice claims, and client trust erosion.
The following is a hypothetical illustrative scenario: A three-attorney personal injury firm is preparing a demand package for a client injured in a rear-end collision. A paralegal, trying to save time, pastes the client's entire treatment history — including a psychiatric evaluation the client had not disclosed to family members — into a consumer chatbot to generate a summary. The summary is useful. But the psychiatric records, the client's full name, date of birth, and treating providers are now sitting in a system with no BAA, no guaranteed data deletion, and no audit log the firm can produce if the client ever asks what happened to their most sensitive health information. The firm did not intend harm. The harm potential exists regardless of intent.
This scenario is hypothetical, but it illustrates a real category of risk that arises whenever PHI is processed through tools not designed for legal compliance.
Why "Deleting the Chat" Does Not Fix the Problem
A common misconception is that deleting a conversation from a consumer AI platform eliminates the data exposure. It does not — for several reasons.
First, deletion from the user-facing interface does not guarantee deletion from backend systems, training pipelines, or server logs. The platform's own data retention policies govern what actually happens to submitted content, and those policies vary significantly across providers and plan tiers.
Second, HIPAA's requirements around PHI are not satisfied by after-the-fact remediation. The obligation is to prevent unauthorized disclosure in the first place. A breach that has already occurred cannot be "un-occurred" by closing a browser tab.
Third, and perhaps most practically: if a client ever asks whether their medical records were shared with any third-party system, "I deleted the chat" is not a satisfying or legally adequate answer. The firm's ability to demonstrate compliant data handling — through audit logs, BAA documentation, and access controls — is the actual standard.
This is precisely why the distinction between consumer AI and purpose-built legal AI matters so much. It is not about which tool produces a better summary. It is about whether the infrastructure surrounding that summary was designed to protect the people whose most sensitive information is inside it.
What Purpose-Built Legal AI Does Differently
Legal AI platforms built specifically for law firm use approach data handling from a fundamentally different starting point. Rather than adapting a general-purpose consumer product for professional use, they are architected around the compliance and confidentiality requirements attorneys actually face.
ProvaLens, for example, signs a Business Associate Agreement before the first document is ever uploaded. Client data is hosted on Microsoft Azure with AES-256 encryption and full audit logging — meaning every access event is recorded and attributable. Documents uploaded to a matter stay within that matter's controlled environment, accessible only to authorized users at the firm.
This architecture is not a marketing differentiator. It is the baseline that distinguishes a tool attorneys can responsibly use from one that creates exposure every time a paralegal tries to save an hour.
Beyond data security, purpose-built legal AI also delivers capabilities that general-purpose chatbots are not structured to provide in a professional context. When ProvaLens reads a case file and returns an answer, it cites the exact page and paragraph in the source document. That citation trail supports attorney review — every claim can be verified against the source document, which is the appropriate standard before any professional reliance on AI output. In a legal context, unverified AI output is not just an inconvenience. It is a liability.
The platform also automatically classifies and organizes documents by type, builds visual filterable timelines from all dates and events across the case file, flags possible inconsistencies and contradictions cited to the source page, and generates itemized medical chronologies and special-damages tallies for demand packages — all exported to Word for attorney review. These are not features a consumer chatbot can replicate in a compliance-ready, citation-grounded way, regardless of how sophisticated its underlying model has become.
Practical Steps to Protect Client PHI Right Now
If your firm has been using consumer AI tools to process medical records or other PHI, the most important immediate step is to stop and assess. Here is a practical framework:
Audit your current workflow. Ask every paralegal and associate where they are using AI tools in case preparation. If consumer chatbots appear anywhere in that workflow, identify which matters may have been affected.
Review the terms of service for any AI tool in use. Specifically look for data retention policies, training data provisions, and whether a BAA is available. If a BAA is not available, PHI should not be processed through that tool.
Implement a firm-wide policy before the next matter opens. Establish clear written guidance on which tools are approved for handling client documents, and require that any AI platform processing PHI has a signed BAA on file.
Evaluate purpose-built legal AI platforms. Look specifically for platforms that offer BAAs, host data on enterprise-grade infrastructure, and maintain full audit logs. These are not optional features — they are the minimum threshold for responsible use.
Disclose appropriately to clients. As AI tools become more central to case preparation, client engagement letters and informed consent practices should reflect how client data is handled. Transparency is both ethically sound and practically protective.
The convenience of consumer AI is real. The compliance gap it creates when applied to PHI is equally real — and far more consequential. Small personal injury and employment law firms, which often handle the most sensitive health and workplace records, have the most to lose from treating general-purpose chatbots as professional-grade tools.
Building a practice on AI that was designed for the work attorneys actually do — with the data protections, citation integrity, and compliance infrastructure that professional responsibility demands — is not just a technology decision. It is a client protection decision. If your firm is ready to make that shift, learn more about ProvaLens and see what compliant, citation-grounded legal AI actually looks like in practice.