AI Agents for Lawyers

Playbooks

AI Privacy Settings for Lawyers: Step-by-Step in ChatGPT, Claude, Gemini, Copilot and Perplexity

Exact steps to turn off training and check retention in ChatGPT, Claude, Gemini, Copilot and Perplexity as of September 2026, plus admin settings and a log.

By Dr. Niklas SchmidtUpdated 13 November 202610 min read
On this page
  1. The tier you are on matters more than any toggle
  2. ChatGPT: the Data Controls toggle and what it leaves behind
  3. Claude: the five-year choice and the Covered Models exception
  4. Gemini: Keep Activity, the 72-hour tail and the human reviewers
  5. Copilot and Perplexity: check the product, then the toggle
  6. What admins should set on Team, Enterprise and Workspace
  7. The monthly re-check: settings move, so check them
  8. What to do this week

Consumer tiers of ChatGPT, Claude, Gemini and Perplexity train on your conversations unless you switch it off; business tiers, the APIs and Microsoft 365 Copilot do not train by default. The switch takes under a minute per tool, but it changes neither retention nor discoverability, so the real protection is the plan you are on. Here are the exact settings, tool by tool, as of September 2026.

The tier you are on matters more than any toggle

Across vendors the rule of thumb, verified on 1 September 2026 by Sonomos, is simple: consumer tiers train unless you opt out; business and API tiers do not; and OpenAI reportedly retains conversations for 30 days for abuse monitoring even after opt-out. The table shows what that means per tool.

ToolConsumer tier defaultBusiness or enterprise tierAfter opting outZero data retention
ChatGPTFree, Go, Plus and Pro train unless opted outBusiness, Enterprise, Edu and API do not train by defaultRetained 30 days for abuse monitoring (reported)Contractual, for eligible API and enterprise customers
ClaudeFree, Pro and Max users must choose; opted-in chats kept up to 5 yearsTeam, Enterprise and API never used for trainingKept 30 days; flagged content up to 2 yearsPer organisation via sales; not available for Team or Enterprise chat; Fable 5.1 and Mythos excluded
GeminiKeep Activity on: 18 months' retention, human review, trainingWorkspace and Vertex excluded by default through the admin consoleChats still kept 72 hours; human-reviewed chats up to 3 yearsEnterprise contract terms
CopilotConsumer Copilot: consumer terms, no DPAMicrosoft 365 Copilot: DPA, Enterprise Data Protection, Purview; no training on promptsNot applicableGoverned by the DPA rather than a toggle
Perplexity"AI data retention" toggle in PreferencesEnterprise data never used for training by defaultNot published in the sources reviewedNot published in the sources reviewed

Why this is worth an hour of a lawyer's time: in United States v. Heppner (S.D.N.Y., opinion 17 February 2026) Judge Rakoff held that a defendant's research memos written with consumer Claude were neither privileged nor work product, "because Claude is not an attorney" and Anthropic's terms permitted logging and disclosure (Harvard Law Review Blog). In the OpenAI copyright litigation, Judge Stein affirmed on 5 January 2026 an order to produce 20 million de-identified ChatGPT logs, reasoning that users had "voluntarily submitted their communications" to OpenAI (National Law Review). The Upper Tribunal in UK v SSHD [2026] UKUT 81 (IAC) said pasting client letters into public AI tools put them "in the public domain". The German BRAK treats tools that store or forward client data for training as a breach of § 43a Abs. 2 BRAO; the Austrian ÖRAK's practice rules include "no client data for training"; ABA Formal Opinion 512 requires informed consent where a self-learning tool may expose client data. The privilege analysis is in attorney–client privilege and AI; this page is about the settings.

ChatGPT: the Data Controls toggle and what it leaves behind

The path below is as verified by Malwarebytes on 15 September 2026 and OpenAI's own Data Controls FAQ.

