Free ChatGPT for Lawyers: Where the Free Plan Stops Being Safe

Only for work with no client-identifiable or confidential content. On Free, ChatGPT may use your chats to improve models unless you switch that off, may show ads, and may keep Temporary Chats up to 30 days. Once client facts, documents or strategy are involved, move to ChatGPT Business or another plan with no training by default, plus your firm's approval.

The setting people point to, "Improve the model for everyone", is only one exposure among several. Ads, retention, shared logins and the absence of any firm-level controls matter as much, and a toggle fixes none of them. Paying for an individual plan doesn't fix it either: OpenAI says data sharing is enabled by default on personal workspaces for Free, Plus and Pro. What makes free ChatGPT usable for a lawyer is a rule about what goes in, not a setting about what happens afterwards.

Follow me on Instagram@sagnikteaches

What the Free plan does with what you type

Everything in this list comes from OpenAI's own pages, read on 1 November 2026: the data controls FAQ, the article on ads in ChatGPT, and the help pages on deleting chats and on file uploads.

Connect on LinkedInSagnik Bhattacharya

• Training default. On personal workspaces with ChatGPT Free, Plus or Pro, data sharing is enabled by default. You can switch off "Improve the model for everyone" under Data controls and keep your chat history. OpenAI's pricing page lists the same pattern, with opt-out available, for Go.
• Temporary Chat. Not used to improve OpenAI models, but "may be retained for up to 30 days for safety purposes". Saving a temporary chat turns it into a normal chat that follows your usual settings.
• Deleting a chat. OpenAI schedules it for permanent deletion within 30 days. That doesn't apply to a chat that was already de-identified, or one OpenAI must keep longer for security or legal obligations.
• Ads. Ads may appear for users on the Free and Go plans. Advertisers receive only aggregated, non-identifying performance information, not your chats or memories. Temporary Chats show no ads, and ads aren't eligible to appear near sensitive topics such as personal health, mental health and politics. Plus, Pro, Business, Enterprise and Edu accounts have no ads.
• The ad-free option. A Free user can switch to an ad-free experience, which OpenAI says removes ads but comes with lower usage limits and no access to some tools, such as image generation.
• Uploads. Free users are limited to 3 file uploads a day, and OpenAI may lower limits at peak times.

None of that is sinister. It is a consumer product doing what consumer products do. The point for a lawyer is that every line describes a relationship between OpenAI and one individual, with no agreement between OpenAI and your firm, no admin who can see what staff paste, and no way to demonstrate to a client what happened to their information.

Subscribe on YouTube@sagnikteaches

Six triggers that end free use for a matter

"It depends" is the honest answer to whether free ChatGPT is safe, but it depends on things you can check in seconds. If any of these six is true, the matter has left the free plan.

1. A client name, matter number or identifiable fact is in the prompt. The threshold is any one. Facts that identify only in combination count too.
2. A client document is attached. Free allows 3 uploads a day, which is a useful accident, but the rule is about the document, not the quota.
3. More than one person uses the account. A shared login means no audit trail, no way to remove one person's access and, on some services, a breach of the terms.
4. The output will be filed, sent or relied on. A draft that goes into a client letter or a court document needs a controlled process behind it, which a personal account can't evidence.
5. The engagement terms or the client restrict third-party processing. Some commercial clients ask about AI in their panel terms or outside-counsel guidelines. Check before you paste, not after.
6. You could not explain to the client where it went. If the honest answer to "who else could see that?" is "I'm not sure", the answer to "may I paste it?" is no.

The sixth is the most practical test because it works for situations nobody has written a rule for. It also explains why a business plan matters beyond the training toggle: the value is an agreement and controls you can describe to a client, not the absence of one setting.

Three kinds of task and the plan each needs

Sorting the work by what it contains settles most cases. Here is the sort, with examples a small firm will recognise.

Most of the value a small firm gets from AI in its first month sits in the first two rows, which is why free ChatGPT is a reasonable place to start. Drafting a letter structure and then adding the facts yourself, in your own document, uses the tool for what it is good at and keeps the sensitive part on your side. The prompts in ChatGPT prompts for lawyers: client updates and confidentiality red lines follow the same split.

A free-safe task, start to finish

Here is a task that sits squarely in the first row of the table: a plain-English explainer for a client newsletter. Nothing in it comes from a matter.

In 200 words, for small-business owners, explain what a limitation
of liability clause does in a services contract and give three
questions to ask before signing one. Plain English, no legal
citations, and say clearly that enforceability depends on the exact
wording and the rules that apply.

