AI Agents for Lawyers

Guides

Which AI Tools Train on Your Data? A Confidentiality Checklist for Lawyers

Which AI tools train on your data, plan by plan as of September 2026: ChatGPT, Claude, Gemini, Copilot, legal vendors, plus bar rules and a 10-point checklist.

By Dr. Niklas SchmidtUpdated 13 November 202611 min read
On this page
  1. Two rulings that frame the question
  2. The plan-by-plan table
  3. Four questions after "does it train"
  4. Switching training off, tool by tool
  5. What the professional rules require
  6. The agent problem
  7. The 10-point confidentiality checklist
  8. Where this leaves you

As of September 2026 the consumer tiers of ChatGPT, Claude and Gemini use your conversations to train models unless you opt out; the business, enterprise and API tiers of all three do not by default; the legal vendors say they never do. Training is only the first question. Retention, human review, discoverability and where the data sits decide whether a tool is fit for client work.

Two rulings that frame the question

In United States v. Heppner (S.D.N.Y., bench ruling 10 February 2026, opinion 17 February 2026), Judge Rakoff held that research memos a defendant had generated in consumer Claude were neither privileged nor work product. Privilege never attached, "because Claude is not an attorney", and Anthropic's consumer terms allow logging and disclosure. The court left open that privilege might attach where counsel directs the AI use in the way a Kovel accountant is directed (Harvard Law Review blog, Mar 2026). Two months earlier, in the OpenAI copyright litigation in the same court, Judge Stein had affirmed an order producing 20 million de-identified ChatGPT logs, reasoning that users "voluntarily submitted their communications" to OpenAI (National Law Review). Chat logs are discoverable ESI.

In England, the Upper Tribunal in UK v SSHD [2026] UKUT 81 (IAC) warned that pasting client letters into public AI tools puts them "in the public domain". The SRA's Warning Notice of 17 August 2026 cites that case, together with BCP v A Mother [2026] EWFC 71 (B) and Cork v Smith [2026] EWHC 1199 (Ch), and records 42 reports of AI misuse between July 2025 and July 2026 (SRA). The notice's test is whether the solicitor understands a tool "well enough to review its output critically". The table below is half of that understanding; the privilege page is the other half.

The plan-by-plan table

All entries are vendor-stated policies as of September 2026 and can change without notice. The plan matters more than the brand: a $20 subscription is a consumer plan whatever the marketing says.

Tool and planTrains on your inputs by default?RetentionOpt-outZero data retention
ChatGPT Free, Go, Plus, ProYes, unless opted outWhile the account exists; a 30-day abuse-monitoring retention is reported even after opt-outSettings > Data Controls > "Improve the model for everyone" off; new chats onlyNo
ChatGPT Business, Enterprise, Edu; OpenAI APINoAPI 30 days for abuse monitoringNot neededContractual for eligible customers; Astra for Law ships a ZDR API (17 Sep 2026)
Claude Free, Pro, MaxOnly if you chose "Help improve Claude" (choice required since 28 Aug 2025)5 years if opted in; 30 days if opted out; incognito chats never usedSettings > Privacy > toggle offNo
Claude Team, EnterpriseNoFlagged content up to 2 years; Compliance API and local session transcripts 6 years by defaultNot neededChat interface not ZDR-eligible; Claude Code under Enterprise is
Claude API, Bedrock, VertexNoNot retained by default, except "Covered Models" (Fable 5/5.1, Mythos) which require 30-day retentionNot neededPer organisation via sales; Fable and Mythos excluded without express authorisation
Gemini app, Google AI Pro/Ultra (personal accounts only)Yes, unless "Keep Activity" is off; a subset of chats is read by human reviewers18 months by default; 72 hours even with activity off; reviewed chats up to 3 yearsmyactivity.google.com/product/geminiNo
Gemini for Workspace, Gemini Enterprise, VertexNo, excluded via the admin consolePer admin policy; EU data residency for NotebookLM Enterprise since 3 Jun 2026Not neededContractual
Microsoft Copilot (consumer, Free/Pro)Consumer terms, no DPAConsumer termsMove to the tenant productNo
Microsoft 365 Copilot ($30/user/month)No; runs under the DPA with Enterprise Data Protection and PurviewTenant-controlledNot neededTenant-bound; note the Word Legal Agent requires Anthropic enabled as a subprocessor
Perplexity Pro, MaxControlled by the "AI data retention" togglePer settingSettings > Preferences > "AI data retention" offNo
Perplexity Enterprise Pro/MaxNo, never by defaultPer contractNot neededContractual
Harvey, Legora, CoCounsel Legal, Lexis+ ProtégéNo, per vendor statements; enterprise contractsPer contract; Lexis offers customer-held encryption keys (May 2026)Not applicablePer contract; confirm which frontier models sit underneath and on what terms