  1. Click your profile name at the bottom-left of the desktop app or web interface, then Settings.
  2. Open Data Controls.
  3. Switch off Improve the model for everyone. The setting applies to the whole account but only to new chats; anything already used for training stays used.
  4. Confirm your plan. Free, Go (about $8, regional), Plus ($20), Pro ($100 or $200) are consumer plans and train unless opted out. Business ($25 per seat monthly, $20 annual), Enterprise and Edu do not train by default, and the API retains for 30 days for abuse monitoring (OpenAI enterprise privacy).
  5. If memory is enabled, the model carries facts between chats. For matter work, clear any entry that names a client or a counterparty, or switch memory off for that account.

What the toggle does not do: it does not shorten retention, and it does not make chats undiscoverable. The 20 million logs ordered produced in the OpenAI litigation were 0.5% of tens of billions preserved. For discovery purposes a lawyer on a Plus account is a member of the public who typed into a website. More on OpenAI's products in ChatGPT for lawyers.

Claude: the five-year choice and the Covered Models exception

Anthropic introduced a training choice for consumer accounts on 28 August 2025, with the deadline to decide extended to 8 October 2025 (Anthropic). The consequences of that choice are larger than most users registered.

  1. Click your name at the bottom-left, then Settings, then Privacy.
  2. Switch off Help improve Claude (labelled Model improvement in some builds).
  3. Note the retention consequence: with the setting on, chats are kept up to five years; with it off, 30 days. Incognito chats are never used for training (Anthropic privacy centre).
  4. Confirm your plan. Pro ($20, $17 annual) and Max ($100 or $200) are consumer plans and carry the toggle. Team ($25 standard or $125 premium per seat, five-seat minimum), Enterprise and the API are never used for training.
  5. If your firm uses the API or a vendor built on it, ask which model. Fable 5.1 and Mythos are "Covered Models": they require 30-day retention and are not eligible for zero data retention without express authorisation. ZDR is arranged per organisation; Team and Enterprise chat interfaces are not eligible, although Claude Code under Enterprise is. Flagged content may be kept up to two years, and the Compliance API feed and local session transcripts for Cowork and Claude Code are retained six years by default (Anthropic data retention docs).
  6. If anyone uses Claude for Chrome, the admin can allowlist and blocklist sites (Claude for Chrome); as of 18 September 2026 the Compliance API reportedly returns Chrome session transcripts for Enterprise accounts, so a browser agent's activity is now part of the firm's record.

Step 5 deserves a second look: the strongest model in the line-up is the one a firm cannot run under ZDR, and the newest surface (Cowork, merged into the Claude app on 17 September 2026) carries the longest default retention. Pick the model and surface deliberately. The line-up is reviewed in Claude for lawyers.

Gemini: Keep Activity, the 72-hour tail and the human reviewers

Google's consumer terms are the most explicit of the five. The Gemini Apps Privacy Hub, updated 10 August 2026, states that a subset of chats is read by human reviewers and adds: "Please don't enter confidential information that you wouldn't want a reviewer to see."

  1. Go to myactivity.google.com/product/gemini, or open Activity from the Gemini menu.
  2. Turn Keep Activity off. The default is on, with activity retained 18 months and used for training.
  3. Understand the tail: with Keep Activity off, chats are still kept 72 hours; any chat already sampled for human review is retained up to three years.
  4. Check the account type. Google AI Plus, Pro and Ultra subscriptions are available only for personal accounts, so paying $20 a month does not move you onto business terms. Workspace customers need a Gemini add-on; Workspace and Vertex data are excluded from training by default through the admin console.
  5. For Gemini Notebook (the product formerly called NotebookLM, renamed 16 July 2026), use it only under a Workspace or enterprise account. NotebookLM Enterprise has had EU data residency since 3 June 2026.

The tempting shortcut, a personal Gemini account for "just the public documents", fails on the 72-hour tail: one pasted paragraph from a client email is outside your control for days at minimum. Workspace setups are in Gemini for lawyers.