Imagine the reply as a clear paragraph on what the clause does (it caps what one party can owe the other, and sometimes excludes certain kinds of loss), then three questions: what the cap is and how it is set, which losses are excluded, and whether the cap applies equally to both sides. What the lawyer fixes: the draft may imply that such clauses are routinely enforced as written, which overstates the position, so you restore the caveat in your own words. It may also call a cap "standard" with no basis; delete the word. Ten minutes of editing turns a competent generic draft into a newsletter piece that carries your name and your judgement.

Anonymising a matter: a filled-in before and after

Take an illustrative payment dispute. The raw instruction note names the client company and its director, the customer, the product, three invoice numbers totalling $184,000, the contract date, the delivery month and the customer's allegation of defects. Every one of those is either an identifier or a fact that points to one. Here is the same matter after anonymising by hand in a text editor, with find-and-replace and a key sheet that stays offline.

The prompt then asks for structure, with the specifics left as placeholders for you to fill in offline.

Draft the structure of a letter before action in a payment dispute.
Party A supplied goods to Party B under a written contract. Party B
withheld payment of a significant sum, alleging defects in one
delivery. Party A disputes the defects and says the contract required
written notice within a set period.
Include: summary of facts, contractual basis, the demand, a deadline,
consequences of non-payment, and a list of facts I must verify.
Use [placeholders] for every specific. Do not cite any law or case.

The skeleton that comes back (illustrative) has six parts with bracketed placeholders and a short "facts to verify" list covering the notice date, the delivery records, the invoice dates and whether any earlier correspondence admitted the defects. Expect to edit the tone: the tool may add one that is more aggressive than you want, and it may suggest a deadline that has no basis in your rules. Pre-action steps and time limits are exactly the kind of rule it can get wrong, so check them against your own jurisdiction's requirements before anything goes out. The instruction not to cite law is deliberate; a skeleton with no citations has nothing to invent.

Then the re-identification test: could opposing counsel, or a colleague who knows the client, guess which matter this paragraph describes? A dispute over a particular niche product in a small market can identify itself without a name, so cut the product and keep the structure. For client-record-heavy work such as injury claims, the redaction rules are stricter; can personal injury lawyers put client records into ChatGPT covers that end.

What a business plan changes, and what it doesn't

ChatGPT Business is the plan OpenAI sells to small teams. On its pricing page OpenAI lists the Standard seat at $20 per user a month billed annually or $25 billed monthly, with a Premium seat at $100 or $125 for higher usage with no 5-hour limit, for teams of 2 to 200. The page states "No training on your business data by default", lists SAML single sign-on and multi-factor authentication, and says Business has completed a SOC 2 Type 2 audit. Data residency, SCIM provisioning, enterprise key management, IP allowlisting and the compliance API are listed for Enterprise only. OpenAI's enterprise privacy page adds that you own your inputs and outputs where the law allows it, and that data is encrypted at rest and in transit.

That is a real difference: an agreement about how your content is handled, an admin who can manage seats, and a workspace you can describe to a client. The cost and settings for a law firm are worked through in ChatGPT Business for law firms: seats, privacy settings and cost. What it doesn't change matters just as much.

• Accuracy. Business ChatGPT can still invent a case. Every citation needs the same check as on Free.
• Your duties. Confidentiality, competence and supervision obligations sit with the lawyer, not the vendor. A business plan makes compliance easier to show; it doesn't discharge anything.
• Client consent and engagement terms. If a client forbids third-party AI processing, a better plan doesn't override it.
• Retention you don't control. Check what the workspace settings say about departed members' chats and about retention before you rely on deletion.
• Privilege questions. Whether using any vendor affects confidentiality or privilege is a matter for your professional body's guidance. Ask your insurer or that body's guidance service rather than assuming.

Business isn't the only route to no-training-by-default. Anthropic's pricing page lists Claude Team at $20 per seat annual or $25 monthly for 2 to 150 seats, with training "none by default". Microsoft says Copilot Chat is included at no extra cost with eligible Microsoft 365 plans, with enterprise data protection for work accounts and prompts not used to train foundation models. Google says Workspace data isn't reviewed by humans or used for generative AI training outside your domain without permission. If your firm already runs Microsoft 365 or Google Workspace, check those first: the tool you need may be switched off rather than unbought. A purpose-built legal assistant is a further option, compared in ChatGPT or a legal AI tool for a small firm.

Three practices, three different answers

The same rules give different outcomes depending on who is typing. Three illustrative setups show where the line falls.