Sources: Anthropic data retention documentation, Anthropic privacy centre, OpenAI data controls FAQ, OpenAI enterprise privacy, Google Gemini Apps privacy hub, updated 10 Aug 2026, and a June 2026 review of Copilot's consumer and enterprise terms (Spellbook, 20 Jun 2026). Google's own page carries the line every lawyer should read twice: "Please don't enter confidential information that you wouldn't want a reviewer to see."

One consequence is counter-intuitive. The strongest Claude models are the least private: Fable 5.1 and Mythos 5.1 are "Covered Models" that require 30-day retention and are barred from zero-data-retention arrangements unless Anthropic expressly authorises it. A firm that needs zero data retention for a matter is, as of September 2026, choosing Opus 5 or Sonnet 5 on the API, or an OpenAI Enterprise or ZDR API arrangement.

Four questions after "does it train"

Retention and human review. "No training" says nothing about how long a transcript exists or who reads it. Gemini keeps human-reviewed consumer chats up to three years; Claude keeps opted-in consumer chats five years; Anthropic's Compliance API keeps Cowork, Claude Code and Claude-in-Chrome session transcripts for six years by default (Anthropic Compliance API docs). Six years of agent transcripts is an asset for supervision and a liability in discovery.

Discoverability. The 20-million-log order shows that a vendor's stored conversations can be pulled into someone else's litigation. Heppner shows that a client's own chatbot use is not shielded. Both point the same way: the lawyer, not the client, should run the tool, on terms the firm has read. The privilege question has its own page.

Where it runs and who else touches it. Microsoft's Legal Agent for Word (30 April 2026) "requires Anthropic enabled as a subprocessor", a fact one commentator noted firm IT departments "have not formally evaluated" (PlatinumIDS, 2026). Claude Managed Agents, Perplexity Computer and remote Cowork sessions run in vendor clouds. For a European firm each of these needs an Art. 28 GDPR processor agreement and, in Germany, a § 43e BRAO contract; the GDPR page sets out the documents. A data-residency toggle changes where data sits, not which government can compel its production, which is one reason Latham & Watkins bought its own servers in a data centre only its staff can enter (reported 10–11 September 2026).

Whether the switch works. In July 2026 a wire-level analysis showed that xAI's Grok Build uploaded entire repositories including an unredacted secrets file, 5.10 GiB where the model used about 192 KB, and that the "Improve the model" toggle did nothing; xAI fixed it server-side on 13 July (The Hacker News, Jul 2026). A privacy setting is a vendor promise, not a control. The controls a firm actually owns are the choice of tier, the contract and what it allows into the tool in the first place.

Switching training off, tool by tool

For any consumer account in the firm, and there are more than IT knows about, the September 2026 paths are:

  1. ChatGPT: profile icon > Settings > Data Controls > switch off "Improve the model for everyone". Applies account-wide, to new chats only.
  2. Claude: name > Settings > Privacy > switch off "Help improve Claude". Incognito chats are never used for training.
  3. Gemini: go to myactivity.google.com/product/gemini and turn "Keep Activity" off. Chats are still held for 72 hours.
  4. Perplexity: Settings > Preferences > "AI data retention" off.
  5. Copilot: there is no consumer switch that produces a DPA; move the user onto Microsoft 365 Copilot in the tenant.

