Can Personal Injury Lawyers Put Client Records Into ChatGPT?

A personal injury lawyer should not paste raw client records into ChatGPT. Use a business plan with no-training defaults, redact names, identifiers, medical and financial details first, keep chats as drafts rather than records, and put every output in front of a lawyer. Free and personal plans suit placeholders only.

Injury files are heavier than most: medical records, claim numbers, insurer correspondence, settlement figures, sometimes children's details. So the rule has to come from the firm owner, not from each person's judgement on a busy afternoon. The plan you buy settles who controls the training setting; it does nothing about what gets pasted, which is why the redaction list carries most of the weight. This is practical guidance, not legal advice.

Follow me on Instagram@sagnikteaches

Business plan versus Free: what actually differs

The plan decides who controls the data setting. On Free and personal paid plans, each person decides whether their chats are used to improve models (the "Improve the model for everyone" control under Data controls), and the control is personal, so it can be changed or forgotten. Temporary Chat keeps a conversation out of training unless the user saves it, and OpenAI may hold it for up to 30 days, but it is still an individual tool outside your control. OpenAI also says ads may appear on the Free and Go plans (not on Plus, Pro or Business), one more reason not to treat a free account as a firm tool.

Connect on LinkedInSagnik Bhattacharya

ChatGPT Business, on the other hand, does not use business content for training by default, costs $20 a seat a month on annual billing or $25 month to month (OpenAI's business pricing page), is sold for teams of 2 to 200 and gives the owner admin control of members. Data residency (choosing the region where content is stored) is listed for Enterprise, not Business. Claude Team (2 to 150 seats, no training on your content by default) and other business plans are comparable. The differences that matter for a firm:

Subscribe on YouTube@sagnikteaches

• Default, not preference. Training exclusion is the default for the workspace, not something each person has to remember.
• Admin control. The owner decides who has a seat and can remove people when they leave.
• Visibility limits. By default owners and admins do not see members' private chats, members have no self-service export of workspace data, and a removed member loses access immediately. So a workspace is not a filing cabinet.
• No output indemnity. The output indemnity in OpenAI's Service Terms, which covers intellectual-property claims, names API and Enterprise-type customers, not ChatGPT Business. Responsibility for what you file stays with the firm on every plan.

The honest summary: a business plan lowers one risk (training use and uncontrolled accounts). It does not lower the risk of sending sensitive data to a third party in the first place. That is what redaction is for. The settings that decide when a free ChatGPT account stops being safe for law-firm work are covered separately.

Sort your data into three tiers

People follow a short, memorable scheme, so give the firm three tiers rather than a long policy.

Red is not "never use AI"; it is "never use this tool for this material". Some injury-specific products are built for records, with contracts that cover them. Whether one fits is a separate procurement decision, covered in the buyer's checklist for personal injury case software with AI.

The redaction list for an injury file

Print this as a card. Each line names something that identifies a person or reveals more than the task needs.

• People: client, family members, witnesses, defendants, drivers, employers, treating clinicians, adjusters and opposing counsel.
• Identifiers: national ID numbers, driving licence and plate numbers, policy and claim numbers, medical record and account numbers, case and docket numbers.
• Places and dates: exact addresses, intersections, accident location, hospital or clinic names, dates of birth, accident date and treatment dates. Use "Day 0", "Day 12" and "Month 3".
• Medical specifics: diagnoses, medications, scan results, mental health or substance history, reproductive health. When the case turns on one, describe it in general terms ("a lower-limb fracture requiring surgery").
• Money: medical bills by provider, wage figures that identify an employer, offers, settlement authority, lien amounts.
• Images and files: scene photos, injury photos, dashcam footage, scans of any document. Text only.
• Metadata: file names, email addresses and document properties.

The check that catches the most: read the redacted text as a stranger. If the paragraph still describes a unique event at a specific time and place, such as a named employer's yard on a particular shift, it identifies the client even without a name. Generalise until it could describe a hundred matters.

Placeholders and a mapping sheet

Use stable labels so the draft can be restored: [CLIENT], [DEFENDANT 1], [CLINIC A], [DAY 0]. Keep the mapping table in your case management system, and do the restoring there. A redaction tool can speed up the first pass, but it misses nicknames, initials and addresses typed in odd formats, so a person reads the result before pasting. The tutorial on redacting personal data with AI, and where it misses covers the method and its limits.

Three safe uses once the data is redacted

1. Plain-English client updates

Clients ask the same questions: what happens next, why is it taking so long, what does an independent medical examination involve. Give the model a redacted status note and ask for a short, kind update at a reading age of twelve.

Write a 150-word update for a client in a personal injury matter. Use
only these facts: [CLIENT] has finished treatment block 1; we have
requested records from [CLINIC A] and [CLINIC B]; we expect the records
in about three weeks; the next step is to review them with [CLIENT].
Do not predict outcomes, values or timelines beyond those given.

Illustrative output: a warm, three-paragraph note that restates the facts and says what the client needs to do. What you would fix: the model may add "we are confident of a strong result", which is a prediction you must delete. Also check that it did not turn "about three weeks" into "within three weeks".