Copilot and Perplexity: check the product, then the toggle

Microsoft ships two things called Copilot. Consumer Copilot (the free app, Copilot Pro and the version inside Windows) runs under consumer terms without a data processing agreement and has been described by Spellbook as "not suitable for confidential or legal work". Microsoft 365 Copilot, the $30 per user per month add-on on an annual commitment, runs under Microsoft's DPA with Enterprise Data Protection and Purview controls and does not train on prompts (Spellbook, 20 June 2026). The check is the account: if you are signed in with a personal Microsoft account, you are on consumer terms, whatever the icon looks like. Two settings sit with the admin rather than the lawyer: the Legal Agent for Word requires Anthropic to be enabled as a subprocessor in the tenant, and Copilot Studio agents are governed through DLP over connectors and Purview audit. Details in Microsoft 365 Copilot for lawyers.

Perplexity has one consumer toggle: Settings, then Preferences, then switch off AI data retention. Pro ($20) and Max ($200) are consumer plans; Enterprise Pro ($40 per seat) and Enterprise Max ($325 per seat) carry the default that enterprise data is never used for training. Perplexity Computer, the long-running agent launched in February 2026, is Max-only, so an individual lawyer using it is on a consumer account. See Perplexity for lawyers for what the research product does well.

What admins should set on Team, Enterprise and Workspace

Individual toggles protect one account. Firm-level protection comes from the admin console and the contract, and it is where most of the 2026 changes landed.

PlatformSetting or decisionWhy it matters
ChatGPT EnterpriseGPT-6 Astra is off by default until an admin enables itAstra is the computer-use model; enabling it is a decision about agents acting on firm systems, not a model upgrade
ChatGPT Business/EnterpriseCompliance API audit trails for ChatGPT WorkChatGPT Work runs "independently for hours"; the audit trail is the only record of what it touched
Claude Team/EnterpriseModel selection; Claude for Chrome allowlist and blocklist; Compliance API retention (six years by default)Fable 5.1 and Mythos cannot be run under ZDR; Chrome transcripts now flow into the compliance feed
Claude EnterpriseRequest ZDR for Claude Code if developers use itThe chat interface is not ZDR-eligible; Claude Code under Enterprise is
Google WorkspaceGemini add-on rather than personal AI Pro/Ultra; confirm training exclusion in the admin console; NotebookLM Enterprise for EU residencyPersonal Google AI subscriptions never inherit Workspace terms
Microsoft 365Anthropic subprocessor toggle (Legal Agent for Word); Copilot Studio DLP; Purview audit; Agent 365 ($15 per user per month or in E7) for agent inventoryAgent 365 detects local agents, including Claude Code, that lawyers may have installed themselves
Perplexity EnterpriseConfirm the no-training default in the contractConsumer Max accounts are the route to Perplexity Computer, which admins cannot see
All vendorsArt. 28 GDPR data processing agreement; in Germany a § 43e BRAO contractA three-lawyer German firm's self-hosted setup still needed both, per the Lyron case study

Vendor questions to ask before signing are in AI vendor due diligence for law firms, and the GDPR analysis in GDPR and AI tools for lawyers.

The monthly re-check: settings move, so check them

This page is dated because the vendors keep moving the furniture. Anthropic changed its consumer training default on 28 August 2025; Google updated the Gemini Privacy Hub on 10 August 2026; OpenAI reportedly removed the standalone ChatGPT agent in early August 2026 without notice (single source); Anthropic merged Cowork into the Claude app on 17 September 2026. A security vendor now re-verifies opt-out paths as routine. A firm should do the same: a 15-minute slot on the first working day of each month, with a log kept in the DMS next to the AI policy.