A sole practitioner in family work

Free ChatGPT is a sensible tool here for plain-English handouts, a newsletter, anonymised letter skeletons and a checklist of questions for a first interview written from a fictional fact pattern. It stops the moment a prompt needs a client's circumstances, because family matters are the clearest case of sensitive information about third parties, including children. The trigger to pay is when more than a few matters a week need real facts in the draft. A sole practitioner can't buy ChatGPT Business alone, because it starts at two seats, so the choice is $40 a month at the annual price, an alternative with no training by default, or the work-account Copilot Chat if the practice runs on Microsoft 365.

A three-lawyer commercial firm

Here the free plan is already a risk, because three people means three personal accounts and no oversight. Three Business Standard seats cost $60 a month on annual billing, $720 a year, or $75 a month on monthly billing, $900 a year. As a sum, if each lawyer saves two hours a month and an hour is worth $150, that is $300 per lawyer, $900 in total, against a $60 cost. The sum is illustrative and no promise, but it shows why the argument about price is the wrong argument. The real work is the firm policy and the training of staff, which is why the free accounts should be closed once the paid ones exist.

In-house counsel at a 40-person company

The "client" is the employer, and the employer usually has an IT policy that already answers the question. A personal free account is often against company policy for confidential material, and the practical route is whatever the company has approved, often Copilot Chat or a workspace plan. Where nothing is approved, the in-house lawyer's best contribution is a short request to the business: pick a tool with no training by default and write the rule. The shadow-AI tutorial on stopping staff pasting client data into free tools is written for exactly that conversation.

What the three have in common is that the answer comes from the contents of the prompt and the number of people, not from the price of the plan. A one-person practice with only generic work can stay free indefinitely. A three-person firm with client facts in play shouldn't be on free accounts at all.

Four leaks people forget, from memory to shared screens

A firm's first incident is rarely a deliberate paste of a client file. It is usually one of these.

• Memory and personalisation. Features that remember details between chats can carry a fact you anonymised in one conversation into another. For any account used near client work, check the personalisation settings and switch memory off, or keep client-adjacent work in a Temporary Chat started as "Unpersonalized", because OpenAI says a personalised temporary chat can still use your saved memories.
• Screenshots. A screenshot of an email chain is an upload with every name in the signatures and the quoted history. Retype the one question you need instead.
• Dictation and voice. Spoken matter detail goes to a vendor the same way typed text does, and a casual "just tell it the story" is where anonymising discipline usually fails. Anonymise in your head before you speak, not after.
• Shared screens. A free chat window with ads in it, shared on a video call with a client, tells the client what you use without telling them what you do with it. Close the window before the call, or use the approved tool.

Citations: the check that no plan removes

Whatever you pay, the model can produce a case name, a court and a citation that look right and don't exist. The routine below takes about five minutes per authority and works the same on every plan.

1. Search the authority in an official or subscribed database by name and by citation, not by asking the chatbot again.
2. Open it and read the passage the tool relied on. Does it say what the draft says it says?
3. Check that it is still good law and that the court and date are right.
4. Delete anything you couldn't find. An authority that "should exist" is a warning, not a gap to fill.

An illustration of the failure: you ask for authority on a point and the tool offers three cases. Two are found and say roughly what was claimed. The third has a plausible name, a believable year and a citation format that matches the series, but a search finds nothing. The third looks exactly like the first two, which is the whole problem, and it is why a lawyer who files an unchecked citation has no defence in saying the tool sounded sure. The safer prompt pattern is to ask the tool for the issues and arguments only, and to find the authorities yourself.

A ten-second test before you hit send

Print this and tape it where the screen is.

1. Could anyone identify the client, the counterparty or the matter from this text?
2. Am I attaching a document that belongs to a client?
3. Is this a free personal account?
4. Will the output be sent, filed or relied on without a line-by-line check by a lawyer?

If the answer to the first, second or fourth is yes, or the answer to the third is yes alongside any of them, stop. Anonymise further, move to the approved tool, or do the task by hand.

One paragraph for your firm policy

A written rule is cheaper than an incident. This wording is a starting point to adapt with your own adviser.

Free or personal AI accounts may be used only for work that contains no client-identifiable or confidential information, and never with attached client documents. Anonymised drafting is allowed in Temporary Chat with model improvement switched off. Client-identifiable work may be done only in the firm-approved AI workspace. Every AI-assisted draft is checked line by line by a lawyer before use, and every citation is verified in an official or subscribed source.

If you'd rather draft the policy and map the tasks with someone, that is what my AI implementation consultation covers. For the confidentiality question in more depth, can solicitors use ChatGPT without breaching confidentiality is the best next read.

Free ChatGPT for lawyers: quick answers

Does switching off model improvement make the Free plan safe?

It stops OpenAI using your content to improve models, which is the biggest single exposure, but it removes nothing else. Chats are still stored, may be retained for safety or legal reasons, and sit in a consumer account with no firm-level controls. Treat the switch as necessary and insufficient, and anonymise as well.

Does a Temporary Chat protect privilege?

A Temporary Chat isn't used to improve models and shows no ads, but OpenAI says it may be kept up to 30 days for safety. Whether sharing client information with a vendor affects confidentiality or privilege is a question for your professional body's guidance, not a setting. Use Temporary Chat for anonymised drafting, never for client facts.

Is ChatGPT Plus enough for a solo lawyer?

Plus removes ads, but OpenAI says data sharing is on by default for Free, Plus and Pro personal workspaces, so the training default is the same. A solo lawyer who needs client-facing use should look at ChatGPT Business, which needs two seats at $20 each a month on annual billing, or at another tool with no training by default.

Can I rely on the citations ChatGPT gives me?

No. Treat every citation as unverified until you have found the authority in an official or subscribed database and read the passage. Lawyers have been sanctioned for filing cases an AI invented. The duty to check is identical on every plan; a free plan just gives you fewer controls and fewer ways to prove how you checked.

Further reads

• Claude v ChatGPT for Lawyers: A Same-Task Test on Your Own Documents — A same-task test of Claude and ChatGPT on your own documents.
• Best AI for Solo Lawyers: A Stack That Fits One Person — An AI stack sized for one lawyer.
• Free ChatGPT for Accountants: When Client Work Needs a Paid Plan — The same free-plan question asked for accountants.
• Is Free ChatGPT Good Enough to Run a Small Business On? — Whether free ChatGPT can run a small business.
• Per-Seat or Pay-As-You-Go? Legal AI Pricing for Small Firms — Per-seat or pay-as-you-go legal AI pricing for small firms.
• Can Lawyers Use AI Note Takers in Client Meetings? — Whether AI note-takers belong in client meetings.
• Claude or ChatGPT for a Law Firm: Which One to Pay For — Which assistant a law firm should pay for.
• Which Legal Tasks Should a Small Firm Never Hand to AI? — Legal tasks a small firm should never hand to AI.
• Best AI for Australian Lawyers: Tools, Privilege and Practice Rules — Legal AI against general assistants for an Australian firm, with a privilege test, a court-by-court map of AI rules and a supervision routine.
• Best AI for Immigration Lawyers: Intake to Filing, Stage by Stage — A nine-stage map of an immigration case with the AI tool type, the data check and the manual step for each, plus a costed worked example.
• AI for Immigration Lawyers: Docketwise or ChatGPT for Forms? — Run one immigration form through Docketwise and ChatGPT: what each catches, the cost per filing, the privacy settings and the steps that stay manual.
• AI Tools for Lawyers in Australia: A Safe Starting Checklist — A checklist for Australian law practices: client data, checking output, supervision, billing and court AI rules, with a filled-in one-page policy.
• AI for Bankruptcy Lawyers: Intake, Schedules and Document Chasing — A document-to-draft route for consumer bankruptcy matters, with missing-evidence checks before any client signs.
• Best AI for Lawyers in Canada: Privacy, Tools and a Safe First Task — ChatGPT, Claude, Copilot and Clio Manage AI compared by task for a small Canadian firm, with confidentiality checks and a low-risk first job.
• AI for Criminal Defence Lawyers: What Is Safe With Client Confidences — A firm-level client-data policy checklist for criminal defence lawyers: safe and unsafe AI tasks, plan settings, discovery handling and a first pilot.
• Can Immigration Lawyers Use ChatGPT? Evidence Letters and Redaction — A firm-owner policy for ChatGPT in immigration practice: which plan, what to redact, how support letters get drafted and who signs them off.
• Using AI in Law: Which Tasks a Small Practice Should Start With — A risk-ranked list of what a small law practice can hand to AI first, with supervision and confidentiality rules for each rung.
• Which AI Is Best for Lawyers in Australia? Choose by Task — A decision rule for choosing AI by task for an Australian law practice, comparing general assistants with the AI built into practice software.
• 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, ads in ChatGPT, file uploads and deleting chats; OpenAI ChatGPT Business pricing and enterprise privacy pages; Anthropic pricing page; Microsoft Learn Copilot Chat page; Google Workspace AI privacy page, read on 1 November 2026 and checked again on 4 November 2026; the June 2023 opinion and order on sanctions in Mata v. Avianca.