2. A chronology skeleton from redacted notes

A chronology table (date, event, source, issue) is time-consuming and repetitive. From a redacted event list, a model can sort entries, spot gaps and format a table. It cannot be trusted to read the records for you. A lawyer or paralegal still checks each row against the source, because models misdate entries and merge visits. For the full workflow, including where specialist tools do better, see can AI build a medical chronology for an injury case?

3. Question outlines and checklists

Draft deposition outlines, witness question lists and discovery checklists from a redacted summary of issues. Ask the model for "ten questions that probe the timeline on [ISSUE]" and then edit. The output is a prompt for your own thinking, and since it is built from placeholders, little can go wrong with the data. What you must guard against is a plausible question premised on a fact you never gave it.

These three have something in common: the model organises or phrases what you supply, and a lawyer owns every fact. Demand letters are riskier because they present damages and liability to the other side, which is why AI demand letters for personal injury: EvenUp, Supio or ChatGPT? compares purpose-built tools and general chat on that job.

What never goes into a general-purpose chat

1. Medical records, bills, imaging reports or any scan of them, whole or in excerpts, unredacted.
2. Insurer letters, claim numbers, policy limits, reserves and settlement authority.
3. Anything about a minor, a person lacking capacity or a client in a vulnerable position.
4. Privileged strategy memos and mediation or negotiation positions.
5. Photos of injuries, accident scenes or people.
6. Client contact details and intake forms.
7. Another party's records produced in discovery, which may carry confidentiality orders.
8. Anything a protective order or sealed filing covers.

Item seven is easy to forget. Records produced by the other side may be subject to a confidentiality agreement or court order that restricts sharing with third parties, and an AI service is a third party. Check the order before any AI use on discovery material.

A small firm's rollout in four weeks

An invented firm: three lawyers, two case managers, one receptionist. The owner starts on monthly billing to keep the exit cheap: five seats (receptionist excluded) at $25 each, $125 a month, moving to annual billing at $20 a seat only if the first quarterly check shows the rules are holding.

• Week one: the owner sorts the firm's ten most common documents into tiers on a single page and writes the redaction card and never-paste list. Two hours.
• Week two: a one-hour briefing with a live exercise: each person redacts the same invented intake note, and the group compares results. People find what they missed within minutes.
• Week three: three approved uses go live (client updates, chronology skeletons, question outlines), each with a prompt saved in a shared document.
• Week four: the owner reviews a sample of ten drafts: was the redaction clean, was a lawyer's review recorded, did the model add anything unsupported? Gaps become rule changes.

If the three lawyers save two hours a month each on outlines and the two case managers four hours each on client updates, that is fourteen hours for $125 plus review time. The numbers are illustrative; what you are really buying is a controlled place for AI use instead of five uncontrolled ones.

A redaction example from an injury note

An invented intake note and the version that may go into a business workspace:

BEFORE (never pasted)
Jane Roe (DOB 12/03/1988) was rear-ended on 14 June at the junction
of Station Road and Main Street while driving a blue hatchback to work at
Example Haulage Ltd. Treated at City General Hospital for whiplash and a
wrist fracture; Dr John Doe referred her to physiotherapy. Claim no. 55-2210.

AFTER (may be pasted)
[CLIENT] was rear-ended at a junction on Day 0 while driving to work.
Treated at [HOSPITAL A] for a neck injury and a wrist fracture; a
clinician referred her to physiotherapy from Day 9. Claim reference
removed.

The after version keeps the injury types in general terms, turns dates into relative days and removes the employer, vehicle and location, none of which a client update needs. Keep the first version in the case system where it belongs.

Sorting ten common documents: a filled-in example

Ten rows, four red, two conditional. A new starter can learn this in ten minutes, and the table is far easier to follow than a policy document.

Questions to ask a vendor that does handle records

Some injury-specific tools exist to read records, build chronologies or draft demands, and their contracts are written for that. Ask each one, in writing: where data is stored and for how long; whether uploads are used to train models; who at the vendor can see them; whether you can delete a matter's data on request and get confirmation; what happens at contract end; and what a lawyer must still check. Prices for several leading tools are not published, so expect a sales conversation. Do not let a demo with the vendor's sample records stand in for a test on a redacted real matter.

A realistic slip and how it showed up

A case manager under deadline pasted a full hospital discharge summary into a personal account to "make it readable for the client". The account used default settings, so it was set to improve models. Nobody noticed until the supervising lawyer found the chat in a browser history during an unrelated check. The firm had no business seats, no redaction card and no rule on personal accounts. The response was the incident routine: record, delete, tell the supervising lawyer, decide on client notification, then fix the cause by buying business seats and issuing the card. The lesson for an owner: the absence of a safe tool is itself a risk, because staff will reach for an unsafe one.

When a client asks whether you use AI

Clients increasingly ask, and the answer should be calm and specific. Illustrative wording: "We use a business-grade tool for routine drafting, such as plain-English updates. We remove your name and details first, and a lawyer checks everything. We do not put your medical records into it." Say only what is true, and make sure every person who may be asked has the same answer. If a client objects, record the instruction in the matter file and respect it.