AI SETTINGS LOG — [Firm] — checked by [name] on [date]

Tool / plan                         Setting                              Expected     Found    Action
ChatGPT [Plus/Business/Enterprise]  Improve the model for everyone       OFF          [ ]      [ ]
ChatGPT [Enterprise]                GPT-6 Astra enabled for users        [OFF/ON]     [ ]      [ ]
Claude  [Pro/Max/Team/Enterprise]   Help improve Claude                  OFF          [ ]      [ ]
Claude  [Team/Enterprise]           Default model (Covered Model?)       [model]      [ ]      [ ]
Claude  [Enterprise]                Chrome allowlist current             YES          [ ]      [ ]
Gemini  [personal/Workspace]        Keep Activity                        OFF          [ ]      [ ]
Gemini  [Workspace]                 Admin console: training excluded     YES          [ ]      [ ]
Copilot [M365]                      Signed in with work account          YES          [ ]      [ ]
Copilot [M365]                      Anthropic subprocessor (Legal Agent) [OFF/ON]     [ ]      [ ]
Perplexity [Pro/Enterprise]         AI data retention                    OFF          [ ]      [ ]
All                                 Vendor terms changed since last log  NO           [ ]      [ ]

Next check due: [date]

Three habits make the log useful: record the plan, not just the tool, because the plan is what a regulator or court will ask about; note any product renamed or merged since the last check, because a merged product tends to inherit the more permissive setting; and have someone other than the person who set it up run the check every third month.

What to do this week

  1. Open each tool you actually use and run the steps above. Thirty minutes covers all five.
  2. List everyone who uses a personal account for work, including a personal Gemini or a Max-tier Perplexity, and move them to a firm plan or off the tool. Clio reported in 2025 that legal-specific tool use had fallen to 40% as lawyers defaulted to general chatbots; personal accounts are the shadow of that number.
  3. Write the plan and the settings into the firm's AI policy, using the law firm AI policy template, so the toggle is a rule rather than a preference.
  4. Ask the admin to confirm, in writing, the three enterprise decisions that lawyers cannot see: which Claude model is default, whether GPT-6 Astra is enabled, and whether Anthropic is a subprocessor in the Microsoft tenant.
  5. Compare the plan you are on against the full confidentiality checklist and decide, tool by tool, whether it is fit for client data at all.

Module 6 of the AI Lab for Lawyers course, on risks and regulation, goes through these vendor settings alongside the professional rules that make them matter.

Frequently asked questions

How do I stop ChatGPT from training on my conversations?

Open your profile menu, go to Settings, then Data Controls, and switch off the option labelled Improve the model for everyone. The change applies to your whole account but only to new chats. On Business, Enterprise, Edu and API plans there is nothing to switch because those tiers do not train by default. Opting out does not shorten retention or make chats undiscoverable.

Does Claude use my chats to train its models?

On Free, Pro and Max plans only if you left the Help improve Claude setting on; Anthropic introduced that choice on 28 August 2025 and chats from opted-in users are kept up to five years. Turn it off under Settings and Privacy and retention drops to 30 days. Team, Enterprise and API usage is never used for training. Incognito chats are not used either.

Is Gemini safe for confidential client work?

Not on a personal account. Consumer Gemini keeps activity 18 months by default, sends a subset of chats to human reviewers and uses them for training unless Keep Activity is off, and Google itself says not to enter confidential information a reviewer should not see. Use Gemini through a Workspace or Gemini Enterprise account, where training is excluded by default through the admin console.

Is Microsoft Copilot the same as Microsoft 365 Copilot for privacy purposes?

No. Consumer Copilot, including the free app and Copilot Pro, runs under consumer terms without a data processing agreement and has been described as unsuitable for confidential or legal work. Microsoft 365 Copilot, the paid add-on for work accounts, runs under Microsoft's DPA with Enterprise Data Protection and Purview controls and does not train on your prompts. Check which account you are signed in with.

What is zero data retention and can a law firm get it?

Zero data retention means the vendor does not store your prompts or outputs after the response is returned. It is a contractual term on API and enterprise plans, not a toggle. As of September 2026 Anthropic arranges it per organisation, but Claude Team and Enterprise chat are not eligible and Fable 5.1 and Mythos are excluded; OpenAI offers it contractually to eligible customers. Most firms will not get it on a chat subscription.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion