---
name: "Use AI at work without leaking company data"
description: "Tells a small business, item by item, whether a piece of work data is safe to paste or upload into ChatGPT, Claude, Gemini or Microsoft Copilot, which settings and plan make it safer, and how to redact it, using the vendors’ own help pages and the data-protection rules of the US (HIPAA and professional confidentiality), UK, EU, UAE and India; other countries are not checked. Used when you or your staff paste work documents, emails or spreadsheets into an AI chat tool."
version: 2
verified_on: 2026-11-05
reverify_by: 2027-02-03
sources:
  - id: t1-oa-datacontrols
    title: "Data controls in ChatGPT (OpenAI Help Center)"
    url: https://help.openai.com/en/articles/7730893-data-controls-in-chatgpt
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-oa-tempchat
    title: "Temporary chat in ChatGPT (OpenAI Help Center)"
    url: https://help.openai.com/en/articles/8914046-temporary-chat-in-chatgpt
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-oa-retention
    title: "Chat and file retention in ChatGPT (OpenAI Help Center)"
    url: https://help.openai.com/en/articles/8983778-chat-and-file-retention-in-chatgpt
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-oa-enterprise
    title: "Enterprise privacy at OpenAI (updated 8 January 2026)"
    url: https://openai.com/enterprise-privacy/
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-oa-nyt
    title: "How we are responding to The New York Times’ data demands (OpenAI, updated 22 October 2025)"
    url: https://openai.com/index/response-to-nyt-data-demands/
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-an-terms
    title: "Updates to Consumer Terms and Privacy Policy (Anthropic, 28 August 2025)"
    url: https://www.anthropic.com/news/updates-to-our-consumer-terms
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-an-retention
    title: "How long do you store my data? (Anthropic Privacy Center, consumer plans)"
    url: https://privacy.claude.com/en/articles/10023548-how-long-do-you-store-my-data
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-an-setting
    title: "How do I change my model improvement privacy settings? (Anthropic Privacy Center)"
    url: https://privacy.claude.com/en/articles/12109829-how-do-i-change-my-model-improvement-privacy-settings
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-an-incognito
    title: "Use incognito chats (Claude Help Center)"
    url: https://support.claude.com/en/articles/12260368-use-incognito-chats
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-an-commercial
    title: "Is my data used for model training? (Anthropic Privacy Center, commercial products)"
    url: https://privacy.claude.com/en/articles/7996868-is-my-data-used-for-model-training
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-an-org-retention
    title: "How long do you store my organization’s data? (Anthropic Privacy Center)"
    url: https://privacy.claude.com/en/articles/7996866-how-long-do-you-store-my-organization-s-data
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-an-baa
    title: "Business Associate Agreements (BAA) for commercial customers (Anthropic Privacy Center)"
    url: https://privacy.claude.com/en/articles/8114513-business-associate-agreements-baa-for-commercial-customers
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-g-privacyhub
    title: "Gemini Apps Privacy Hub (Google, last updated 26 October 2026)"
    url: https://support.google.com/gemini/answer/13594961?hl=en
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-g-activity
    title: "Manage and delete your activity in Gemini Apps (Google)"
    url: https://support.google.com/gemini/answer/13278892?hl=en
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-g-workspace
    title: "Generative AI in Google Workspace Privacy Hub (Google, last updated 6 October 2026)"
    url: https://knowledge.workspace.google.com/admin/generative-ai/generative-ai-in-google-workspace-privacy-hub
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-ms-edp
    title: "Enterprise data protection in Microsoft Copilot and Microsoft Copilot Chat (Microsoft Learn, updated 4 November 2026)"
    url: https://learn.microsoft.com/en-us/copilot/plan/enterprise-data-protection
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-ms-chat
    title: "Microsoft Copilot Chat privacy and protections (Microsoft Learn, updated 4 November 2026)"
    url: https://learn.microsoft.com/en-us/copilot/plan/privacy-and-protections
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-ms-privacy
    title: "Data, privacy and security for Microsoft Copilot (Microsoft Learn, updated 4 November 2026)"
    url: https://learn.microsoft.com/en-us/copilot/plan/microsoft-365-copilot-privacy
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-eu-gdpr
    title: "Regulation (EU) 2016/679, General Data Protection Regulation: Recital 26, Articles 6, 9, 28, 33, 44 to 46 (EUR-Lex)"
    url: https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32016R0679
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-uk-gdpr-9
    title: "UK GDPR Article 9: processing of special categories of personal data (legislation.gov.uk)"
    url: https://www.legislation.gov.uk/eur/2016/679/article/9
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-uk-gdpr-28
    title: "UK GDPR Article 28: processor (legislation.gov.uk)"
    url: https://www.legislation.gov.uk/eur/2016/679/article/28
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-eu-us-dpf
    title: "EU-US data transfers and the Data Privacy Framework (European Commission)"
    url: https://commission.europa.eu/law/law-topic/data-protection/international-dimension-data-protection/eu-us-data-transfers_en
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-eu-ailit
    title: "AI Literacy: questions and answers (European Commission, last update 20 October 2026)"
    url: https://digital-strategy.ec.europa.eu/en/faqs/ai-literacy-questions-answers
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-in-dpdp-act
    title: "Digital Personal Data Protection Act, 2023, section 8 (MeitY)"
    url: https://www.meity.gov.in/static/uploads/2024/06/2bf1f0e9f04e6fb4f8fef35e82c42aa5.pdf
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-in-dpdp-commencement
    title: "Notification G.S.R. 843(E), 13 November 2025, appointing commencement dates under section 1(2) of the Digital Personal Data Protection Act, 2023 (MeitY, Gazette of India Extraordinary)"
    url: https://www.meity.gov.in/static/uploads/2025/11/c56ceae6c383460ca69577428d36828b.pdf
    tier: T1
    accessed_on: 2026-11-05
  - id: t1-in-dpdp-rules
    title: "Digital Personal Data Protection Rules, 2025, G.S.R. 846(E), rule 1 commencement (MeitY)"
    url: https://www.meity.gov.in/static/uploads/2025/11/53450e6e5dc0bfa85ebd78686cadad39.pdf
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-in-pib
    title: "DPDP Rules, 2025 notified (Press Information Bureau backgrounder, 17 November 2025)"
    url: https://static.pib.gov.in/WriteReadData/specificdocs/documents/2025/nov/doc20251117695301.pdf
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-uae-health
    title: "Federal Law No. 2 of 2019 on the use of ICT in health fields, Articles 2, 4, 13 and 24 (Emirates Health Services copy)"
    url: https://www.ehs.gov.ae/app_content/legislations/php-law-en-77/files/basic-html/page5.html
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-uae-dp
    title: "Data protection laws, including Federal Decree-Law No. 45 of 2021 (UAE Government portal u.ae, updated 4 December 2025)"
    url: https://u.ae/en/about-the-uae/digital-uae/data/data-protection-laws
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-difc-dp
    title: "Commissioner of Data Protection: DIFC Data Protection Law No. 5 of 2020 (DIFC)"
    url: https://www.difc.com/business/registrars-and-commissioners/commissioner-of-data-protection
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-adgm-dp
    title: "Office of Data Protection: ADGM Data Protection Regulations 2021 (ADGM)"
    url: https://www.adgm.com/operating-in-adgm/office-of-data-protection
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-us-cfr-164502
    title: "45 CFR 164.502(e): disclosures to business associates (eCFR, up to date as of 2 November 2026)"
    url: https://www.ecfr.gov/current/title-45/subtitle-A/subchapter-C/part-164/subpart-E/section-164.502
    tier: T1
    accessed_on: 2026-11-04
  - id: t2-aba-512
    title: "ABA issues first ethics guidance on a lawyer’s use of AI tools: Formal Opinion 512 (American Bar Association, 29 July 2024)"
    url: https://www.americanbar.org/news/abanews/aba-news-archives/2024/07/aba-issues-first-ethics-guidance-ai-tools/
    tier: T2
    accessed_on: 2026-11-04
  - id: t1-ms-consumer
    title: "Microsoft Copilot for individuals: your privacy controls and choices (Microsoft Support, new Copilot app)"
    url: https://support.microsoft.com/en-us/privacy/microsoft-copilot/privacy-controls
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-ms-consumer-history
    title: "Microsoft Copilot for individuals: your activity history (Microsoft Support)"
    url: https://support.microsoft.com/en-us/privacy/microsoft-copilot/activity-history
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-ms-consumer-old
    title: "Microsoft Copilot privacy controls (Microsoft Support, older Copilot app only)"
    url: https://support.microsoft.com/microsoft-copilot/microsoft-copilot-privacy-controls
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-an-dpa
    title: "How do I view and sign your Data Processing Addendum (DPA)? (Anthropic Privacy Center)"
    url: https://privacy.claude.com/en/articles/7996862-how-do-i-view-and-sign-your-data-processing-addendum-dpa
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-an-covered
    title: "Data retention practices for Covered Models (Anthropic Privacy Center)"
    url: https://privacy.claude.com/en/articles/15425996-data-retention-practices-for-covered-models
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-oa-nyt-20m
    title: "OpenAI’s Rule 72(a) objections to the orders compelling production of 20 million ChatGPT logs, In re OpenAI Copyright Infringement Litigation, 25-md-03143, ECF 935 (S.D.N.Y., filed 10 December 2025)"
    url: https://cdn.openai.com/pdf/2025-12-10-memorandum-of-law-dckt-935-0.pdf
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-uk-us-bridge
    title: "UK-US data bridge: supporting documents (UK Government, gov.uk)"
    url: https://www.gov.uk/government/publications/uk-us-data-bridge-supporting-documents
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-uk-gdpr-46
    title: "UK GDPR Article 46: transfers subject to appropriate safeguards (legislation.gov.uk)"
    url: https://www.legislation.gov.uk/eur/2016/679/article/46
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-uae-pdpl
    title: "Federal Decree-Law No. 45 of 2021 on the Protection of Personal Data, Article 2 (official Arabic text, linked from u.ae)"
    url: https://assets.u.ae/api/public/content/954629a06cc64fd887771ce3273d6154?v=21d88839
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-uae-mr51
    title: "Ministerial Resolution No. 51 of 2021 on the cases in which health data may be stored or transferred outside the UAE (Ministry of Health and Prevention, last updated 7 February 2026)"
    url: https://uaephl.mohap.gov.ae/en/health-policies-and-legislations-advocacy/health-legislations?itemId=726efdef-580e-4817-bf5a-fbf8e53fddfc
    tier: T1
    accessed_on: 2026-11-04
  - id: t1-court-nyt-20m-order
    title: "Order denying OpenAI’s objections and affirming the orders to produce the 20 million log sample, In re OpenAI, Inc. Copyright Infringement Litigation, 25-md-3143, ECF 1021 (S.D.N.Y., 5 January 2026; RECAP copy on CourtListener)"
    url: https://storage.courtlistener.com/recap/gov.uscourts.nysd.640396/gov.uscourts.nysd.640396.1021.0.pdf
    tier: T1
    accessed_on: 2026-11-05
  - id: t1-eu-c703-25-oj
    title: "Case C-703/25 P: appeal brought on 31 October 2025 by Philippe Latombe against the judgment of the General Court of 3 September 2025 in Case T-553/23 on the EU-US Data Privacy Framework adequacy decision (Official Journal C/2025/6610, EUR-Lex)"
    url: https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:62025CN0703
    tier: T1
    accessed_on: 2026-11-05
  - id: t1-curia-c703-25
    title: "Case C-703/25 P, Latombe v Commission: case details, appeal from Case T-553/23 (Court of Justice of the European Union, InfoCuria)"
    url: https://curia.europa.eu/juris/liste.jsf?num=C-703/25
    tier: T1
    accessed_on: 2026-11-05
  - id: t1-us-cfr-7216
    title: "26 CFR 301.7216-1: penalty for disclosure or use of tax return information (eCFR, up to date as of 3 November 2026)"
    url: https://www.ecfr.gov/current/title-26/chapter-I/subchapter-F/part-301/subpart-ZZ/section-301.7216-1
    tier: T1
    accessed_on: 2026-11-05
---