A redaction drill on an invented crash file

Rules stick when people practise them on material that looks like the real thing. Build one invented crash file: an intake note like the one above, a two-page discharge summary, an insurer's acknowledgement letter and a short witness account, all with made-up names, plates, claim numbers and dates. Give each person twenty minutes to produce the version that may be pasted, then swap and mark each other's work. The usual misses are a vehicle registration inside a witness account, a hospital ward name that narrows the place, a treating clinician's surname in a referral line, and an exact date left in a heading. Score one point per miss and talk through the top three. Then spend ten minutes on the never-paste list and the incident routine, and collect signed acknowledgements. Run the drill again for every new starter in week one with a fresh invented file, and keep the signed sheets with your compliance records.

Retention, departures and the record

Because owners cannot see members' private chats by default, members cannot export their own workspace data and a removed member loses access immediately, use these habits. Save the output you keep into the case system as you go. Delete working chats at the end of each matter stage. Make "copy any draft you still need into the case system" part of the leaving checklist, before the seat is removed. Never rely on chat history as evidence of what the firm did.

The record that matters is yours: the redacted prompt (or a note that one was used), the final document, who reviewed it and when. Many firms add a single line to the matter file: "AI-assisted draft: redacted, reviewed by [initials], [date]". It takes ten seconds and answers most questions later.

Engagement letter wording

Our firm uses business-grade AI tools for routine drafting, such as
plain-English updates and checklists. Your medical records, photographs
and insurer documents are never put into these tools. Before any detail
of your matter is used, we remove your name and anything that identifies
you, and a lawyer checks every document before it is sent or filed. If
you would prefer us not to use these tools on your matter, tell us and
we will record that instruction on your file.

Only include a sentence if it is true in practice. If someone uses a personal account for convenience, the paragraph is a promise the firm breaks every day.

Where the model fails in injury work

Models write confident medical and legal prose, which is exactly what makes errors hard to see. Typical failures: inventing a causation link the records do not support, mixing up left and right or before and after, summarising a normal result as an abnormal one, citing cases that do not exist and softening "the claimant reported" into "the claimant suffered". None of these is rare. They are why the three safe uses are organisational and why a lawyer reads every factual statement against its source. A good test of your process: seed one deliberate error into a practice draft and see whether the reviewer finds it.

A quarterly check the owner can do in an hour

Every three months, sample ten AI-assisted drafts. Check: redaction clean, reviewer recorded, no unsupported claims, nothing in the never-paste list, and a seat list that matches current staff. Look at the chats you can see for stray identifiers. Ask staff which rule was hardest to follow and fix that one. Adjust the tiers if a new document type has appeared, and update the engagement wording if your practice changes. For prompt wording on the amber tier, ChatGPT prompts for lawyers: client updates and confidentiality red lines is a useful companion.

Drafting the tiers, the plan settings and a rule sheet the team can run is the kind of job my AI implementation consultation covers; either way, start by sorting your ten most common documents this week.

Questions a firm owner should answer before rollout

Is ChatGPT Business safe enough for medical records?

It is the better option because it does not train on your content by default and gives admin controls, but safe enough depends on your professional duties, your client agreements and the vendor's terms. Medical records are among the most sensitive material you hold, so redact first and ask your insurer or ethics adviser what they expect.

What is the actual difference between Free and Business?

On Free and other personal plans, whether chats are used to improve models is a user setting that can be changed or forgotten. On Business, content is not used for training by default and the firm controls members. Neither plan changes your duty to protect client confidentiality, and neither replaces redaction.

Can we upload a client's medical records as files?

Not unredacted. Upload creates a copy of the record in a third-party service. Either redact it first, or use a tool built for legal and medical records with a contract that covers records handling, such as the injury-specific products discussed in related tutorials. Check the vendor's terms before you decide.

Do clients need to consent to AI use?

Many firms add a short engagement-letter paragraph about AI-assisted drafting, and some professional guidance points towards informing clients when their confidential information may be shared. Check the rules that bind you. If your process redacts everything identifying, say so plainly in the same paragraph.

What if a staff member already pasted records into a personal account?

Treat it as an incident. Record what was pasted and where, delete the chat, tell the supervising lawyer the same day and decide whether the client must be told under your obligations. Then find out why it happened, for example no business seats, and fix that before blaming the person.

Further reads

• AI for Personal Injury Lawyers: Seven Jobs to Hand Over First — Seven jobs worth handing to AI first.
• Personal Injury Lawyer Chatbot: Intake Questions, Limits and Cost — Intake chatbots: questions, limits and cost.
• Can Immigration Lawyers Use ChatGPT? Evidence Letters and Redaction — The same rule-setting for an immigration practice.
• Claude v ChatGPT for Lawyers: A Same-Task Test on Your Own Documents — Compare assistants on your own documents.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI help articles on data controls, ChatGPT Business and ads, OpenAI business pricing page and Service Terms, Claude pricing page, injury-practice AI vendor sites; professional rules vary and should be checked with your regulator or bar body.