Screenshots and admin-console equivalents are in AI privacy settings step by step. Opting out is the floor. It does not turn a consumer plan into a plan on which client documents belong, and none of the bar guidance below treats it as sufficient.

What the professional rules require

JurisdictionRule or guidanceWhat it asks of you
USA (ABA)Formal Opinion 512 (29 Jul 2024)Competence with the tool; confidentiality; informed client consent before using a self-learning tool that may expose client data; supervision of the output
CaliforniaCOPRAC comments approved 13 Mar 2026; Practical Guidance on generative and agentic AI, 14 May 2026Proposed Rule 1.6 comment treats exposing client data to AI as "revealing" it where there is "a material risk of unauthorized access, retention, or use"
TexasOpinion 705 (Feb 2025)A "reasonable and current understanding of the technology"; no blind reliance
England and WalesSRA Warning Notice (17 Aug 2026); Judiciary guidance (31 Oct 2025)Understand the tool "well enough to review its output critically"; no private information in public AI tools
GermanyBRAK "Hinweise zum Einsatz von KI" (Dec 2024)Tools that store or forward client data for model training breach § 43a Abs. 2 BRAO; service providers need a § 43e BRAO contract
AustriaÖRAK practice rules (KI-Tag, 19 Mar 2026); § 40 Abs 3 RL-BA 2015Client data stays on EU servers; disclose AI use to clients; the lawyer keeps final responsibility; no client data for training
Europe (CCBE)Guide on generative AI (2 Oct 2025); Technical Guide (adopted 27 Mar 2026)Inform clients and let them object; assess data-retention risk; choose the deployment model, on-premise to SaaS, deliberately
EU data protectionEDPB Opinion 28/2024 (17 Dec 2024)Deployers must assess whether the model was trained unlawfully; legitimate-interest test for processing

Links: ABA Opinion 512, California 2026 guidance summary, BRAK Hinweise (PDF), ÖRAK KI-Tag report, CCBE guide summary. Fuller treatment is in the pages on ABA Opinion 512, the SRA notice and Austrian and German bar guidance.

The rules converge on four demands: a contract that excludes training and fixes retention, data in a jurisdiction the bar accepts, a client who has been told, and a lawyer who understands the tool. The German and Austrian rules are most explicit that a training default is itself a breach; the California comment is most explicit that retention alone can be "revealing".

The agent problem

The table above assumes a chatbot: you paste, it answers. A connector that lets Claude, ChatGPT or Copilot search the document management system exposes the whole matter, not one document. The two DMS vendors have built for this: iManage's MCP Server (14 May 2026) is permission-bound, fully logged and "respects existing ethical walls" (iManage); NetDocuments' ndMAX connector runs every call under the user's own credentials, so matter restrictions apply automatically, and "documents never leave the platform" (NetDocuments). Demand those properties from any connector; the DMS connection playbook shows how to check them. Prompt injection, where a poisoned email instructs the agent to exfiltrate data, is covered in agent security for lawyers.

The 10-point confidentiality checklist

Run this for every AI tool in the firm, including the ones individual lawyers pay for themselves.

  1. Name the plan, not the product. "We use Claude" is not an answer; "Claude Team, five seats, no training, no ZDR" is.
  2. Training is off or contractually excluded, and the setting has been checked on every account, not assumed.
  3. Retention is known and acceptable, human review is excluded, and the model in use is not one that forces retention if the matter needs ZDR.
  4. A processor agreement exists: Art. 28 GDPR, § 43e BRAO where relevant, and a subprocessor list that has been read, including Anthropic behind Microsoft's Word agent.
  5. Data location satisfies the bar: EU servers for Austrian matters under the ÖRAK rules; on-premise or ZDR for the most sensitive files.
  6. Logs are understood: who can read them, for how long, whether they capture agent sessions, and that they are discoverable ESI.
  7. The lawyer directs the tool. Clients are told not to prepare their case in consumer chatbots; the firm's use is documented so that privilege arguments of the Kovel kind are available.
  8. Clients are informed, in the engagement letter or on the matter, and can object. Clio's 2025 report found 78% of clients want disclosure while 35% of lawyers rarely give it; the disclosure page has wording.
  9. Agents are scoped: connectors run under the user's credentials and ethical walls, write and send actions are gated, and no agent has file-system access to a matter folder without a human approval step.
  10. It is written down in the firm's AI policy with a named owner and re-checked quarterly; vendor terms changed three times in the past year alone.

The vendor due-diligence page turns items 2 to 6 into questions for the procurement call. Module 6 of the AI Lab for Lawyers course works through these settings live, tool by tool, so that the checklist is something you have done rather than something you have read.

Where this leaves you

For most firms the practical answer in September 2026 is a business or enterprise tier of one general model (Claude or ChatGPT) with a signed DPA and training excluded, a legal or DMS-embedded tool for matter documents, ZDR or on-premise for the few matters that justify it, and a firm-wide ban on client data in any consumer plan, opted out or not. The ban is the one rule never to soften; it is the only line a court has already drawn.

Frequently asked questions

Does ChatGPT train on my data if I am a lawyer?

On Free, Go, Plus and Pro it does unless you switch off the setting labelled Improve the model for everyone, and a 30-day abuse-monitoring retention is reported even after opting out. On Business, Enterprise, Edu and the API, OpenAI does not train by default and zero data retention can be arranged contractually. As of September 2026, the paid consumer plan is still a consumer plan.

Is Claude safer than ChatGPT for confidential client documents?

On comparable tiers the policies are similar. Claude Free, Pro and Max users had to choose whether chats train models; opted-in data is kept five years, opted-out data 30 days. Team, Enterprise and API data is never used for training. One difference matters for law firms: the Fable and Mythos models require 30-day retention and are not eligible for zero data retention without express authorisation, whichever tier you are on.

What is zero data retention and does my firm need it?

Zero data retention is a contractual arrangement under which the provider does not store your prompts or outputs after the response is delivered. It is available on OpenAI Enterprise and API plans and on Anthropic API plans per organisation, not on consumer plans or on Claude Team and Enterprise chat. Most firms need it for privileged litigation material and regulated data; for routine drafting a no-training enterprise tier with a DPA is usually enough.

Does using an AI tool waive attorney-client privilege?

It can, depending on who uses it and on what terms. In United States v. Heppner a federal judge held that a defendant's own research memos generated in consumer Claude were never privileged because Claude is not a lawyer and the terms permit logging and disclosure. The court indicated privilege may attach where counsel directs the AI use, so lawyer-directed use on enterprise terms is the safer structure.

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

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.

10 min read

Glossary

Zero Data Retention (ZDR) — Definition for Lawyers

Zero data retention means the AI vendor stores nothing after replying. Which plans offer ZDR, the Covered Models exception, and why it matters for privilege and GDPR.

5 min read

Regulation & Ethics

Attorney–Client Privilege and AI: Heppner, Warner, the NYT Logs and Public Chatbots

Consumer chatbot memos are not privileged. US v. Heppner, Warner v. Gilbarco and the 20-million-log OpenAI order, plus rules that keep privilege intact.

10 min read

Regulation & Ethics

GDPR and AI Tools for Lawyers: Lawful Basis, DPIAs and Vendor Terms

How GDPR applies when law firms use Claude, ChatGPT or Copilot: controller and processor roles, lawful basis, DPIA triggers, transfers and vendor terms.

11 min read

Regulation & Ethics

AI Rules for Austrian and German Lawyers: ÖRAK, BRAK and CCBE Guidance

ÖRAK's four AI rules and § 40 Abs 3 RL-BA bind Austrian lawyers; BRAK's § 43a BRAO line binds German ones; CCBE guides cover both. With a bilingual checklist.

11 min read

Questions

Does Using ChatGPT Waive Attorney–Client Privilege?

Using ChatGPT rarely waives privilege in the technical sense; the courts' answer is worse. In Heppner, consumer-chatbot memos were never privileged at all.

6 min read