> **How to use this file:** attach it to ChatGPT, Claude, Gemini or Copilot, replace every [PLACEHOLDER] with your details, and add any data or Deepstory evidence files the method needs.

**Best with:** Any chat AI. Describe the data or paste a redacted sample; do not paste the real data into the chat you use to run this check.

## Quick start

1. Before you paste anything into ChatGPT, Claude, Gemini or Copilot, ask: whose information is this? This check covers the US, UK, EU, UAE and India only.
2. Never paste passwords, API keys, card numbers, health or patient details, or client files your contract keeps private. If you already did, change the password or key now.
3. On a personal plan, switch off training: ChatGPT Settings > Data controls > "Improve the model for everyone"; Claude Settings > Privacy > "Help Improve our AI models"; Gemini Settings & help > Activity > turn off Keep Activity. Personal Microsoft Copilot (new app) has no training switch, and Microsoft says your prompts, responses and files are not used to train its foundation models; still turn off Saved memories and Allow ads personalization under Settings > Personalization.
4. Switching off training is not enough for customer or staff data. Use a business plan (ChatGPT Business or Enterprise, Claude Team or Enterprise, Google Workspace, Copilot with the green shield) and make sure the AI company’s data processing agreement (DPA) is in place: OpenAI asks you to complete its DPA form; Anthropic, Microsoft and Google include theirs in the business terms.
5. Remove names and details first: write [CUSTOMER_1] instead of a name, and delete phone numbers, ID numbers and dates of birth. Or ask the AI for a template and fill it in yourself.
6. Do not click thumbs up or down on chats with business data. "Deleted" chats can still be kept for up to 30 days, or longer.
7. This is a first-pass check, not legal or tax advice. Confirm with a qualified accountant or lawyer for your country.

## Before you start

- You want to choose AI tools, set up projects, custom instructions or memory, or write a team AI policy. This method only covers what data is safe to paste and the privacy settings. See: Set up your AI workspace (projects, instructions, memory) (https://deepstoryresearch.com/method/ai-workspace-setup)
- You are putting an AI chatbot in front of customers on your website, help desk or WhatsApp. See: Set up an AI support chatbot that customers trust (https://deepstoryresearch.com/method/ai-support-chatbot-setup)
- You need a full data map, records of processing or a privacy compliance review across the whole business. See: Check data-privacy compliance gaps (https://deepstoryresearch.com/method/data-privacy-gap-check)
- You need to update your website privacy notice to say you use AI vendors. See: Check your website’s privacy policy, cookies, terms and refund policy (https://deepstoryresearch.com/method/website-privacy-terms)
- Data has already leaked (for example a password, customer list or patient file pasted into a personal AI account). Treat it as a possible data breach. Rule of thumb: change any exposed password or key now; note which account, roughly when and what kind of data, without copying the data itself; then delete the chats, any saved memories and any project or GPT files that hold it (if it is an employee’s account, they do this while you watch). Speak to your data-protection lead or a lawyer today about whether you must report it: in the EU and UK a reportable breach must normally reach the regulator within 72 hours of you becoming aware of it, and in India, from 13 May 2027, the DPDP Act requires the business that decides why and how the data is used (the Data Fiduciary) to tell the Data Protection Board and each affected person. Elsewhere, including the UAE, ask a lawyer whether you must report it.
- Your business or the people in the data are outside the US, UK, EU, UAE and India. The method will only flag that local law was not checked.

## Intake questions

Ask me these first:

1. **Which AI tool do you use, which plan (for example ChatGPT Free, Go, Plus, Pro, Business or Enterprise; Claude Free, Pro, Max, Ultra, Team or Enterprise; Gemini with a personal or work Google account; Microsoft Copilot with a personal or a work account), and are you signed in with a personal or a work account?**
   Why: Consumer plans may use chats to train models and come with no data processing agreement; business plans do not train on your data by default and offer one.
   If unknown, assume: Assume a free personal account with training switched on, which is the riskiest case, and say so.
2. **What exactly do you want to paste or upload, and what do you want the AI to do with it? Describe it or give a made-up example; do not paste the real data here.**
   Why: The task often works just as well with a template, dummy data or a redacted copy.
   If unknown, assume: Treat the material as containing personal data until the user confirms otherwise.
3. **Whose information is in it: your own public material, customers, patients, staff, or a client’s business? Does it include health, biometric, religion, ethnicity, sexual life, union membership, ID numbers, bank or card details, passwords or keys?**
   Why: Health and the other special categories have stricter rules in the EU, UK and UAE, and passwords, keys and card numbers are never safe to paste.
   If unknown, assume: Assume it may include health or financial details and ask for one example field list.
4. **Where is your business set up (for the UAE: mainland, DIFC or ADGM), and where do the people in the data live?**
   Why: The EU and UK GDPR, India’s DPDP Act, UAE law and US HIPAA apply by location and sector; DIFC and ADGM have their own data-protection laws.
   If unknown, assume: Apply the strictest of the covered places involved and list the rest as not checked.
5. **Has your company signed a data processing agreement (DPA) or, for US health data, a business associate agreement (BAA) with the AI company, and who controls the workspace settings?**
   Why: Without a contract the AI company is not bound to act only on your instructions, which GDPR, UK GDPR, India’s DPDP Act and HIPAA expect for this kind of processing.
   If unknown, assume: Assume no DPA or BAA, which puts personal data in red until one is signed.
6. **Do you work in a regulated or confidential profession (health, law, accounting, finance), or do client contracts or NDAs limit who you may share their information with?**
   Why: Professional confidentiality and contract terms can forbid sharing even where data-protection law would allow it.
   If unknown, assume: Assume client information is confidential and list "check the confidentiality clause" as a question for the lawyer.
7. **Have you clicked thumbs up or down, connected email or file storage, saved files into projects, or left memory on (saved memories, or past chats used in new ones) in this tool?**
   Why: Feedback can send the whole conversation for training even when training is off, connectors and saved files keep data longer than a single chat, and memory carries details into later chats.
   If unknown, assume: Assume memory is on and the rest is not. Tell the user not to use the feedback buttons on chats with business data, to turn off memory for chats with amber data, and to delete saved memories that hold business data.

## Decision rules

| If | Then | Threshold | Type | Source |
|---|---|---|---|---|
| The user is on ChatGPT Free, Go, Plus or Pro in a personal workspace. | Data sharing for model training is on by default (OpenAI states this for Free, Plus and Pro; treat Go, also a personal plan, the same and check the toggle). Turn off "Improve the model for everyone" (Settings > Data controls); it applies to the signed-in account across web and mobile. Voice mode and Codex have separate training controls. Turning it off does not delete saved chats. Giving thumbs up or down feedback can still send the entire conversation for training. As of November 2026 OpenAI also says that during Trusted Access sessions (such as Daybreak Blue) content may be used for training even with the setting off. | - | Platform rule | [t1-oa-datacontrols] |
| The user wants a ChatGPT chat that is not kept in history or used for training, or wants to know how long deleted chats stay. | Use Temporary chat (new chat > Temporary): it stays out of history, does not create memories and is not used for training while it stays temporary, but OpenAI may keep a copy for up to 30 days for safety. Saving it turns it into a regular chat. A deleted saved chat leaves your view at once and is scheduled for permanent deletion within 30 days, unless already de-identified or OpenAI must keep it longer for security or legal reasons. Rule of thumb: Temporary chat does not make red data safe. | - | Platform rule | [t1-oa-tempchat], [t1-oa-retention] |
| The business uses ChatGPT Business, Enterprise, Edu or ChatGPT for Healthcare, or the OpenAI API. | OpenAI does not train on this business data by default (unless you opt in, for example through feedback). OpenAI will sign a data processing agreement (DPA) for ChatGPT Business, ChatGPT Enterprise and the API through its DPA form; workspace admins set retention, and deleted conversations are removed within 30 days unless the law requires longer or (for ChatGPT Business) OpenAI needs them to protect its services or others from harm. For ChatGPT Business and the API, OpenAI’s access is limited to authorised staff (for engineering support, abuse investigations and legal compliance) and specialised contractors bound by confidentiality who review only for abuse and misuse; for ChatGPT Enterprise, Edu and ChatGPT for Healthcare, OpenAI says authorised staff access conversations only to resolve incidents, to recover chats with your explicit permission, or where the law requires. API inputs and outputs may be kept up to 30 days for abuse checks. OpenAI says it can sign a HIPAA business associate agreement (stated for the API Platform); ask OpenAI in writing which product your BAA covers. Rule of thumb: this is the minimum for amber data. | - | Platform rule | [t1-oa-enterprise], [t1-oa-datacontrols] |
| Someone says "deleted ChatGPT chats are kept forever because of a court order". | In 2025 a US court preservation order in the New York Times case made OpenAI keep consumer ChatGPT and API content. OpenAI says the obligation ended on 26 September 2025, so new data is no longer kept indefinitely, but it still stores limited historical April to September 2025 user data under legal hold. In November and December 2025 the court also ordered OpenAI to produce a sample of 20 million de-identified consumer ChatGPT logs to the plaintiffs. On 5 January 2026 the district judge denied OpenAI’s objections and affirmed those orders, finding that users’ privacy had been adequately weighed in light of the protections already in place (the smaller sample, de-identification and the case’s protective order). Later orders in the case may change this, so check its current status. ChatGPT Enterprise, Edu and Zero Data Retention API customers were not affected. Lesson: "deleted" always means "unless the law requires longer", so never rely on deletion to undo pasting red data. | - | Platform rule | [t1-oa-nyt], [t1-oa-nyt-20m], [t1-court-nyt-20m-order] |
| The user is on Claude Free, Pro, Max or Ultra (consumer plans), including Claude Code with those accounts. | Since the consumer terms update of 28 August 2025, Anthropic trains on chats and coding sessions when the model-training setting is on. Turn off "Help Improve our AI models" (name menu > Settings > Privacy). With it on, de-identified data from new or resumed chats can be kept for up to 5 years; if you do not allow training, the 30-day retention applies, and a deleted chat is removed from back-end storage within 30 days. Chats flagged for Usage Policy violations are kept up to 2 years, feedback (thumbs up or down) for 5 years, and data already used in training is not removed. Incognito chats are not used for training. Even with the setting off, chats flagged by Anthropic’s safety classifiers may still be used for its trust and safety work and safety research. | - | Platform rule | [t1-an-terms], [t1-an-retention], [t1-an-setting], [t1-an-incognito] |
| The business uses Claude for Work (Team or Enterprise), Claude Gov or the Anthropic API. | Anthropic does not use inputs or outputs from commercial products to train models by default, unless you send feedback or otherwise opt in. API inputs and outputs are deleted within 30 days unless a longer-retention feature, a zero-data-retention agreement, Usage Policy enforcement or the law requires otherwise. Since 9 June 2026, prompts and outputs of the models Anthropic designates as Covered Models (Mythos-class models) are kept for 30 days on every platform, so an organisation with zero data retention must turn retention on to use them, unless Anthropic has told it that it is eligible to use them with zero data retention. Owners of Team or Enterprise plans can stop staff sending feedback with the "Rate chats" setting. | - | Platform rule | [t1-an-commercial], [t1-an-org-retention], [t1-an-covered] |
| The user uses Gemini (gemini.google.com or the Gemini app) with a personal Google account. | Keep Activity is on by default for users 18 or over: chats are used to improve Google services including training, and a subset is read by human reviewers, including service providers. Reviewed chats are disconnected from the account and kept for up to 3 years, even if you delete your activity. Google’s own advice is not to enter confidential information you would not want a reviewer to see. Turning Keep Activity off (Settings & help > Activity) stops future chats being used to train unless you send feedback, but chats are still kept for up to 72 hours, and Google still uses them to protect its users and the public, with help from human reviewers. Temporary chats are not used for training. Sending feedback with Keep Activity off also gives Google the last 24 hours of your chats as context. Activity auto-deletes after 18 months by default (can be 3 or 36 months, or never). | - | Platform rule | [t1-g-privacyhub], [t1-g-activity] |
| The business uses Gemini with a qualifying Google Workspace (work or school) account. | Google says Workspace content and prompts are not human reviewed or used to train generative AI models outside your domain without permission, and existing Workspace protections apply. Admins set retention: Gemini in Workspace apps keeps prompts and responses for 90 days to indefinitely, the Gemini app for up to 36 months. Check you are signed in with the work account, not a personal Gmail. | - | Platform rule | [t1-g-workspace] |
| The business uses Microsoft Copilot or Copilot Chat with a work account. | Enterprise data protection applies under Microsoft’s Data Protection Addendum and Product Terms, with Microsoft acting as data processor; prompts, responses and Microsoft Graph data are not used to train foundation models, and Copilot has opted out of human-review abuse monitoring. Copilot Chat shows a green shield next to New Chat when this applies. Rule of thumb: no shield means do not paste work data. Web search queries sent to Bing are outside the DPA and the BAA (Microsoft is an independent controller for them), and Anthropic models are excluded from the EU Data Boundary. | - | Platform rule | [t1-ms-edp], [t1-ms-chat], [t1-ms-privacy] |
| The user uses Microsoft Copilot (copilot.com or the Copilot app) signed in with a personal Microsoft account. | This is a consumer service under the Microsoft Services Agreement and the Copilot Terms of Use: there is no green shield, no enterprise data protection and no Microsoft data processing agreement. Rule of thumb: allow green data only. In the new Copilot app (available since 18 August 2026 and used on the web), Microsoft says prompts, responses and file contents are not used to train foundation models, and its privacy controls page lists no training switch, so the older training opt-out could not be confirmed for it. Under Settings > Personalization, turn off Saved memories (this does not delete memories already saved: use Manage > Delete all memories), One shared experience (which lets your chats personalise Bing, Edge and MSN) and Allow ads personalization (personalised ads can use your chat history and saved memories). If you still use the older app, Microsoft’s page for it gives profile icon > Settings > Privacy > "Training on conversation activity" and "Training on voice conversations" (opting out does not stop use for other product improvement, advertising, safety, security and compliance) and says history is kept for 18 months. | - | Platform rule | [t1-ms-consumer], [t1-ms-consumer-history], [t1-ms-consumer-old] |
| The data is about people in the EU or UK and includes health, genetic or biometric data, racial or ethnic origin, political opinions, religious or philosophical beliefs, trade union membership, sex life or sexual orientation. | This is special category data. Under GDPR Article 9 (and UK GDPR Article 9) processing it is prohibited unless one of the listed exceptions applies, in addition to a lawful basis under Article 6. Rule of thumb: treat it as red for any consumer AI plan, and as amber at best in a business plan only after a lawyer confirms the Article 9 condition and a DPA is signed. | - | Legal rule | [t1-eu-gdpr], [t1-uk-gdpr-9] |
| Personal data of customers, staff or patients in the EU or UK goes into an AI tool on the business’s behalf. | The AI company acts as a processor, so GDPR and UK GDPR Article 28 require a binding contract (a data processing agreement) that, among other things, makes it process the data only on your documented instructions. Consumer plans used under personal terms do not give you this. Rule of thumb: personal data is red there, and business plans with a signed DPA can be amber: OpenAI offers one for ChatGPT Business, Enterprise and the API; Microsoft Copilot runs under the Microsoft Data Protection Addendum; Gemini in Google Workspace follows the Cloud Data Processing Addendum. For Claude for Work (Team or Enterprise) and the Anthropic API, Anthropic’s DPA, with standard contractual clauses, is part of the Commercial Terms and is accepted with them; keep a copy. Claude used through another company’s platform falls under that platform’s terms. | - | Legal rule | [t1-eu-gdpr], [t1-uk-gdpr-28], [t1-oa-enterprise], [t1-ms-edp], [t1-g-workspace], [t1-an-dpa] |
| EU or UK personal data will be processed by an AI company outside the EU or UK (for example in the US). | A transfer outside the EU is allowed only under GDPR Chapter V: an adequacy decision (for the US, only to companies certified under the EU-US Data Privacy Framework) or appropriate safeguards such as standard contractual clauses. For UK data the same idea applies under UK GDPR Chapter V: US companies that take part in the UK Extension to the EU-US Data Privacy Framework (in effect since 12 October 2023), or UK standard data protection clauses under UK GDPR Article 46. Since 5 February 2026 (Data (Use and Access) Act 2025), a UK transfer is made subject to appropriate safeguards only if, besides safeguards such as those clauses, the exporter, acting reasonably and proportionately, considers that the "data protection test" is met: protection after the transfer must not be materially lower than under UK data-protection law. The EU General Court’s 3 September 2025 judgment on the EU-US Data Privacy Framework (Case T-553/23, Latombe v Commission) is under appeal at the EU Court of Justice (Case C-703/25 P, pending). Check that the vendor’s DPA names the transfer mechanism for each. Rule of thumb: transfer frameworks can change, so prefer a DPA that also includes standard contractual clauses, ask the vendor for its UK transfer assessment, and re-check the official case and Commission pages before relying on the framework. | - | Legal rule | [t1-eu-gdpr], [t1-eu-us-dpf], [t1-uk-us-bridge], [t1-uk-gdpr-46], [t1-curia-c703-25], [t1-eu-c703-25-oj] |
| The user thinks replacing names with codes makes the data anonymous. | Under GDPR Recital 26, pseudonymised data that could be linked back to a person with extra information (such as your key of codes to names) is still personal data for you, because you hold the key; only data where the person is no longer identifiable is outside the law. Rule of thumb: redaction lowers risk but does not by itself move amber or red data to green. | - | Legal rule | [t1-eu-gdpr] |
| The business is in India, or processes digital personal data of people in India. | Under section 8 of the DPDP Act 2023 the business (Data Fiduciary) will stay responsible for processing done on its behalf by a Data Processor, may use a processor only under a valid contract, must take reasonable security safeguards to prevent a personal data breach, and must tell the Data Protection Board and each affected person about a breach. Commencement notification G.S.R. 843(E), published on 13 November 2025, brings most of the Act into force on 13 May 2027 (eighteen months after publication), including all of section 8 and the other duties of businesses; section 6(9) (registration of consent managers) and section 27(1)(d) on 13 November 2026 (one year after); the definitions (section 2) and the sections setting up the Data Protection Board (sections 18 to 26) have applied since publication; the Board’s powers to inquire and impose penalties start on 13 May 2027. The DPDP Rules 2025 follow the same phasing: rules 1, 2 and 17 to 21 from publication, rule 4 (consent managers) after one year, and rules 3, 5 to 16, 22 and 23 (including security safeguards and breach notice) after 18 months; the government calls this an eighteen-month period for phased compliance. Rule of thumb: "not yet in force" is not permission to paste. The traffic light in this check does not depend on these dates, so put processor contracts and safeguards in place now. Rule of thumb: the government can change these dates by a new notification or by amending the Rules, so confirm the exact dates on MeitY’s website or in the Official Gazette before relying on them. | - | Legal rule | [t1-in-dpdp-act], [t1-in-dpdp-commencement], [t1-in-dpdp-rules], [t1-in-pib] |
| The business provides health services in the UAE (including free zones). | Under Federal Law No. 2 of 2019 on ICT in health fields, health data must be kept confidential, and health information and data related to health services provided in the UAE may only be stored, processed, generated or transferred outside the UAE in cases set by a Health Authority decision in coordination with the Ministry. Ministerial Resolution No. 51 of 2021 sets out those cases, for example data on patients treated abroad within the limits of their treatment, samples sent to laboratories abroad, information insurers need under set conditions, and other data the health authority approves, with conditions that include keeping copies in the UAE. The PDPL does not apply to health data covered by this law. Breaking Article 13 carries a fine of AED 500,000 to AED 700,000. Rule of thumb: unless the vendor confirms in writing that both storage and processing by the AI model stay in the UAE for the exact plan you use, or a lawyer confirms an exception applies, patient data is red for any AI tool. Storing data in the UAE is not the same as processing it there, so check where your plan stores and processes data. | - | Legal rule | [t1-uae-health], [t1-uae-mr51], [t1-uae-pdpl] |
| A UAE healthcare provider’s staff want to use AI on patients’ messages or notes with the names removed. | Rule of thumb: treat anything copied from a patient’s message or record (symptoms, treatment, dates, booking requests), even with the name removed, as health data under the UAE health rule above, so it is red for any AI tool without written in-country or exception confirmation. Green is only a general question staff write fresh, with no detail from a real patient, or a placeholder template. Whether patient text with names removed is still health data under Federal Law No. 2 of 2019 is a question for a UAE health-law lawyer. | - | Heuristic (rule of thumb) | [t1-uae-health] |
| The business is in the UAE, or processes personal data of people in the UAE, and the data is not health data (for a healthcare provider, first apply the rule of thumb on patient text above). | Federal Decree-Law No. 45 of 2021 (the PDPL, in force since 2 January 2022) bans processing personal data without the person’s consent except in limited cases, and sets requirements for sending data abroad. Under its Article 2(1) it applies to a business in the UAE processing data of people inside or outside the UAE, and also to a business outside the UAE processing data of people in the UAE. Businesses in DIFC are supervised under the DIFC Data Protection Law No. 5 of 2020, and those in ADGM under the ADGM Data Protection Regulations 2021; ask a lawyer which law applies to you. Rule of thumb: treat customer and staff data as amber at best, and confirm consent and transfer rules with a UAE lawyer. Under its Article 2, the PDPL does not apply to health personal data or banking and credit data that have their own law, or to companies in free zones that have their own data-protection law (such as DIFC and ADGM). | - | Legal rule | [t1-uae-dp], [t1-difc-dp], [t1-adgm-dp], [t1-uae-pdpl] |
| The business is a US healthcare provider, health plan or clearing house (a HIPAA covered entity), or works for one, and the data is protected health information (PHI). | A covered entity may let a business associate create, receive, maintain or transmit PHI only with satisfactory assurances documented in a written contract, the business associate agreement (BAA). No BAA, no PHI. Rule of thumb: consumer plans (ChatGPT Plus, Claude Pro, personal Gemini) are red. Anthropic’s BAA covers only HIPAA-ready services (its first-party API, Claude in Microsoft Foundry, or Enterprise plans where the Primary Owner activates HIPAA compliance), and excludes listed features such as beta features. Claude Team is not among the HIPAA-ready services Anthropic names. Microsoft says Copilot supports HIPAA for properly configured setups, but not web search queries. Google says Gemini can support HIPAA workloads and is included in its Google Workspace HIPAA Included Functionality, but Gemini Notebook and Gemini in Chrome are not covered by Google’s Business Associate Agreement. OpenAI describes ChatGPT for Healthcare as a secure workspace designed to support HIPAA compliance. Rule of thumb: in every case, get the signed BAA and check that it names the exact product and features you will use. | - | Legal rule | [t1-us-cfr-164502], [t1-an-baa], [t1-ms-edp], [t1-g-workspace], [t1-oa-enterprise] |
| The user is a US lawyer, or handles client information under a professional duty of confidentiality. | The American Bar Association’s Formal Opinion 512 (July 2024) says lawyers using generative AI must keep all information about a client’s matter confidential unless the client gives informed consent, and must understand the benefits and risks of the tool. Rule of thumb: treat client matter information as red unless the tool is a business plan with no training and the client has agreed, and check your own bar or professional body’s rules. | - | Professional guidance | [t2-aba-512] |
| The user prepares US tax returns and the item is a client’s tax return information. | Under Internal Revenue Code section 7216, as described in 26 CFR 301.7216-1, a tax return preparer who knowingly or recklessly discloses or uses tax return information for a purpose other than preparing a tax return commits a misdemeanor (up to one year in prison, a fine of up to $1,000, or both), and section 6713 adds a civil penalty of $250 for each disclosure or use, up to $10,000 a year; the regulations set out exceptions. Tax return information includes the client’s name, address and identifying number given for preparing the return. Rule of thumb: treat client tax return information as red until your tax adviser confirms that the AI use fits an exception in the regulations. | - | Legal rule | [t1-us-cfr-7216] |
| The business or its staff use AI tools in the EU. | EU AI Act Article 4 has applied since 2 February 2025: providers and deployers (businesses that use AI) must take measures to support the AI literacy of staff using AI on their behalf. The Digital Omnibus on AI, in force since July 2026, kept the duty but removed the requirement for a "sufficient" level. National enforcement applies from 2 August 2026. The Commission says staff using ChatGPT for work should be told about its specific risks. Use this checklist as part of that briefing. | - | Legal rule | [t1-eu-ailit] |
| The item is a password, API key, private key, login code, card number with security code, or bank login. | Red in every tool and every plan. Rule of thumb: if it was pasted, treat it as exposed: change the password or rotate the key now, then delete the chat. After rotating, check the key’s usage or audit log with its provider for activity you do not recognise. Replace secrets with placeholders such as [API_KEY] before asking for help with code or config, and keep them in environment variables or a secrets manager, with secret scanning turned on in your code repository. | - | Heuristic (rule of thumb) | - |
| The item is client information covered by a contract or NDA. | Rule of thumb: red until you have read the confidentiality and subcontractor clauses. If the contract forbids sharing with third parties or needs the client’s consent, get the client’s written agreement or redact it to green first. | - | Heuristic (rule of thumb) | - |
| The data belongs to your own customers and you hold or process it for them (for example contact lists in your software, support tickets, logs or database extracts). | Rule of thumb: you are probably their processor. Your contract or DPA with them usually limits what you may do with the data and lists the sub-processors you may use, so an AI tool not on that list, or any personal account, is red. Test and debug with made-up or masked data, and add the AI vendor to your sub-processor list (telling customers first if the contract says so) before using a business plan on real customer data. | - | Heuristic (rule of thumb) | - |
| Sorting any other item. | Rule of thumb. Green: already-public material, your own drafts with no personal or confidential details, made-up or fully anonymised examples (but see the UAE patient rule above), and small, generic code snippets with secrets replaced by placeholders and no customer data, internal hostnames or logs. Amber: names, contact details, orders, internal figures, non-public business plans, or larger parts of your own non-public source code and config, only in a business plan with training off, a signed DPA and admin controls. Red: secrets, special category or health data in a consumer plan, PHI without a BAA, UAE health data without written in-country or exception confirmation, and client data the contract does not allow. | - | Heuristic (rule of thumb) | - |
| The user wants to redact before pasting. | Rule of thumb: replace each name with a role and number ([PATIENT_1], [CLIENT_A]); remove ID numbers, emails, phone numbers, exact addresses, dates of birth and account numbers; round exact dates and amounts; remove rare details that point to one person (a job title in a small firm, an unusual diagnosis); keep the key that links codes to people outside the AI tool. Better still, ask the AI for a template, rubric or formula with placeholders, using column names and made-up rows only, and apply it yourself in the system where the data already lives (your CRM, spreadsheet or clinic system), or use that system’s own features under the contract you already have with its vendor. | - | Heuristic (rule of thumb) | - |
| The user uploads a file rather than pasting text. | Rule of thumb: files carry more than you see (hidden sheets and columns, comments, tracked changes, author names and other metadata). Export only the columns or pages the task needs into a new file. Files saved to projects, custom GPTs or libraries stay until you delete them there; deleting the chat does not delete them. | - | Heuristic (rule of thumb) | - |
| The chat uses connectors or apps (email, Drive, Slack), custom GPTs with actions, memory, or a share link. | Rule of thumb: a connector lets the AI read more than you pasted, and GPT actions or third-party connectors send data to another company under its own privacy terms. OpenAI says such a recipient may keep it longer than 30 days, and Anthropic says sending data to third parties through connectors is not covered by its BAA. Memory carries details into later chats. A share link makes the chat readable by anyone with the link, and Google says deleting a shared chat does not delete its public link. Turn off connectors and memory for chats with amber data, never share links to chats with personal or client data, and delete share links separately. | - | Heuristic (rule of thumb) | [t1-oa-tempchat], [t1-g-activity], [t1-an-baa] |
| Staff use AI on personal accounts for work, or the plan is unknown. | Rule of thumb: treat the account as a consumer plan and allow green items only. Personal accounts are outside the company’s DPA and admin controls even if the person pays for them. | - | Heuristic (rule of thumb) | - |
| The people in the data are in a US state, or in a country not covered here. | Many US states now have their own consumer privacy laws, some with contract rules for vendors; they were not checked one by one. Countries outside the US, UK, EU, UAE and India were not checked. Say so, and treat personal data as amber at best until a local lawyer confirms. | - | Heuristic (rule of thumb) | - |
| Settings or terms may have changed since this check. | Rule of thumb: vendors change defaults (Anthropic moved consumer plans to train when the setting is on in 2025). Re-check the settings pages above every 90 days and whenever the vendor emails about new terms. | - | Heuristic (rule of thumb) | - |

## Procedure

Copyable checklist:

- [ ] Identify the tool and plan: Tool, plan, personal or work account, who is admin, any DPA or BAA signed.
- [ ] Describe the data, not paste it: List the fields or give a made-up example; note whose data it is and where they live.
- [ ] Sort into green, amber, red: Use the rules above; anything special category, health or secret is red unless a rule says otherwise.
- [ ] Fix the settings: Training off, feedback off, memory off for amber data, Temporary or Incognito chat for one-off work, auto-delete shortened; work account with the shield or Workspace sign-in.
- [ ] Fix the contract for amber data: Business plan, signed DPA (and BAA for US health data), transfer mechanism for EU or UK data, admin retention set.
- [ ] Redact or redesign the task: Placeholders, dummy data, or ask for a template and fill in real details outside the AI.
- [ ] Record and brief: Write down the decision, tell staff the red list, and keep it with your AI literacy records.
- [ ] Prepare for the lawyer: Countries checked and not checked, contracts to read, and open questions.

## Output template

```
## AI data safety check: [BUSINESS], [TOOL AND PLAN], [DATE]

Covered: [only those involved among US (HIPAA, professional confidentiality), UK, EU, UAE (mainland, DIFC, ADGM, health ICT law), India]. Not checked: [every other country or US state law involved].

### Traffic light for your data
Item | Whose data | Green / Amber / Red | Why | What to do instead

Key: Green = fine to paste. Amber = only in a business plan with training off, a signed DPA (and BAA for US health data) and admin controls. Red = never paste (secrets, card numbers, special category or health data in consumer plans, PHI without a BAA, client data the contract does not allow).

### Settings to change now
Tool | Where | Setting | Done (yes / no)

### Plan and contract needed for amber items
Need | Have it? | Next step

### Redacted version or safe template
[Placeholder template or redacted text, with no real names or identifiers]

### Questions for your lawyer or the AI vendor
1. [Question]

For data-protection questions, a data-protection lawyer or adviser is the right person.

This is a first-pass check, not legal or tax advice. Confirm with a qualified accountant or lawyer for your country.
```

## Failure modes

| Mistake | Symptom | Fix |
|---|---|---|
| Turning off training on a consumer plan and treating it as safe for customer or patient data. | Personal data sits with a vendor under personal terms, with no data processing agreement and up to 30 days’ retention after deletion. | Training off only lowers risk. Personal data needs a business plan with a signed DPA (and a BAA for US health data); otherwise redact to green. |
| Staff using personal accounts (personal Gmail with Gemini, Copilot without the green shield, a private ChatGPT login) for work. | Work data outside the company’s contract and admin controls, and invisible to retention or audit. | Sign in with the work account; check the shield or Workspace account before pasting; allow only green data on personal accounts. |
| Clicking thumbs up or down on a chat that contains business data. | The whole conversation is sent as feedback: OpenAI may use it for training, Anthropic keeps feedback for 5 years, Google keeps reviewed feedback up to 3 years and, with Keep Activity off, also takes your last 24 hours of chats as context. | Do not rate chats with business data; on Claude Team or Enterprise turn off "Rate chats". |
| Redacting names but leaving ID numbers, dates of birth, phone numbers or rare details. | People can still be identified, so the data is still personal data. | Remove all direct identifiers, generalise dates and amounts, drop rare details, keep the code key outside the AI, or use a template instead. |
| Uploading a whole spreadsheet or document when only part is needed. | Hidden sheets, comments, tracked changes or extra columns reach the AI. | Export only the needed columns or pages to a new clean file. |
| Sharing a chat link, or using a custom GPT, connector or plugin, on a chat with business data. | Customer or client details readable by anyone with the link, or sent to a third-party service outside the AI company’s DPA or BAA. | Do not share links to chats with personal or client data; delete share links separately; turn off connectors and third-party actions for amber data. |
| Pasting code or config files that contain API keys or passwords. | A live secret sits in a third-party chat history. | Rotate the key or change the password at once, delete the chat, and use placeholders next time. |
| Assuming "delete" means gone immediately. | Data still held for up to 30 days, under legal holds such as the 2025 New York Times order, or for 3 years if a Gemini reviewer saw it. | Decide before pasting; deletion is a clean-up, not an undo. |
| Ignoring sector law or client contracts (UAE health data, US PHI, NDAs, professional rules). | A breach of contract or professional duty even where general privacy law looked fine. | Check the sector rule and the contract first; treat the item as red until both allow it. |
| Treating this checklist as legal advice, or relying on old settings screenshots. | A decision based on a setting or term the vendor has since changed. | Rule of thumb: re-check the vendor pages every 90 days, and confirm with a lawyer before using amber data. |

## Verification

- [ ] Every item the user named is in the traffic-light table with a reason.
- [ ] Each tool in use has its settings listed with the menu path from the vendor’s help page.
- [ ] No red item is suggested for any plan, and no amber item for a consumer plan.
- [ ] Countries covered and not checked are listed at the top.
- [ ] A redacted version or a placeholder template is offered for every red or amber item.
- [ ] Volumes and counts come only from the user; if the user gave none, no volume is estimated and no arithmetic is shown.
- [ ] The output says it is a first-pass check, not legal or tax advice, and ends with the confirm line.

## Agent mode

If an AI agent is operating a browser or API on your behalf:

**May do without asking:**
- Sort a description or a redacted sample of the data into green, amber and red.
- Draft a redacted version or a placeholder template from text the user has already redacted.
- Open the vendors’ official help pages and list the settings and their menu paths.
- Draft questions for the AI vendor or a lawyer about a DPA, a BAA or where data is processed.

**Must stop before (the user decides and does these):**
- Uploading or pasting any real file, customer, patient, staff or client data into any AI tool.
- Changing account, privacy, workspace or admin settings (tell the user which toggle to change and let them do it).
- Accepting terms, a DPA or a BAA, or buying, upgrading or changing a plan (any spend or billing).
- Deleting chats, activity, files or accounts.
- Sending or publishing any message drafted from the data, such as patient or customer follow-ups.
- Contacting a vendor, regulator, client or patient on the user’s behalf.

**Setup advice:**
- Give the agent a description or a redacted sample, never the real file or a login to the AI admin console.
- Have a lawyer or your data-protection lead sign off amber uses before staff start.

## Worked example

The figures below are illustrative. Never reuse them: use only the user’s own numbers and facts, and label anything they have not given as an ASSUMPTION.

**Scenario:** A dental clinic owner on the Dubai mainland wants to paste each week’s patient appointment notes into ChatGPT Plus, on a personal account, to draft follow-up messages. The clinic sends about 40 follow-ups a week. No contract has been signed with OpenAI.

**Input:** Tool: ChatGPT Plus, personal account, "Improve the model for everyone" never changed. Data: patient names, phone numbers, visit date, treatment done, next step. Business: health services in Dubai, UAE (mainland, not DIFC or ADGM). Patients live in the UAE. Contracts: none.

**Output excerpt:** Covered: UAE (Federal Law No. 2 of 2019 on ICT in health fields; the PDPL does not apply to health data that has its own law). Not checked: Dubai Health Authority licensing rules and any patient lives abroad. Verdict: RED. These are health data about services in the UAE. Under Federal Law No. 2 of 2019, Article 13, they may only be stored or processed outside the UAE in cases set by a Health Authority decision, and breaking that rule carries a fine of AED 500,000 to AED 700,000. ChatGPT Plus also trains on chats by default, is not one of the plans OpenAI offers its data processing agreement for (ChatGPT Business, Enterprise and the API), and even a Temporary chat may be kept up to 30 days. At the 40 follow-ups a week you gave, that is 40 x 52 = 2,080 patient notes a year leaving the clinic. Safe way: (1) In ChatGPT, turn off "Improve the model for everyone" today anyway. (2) Ask it for three follow-up templates with placeholders only, for example "Hi [FIRST_NAME], thank you for visiting on [DATE]. Your next step is [NEXT_STEP]. Reply to book." No patient details go in, so this is green. (3) Fill the placeholders inside your clinic system or messaging tool. (4) If you want AI to read real notes, ask a UAE health-law lawyer whether one of the cases in Ministerial Resolution No. 51 of 2021 applies and ask vendors in writing whether processing stays in the UAE before any pilot. This is a first-pass check, not legal or tax advice. Confirm with a qualified accountant or lawyer for your country.

## References

### Platform settings

| Platform | Path | Setting | Note | Source |
|---|---|---|---|---|
| ChatGPT (Free, Go, Plus, Pro) | Account menu > Settings > Data controls | Improve the model for everyone: Off | On by default in personal workspaces. Voice mode and Codex have separate training controls. | [t1-oa-datacontrols] |
| ChatGPT | New chat > Temporary (choose Personalized or Unpersonalized before the first message) | Temporary chat for one-off work | Not in history or used for training while temporary; a copy may be kept up to 30 days for safety. | [t1-oa-tempchat] |
| ChatGPT Business, Enterprise, API | OpenAI Enterprise privacy page > Data Processing Addendum form | Sign the DPA before putting personal data in | The page offers the DPA for ChatGPT Business, ChatGPT Enterprise and the API. Admins set workspace retention. | [t1-oa-enterprise] |
| Claude (Free, Pro, Max, Ultra) | Name menu > Settings > Privacy | Help Improve our AI models: Off | With it on, de-identified data can be kept up to 5 years. Use Incognito chats for one-off work. | [t1-an-setting], [t1-an-retention], [t1-an-incognito] |
| Claude Team and Enterprise | Organization settings > Data and Privacy | Rate chats: Off (stops staff sending feedback that can be used for training) | Owner or Primary Owner only. | [t1-an-commercial] |
| Claude Enterprise (US health data) | Organization settings > Data and privacy | Activate HIPAA compliance and accept Anthropic’s BAA (Primary Owner) | Standard Enterprise plans have no BAA coverage until this is done; some features stay outside the BAA. | [t1-an-baa] |
| Gemini (personal Google account) | gemini.google.com > Settings & help > Activity > On > Turn off (or Turn off and delete activity) | Keep Activity: Off | Chats are still kept up to 72 hours. Some connected apps, such as Google Workspace, stop working. | [t1-g-activity], [t1-g-privacyhub] |
| Gemini (personal Google account) | Settings & help > Activity > Deleting activity older than ... months | Auto-delete: 3 months (if you keep activity on) | Default is 18 months. Chats already reviewed by humans are kept up to 3 years regardless. | [t1-g-activity] |
| Microsoft Copilot Chat | Sign in with the work account > look along the top of the screen next to New Chat | Green shield shown (enterprise data protection on) | No shield means personal terms: do not paste work data. | [t1-ms-chat] |
| Microsoft Copilot (admin) | Microsoft Purview > retention policies | Set retention for Copilot interactions | Prompts and responses are stored for audit and eDiscovery in the tenant. | [t1-ms-privacy] |
| Microsoft Copilot (personal Microsoft account, new app) | Copilot > Settings > Personalization | Saved memories: Off; One shared experience: Off; Allow ads personalization: Off | Consumer terms and no DPA, so green data only. Microsoft says the new app does not use prompts, responses or files to train foundation models, and it shows no training switch. Turning memory off does not delete saved memories: use Manage > Delete all memories. | [t1-ms-consumer], [t1-ms-consumer-history] |
| Microsoft Copilot (personal Microsoft account, older app only) | Windows or macOS app: profile icon > Settings > Privacy; mobile app: menu > profile icon > Account > Privacy | Training on conversation activity: Off; Training on voice conversations: Off | Microsoft says this applies only to the app version from before 18 August 2026. Opting out covers future chats only; history is kept for 18 months. | [t1-ms-consumer-old] |

### Sources

| ID | Title | Tier | Accessed on | URL |
|---|---|---|---|---|
| t1-oa-datacontrols | Data controls in ChatGPT (OpenAI Help Center) | T1 | 2026-11-04 | https://help.openai.com/en/articles/7730893-data-controls-in-chatgpt |
| t1-oa-tempchat | Temporary chat in ChatGPT (OpenAI Help Center) | T1 | 2026-11-04 | https://help.openai.com/en/articles/8914046-temporary-chat-in-chatgpt |
| t1-oa-retention | Chat and file retention in ChatGPT (OpenAI Help Center) | T1 | 2026-11-04 | https://help.openai.com/en/articles/8983778-chat-and-file-retention-in-chatgpt |
| t1-oa-enterprise | Enterprise privacy at OpenAI (updated 8 January 2026) | T1 | 2026-11-04 | https://openai.com/enterprise-privacy/ |
| t1-oa-nyt | How we are responding to The New York Times’ data demands (OpenAI, updated 22 October 2025) | T1 | 2026-11-04 | https://openai.com/index/response-to-nyt-data-demands/ |
| t1-an-terms | Updates to Consumer Terms and Privacy Policy (Anthropic, 28 August 2025) | T1 | 2026-11-04 | https://www.anthropic.com/news/updates-to-our-consumer-terms |
| t1-an-retention | How long do you store my data? (Anthropic Privacy Center, consumer plans) | T1 | 2026-11-04 | https://privacy.claude.com/en/articles/10023548-how-long-do-you-store-my-data |
| t1-an-setting | How do I change my model improvement privacy settings? (Anthropic Privacy Center) | T1 | 2026-11-04 | https://privacy.claude.com/en/articles/12109829-how-do-i-change-my-model-improvement-privacy-settings |
| t1-an-incognito | Use incognito chats (Claude Help Center) | T1 | 2026-11-04 | https://support.claude.com/en/articles/12260368-use-incognito-chats |
| t1-an-commercial | Is my data used for model training? (Anthropic Privacy Center, commercial products) | T1 | 2026-11-04 | https://privacy.claude.com/en/articles/7996868-is-my-data-used-for-model-training |
| t1-an-org-retention | How long do you store my organization’s data? (Anthropic Privacy Center) | T1 | 2026-11-04 | https://privacy.claude.com/en/articles/7996866-how-long-do-you-store-my-organization-s-data |
| t1-an-baa | Business Associate Agreements (BAA) for commercial customers (Anthropic Privacy Center) | T1 | 2026-11-04 | https://privacy.claude.com/en/articles/8114513-business-associate-agreements-baa-for-commercial-customers |
| t1-g-privacyhub | Gemini Apps Privacy Hub (Google, last updated 26 October 2026) | T1 | 2026-11-04 | https://support.google.com/gemini/answer/13594961?hl=en |
| t1-g-activity | Manage and delete your activity in Gemini Apps (Google) | T1 | 2026-11-04 | https://support.google.com/gemini/answer/13278892?hl=en |
| t1-g-workspace | Generative AI in Google Workspace Privacy Hub (Google, last updated 6 October 2026) | T1 | 2026-11-04 | https://knowledge.workspace.google.com/admin/generative-ai/generative-ai-in-google-workspace-privacy-hub |
| t1-ms-edp | Enterprise data protection in Microsoft Copilot and Microsoft Copilot Chat (Microsoft Learn, updated 4 November 2026) | T1 | 2026-11-04 | https://learn.microsoft.com/en-us/copilot/plan/enterprise-data-protection |
| t1-ms-chat | Microsoft Copilot Chat privacy and protections (Microsoft Learn, updated 4 November 2026) | T1 | 2026-11-04 | https://learn.microsoft.com/en-us/copilot/plan/privacy-and-protections |
| t1-ms-privacy | Data, privacy and security for Microsoft Copilot (Microsoft Learn, updated 4 November 2026) | T1 | 2026-11-04 | https://learn.microsoft.com/en-us/copilot/plan/microsoft-365-copilot-privacy |
| t1-eu-gdpr | Regulation (EU) 2016/679, General Data Protection Regulation: Recital 26, Articles 6, 9, 28, 33, 44 to 46 (EUR-Lex) | T1 | 2026-11-04 | https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32016R0679 |
| t1-uk-gdpr-9 | UK GDPR Article 9: processing of special categories of personal data (legislation.gov.uk) | T1 | 2026-11-04 | https://www.legislation.gov.uk/eur/2016/679/article/9 |
| t1-uk-gdpr-28 | UK GDPR Article 28: processor (legislation.gov.uk) | T1 | 2026-11-04 | https://www.legislation.gov.uk/eur/2016/679/article/28 |
| t1-eu-us-dpf | EU-US data transfers and the Data Privacy Framework (European Commission) | T1 | 2026-11-04 | https://commission.europa.eu/law/law-topic/data-protection/international-dimension-data-protection/eu-us-data-transfers_en |
| t1-eu-ailit | AI Literacy: questions and answers (European Commission, last update 20 October 2026) | T1 | 2026-11-04 | https://digital-strategy.ec.europa.eu/en/faqs/ai-literacy-questions-answers |
| t1-in-dpdp-act | Digital Personal Data Protection Act, 2023, section 8 (MeitY) | T1 | 2026-11-04 | https://www.meity.gov.in/static/uploads/2024/06/2bf1f0e9f04e6fb4f8fef35e82c42aa5.pdf |
| t1-in-dpdp-commencement | Notification G.S.R. 843(E), 13 November 2025, appointing commencement dates under section 1(2) of the Digital Personal Data Protection Act, 2023 (MeitY, Gazette of India Extraordinary) | T1 | 2026-11-05 | https://www.meity.gov.in/static/uploads/2025/11/c56ceae6c383460ca69577428d36828b.pdf |
| t1-in-dpdp-rules | Digital Personal Data Protection Rules, 2025, G.S.R. 846(E), rule 1 commencement (MeitY) | T1 | 2026-11-04 | https://www.meity.gov.in/static/uploads/2025/11/53450e6e5dc0bfa85ebd78686cadad39.pdf |
| t1-in-pib | DPDP Rules, 2025 notified (Press Information Bureau backgrounder, 17 November 2025) | T1 | 2026-11-04 | https://static.pib.gov.in/WriteReadData/specificdocs/documents/2025/nov/doc20251117695301.pdf |
| t1-uae-health | Federal Law No. 2 of 2019 on the use of ICT in health fields, Articles 2, 4, 13 and 24 (Emirates Health Services copy) | T1 | 2026-11-04 | https://www.ehs.gov.ae/app_content/legislations/php-law-en-77/files/basic-html/page5.html |
| t1-uae-dp | Data protection laws, including Federal Decree-Law No. 45 of 2021 (UAE Government portal u.ae, updated 4 December 2025) | T1 | 2026-11-04 | https://u.ae/en/about-the-uae/digital-uae/data/data-protection-laws |
| t1-difc-dp | Commissioner of Data Protection: DIFC Data Protection Law No. 5 of 2020 (DIFC) | T1 | 2026-11-04 | https://www.difc.com/business/registrars-and-commissioners/commissioner-of-data-protection |
| t1-adgm-dp | Office of Data Protection: ADGM Data Protection Regulations 2021 (ADGM) | T1 | 2026-11-04 | https://www.adgm.com/operating-in-adgm/office-of-data-protection |
| t1-us-cfr-164502 | 45 CFR 164.502(e): disclosures to business associates (eCFR, up to date as of 2 November 2026) | T1 | 2026-11-04 | https://www.ecfr.gov/current/title-45/subtitle-A/subchapter-C/part-164/subpart-E/section-164.502 |
| t2-aba-512 | ABA issues first ethics guidance on a lawyer’s use of AI tools: Formal Opinion 512 (American Bar Association, 29 July 2024) | T2 | 2026-11-04 | https://www.americanbar.org/news/abanews/aba-news-archives/2024/07/aba-issues-first-ethics-guidance-ai-tools/ |
| t1-ms-consumer | Microsoft Copilot for individuals: your privacy controls and choices (Microsoft Support, new Copilot app) | T1 | 2026-11-04 | https://support.microsoft.com/en-us/privacy/microsoft-copilot/privacy-controls |
| t1-ms-consumer-history | Microsoft Copilot for individuals: your activity history (Microsoft Support) | T1 | 2026-11-04 | https://support.microsoft.com/en-us/privacy/microsoft-copilot/activity-history |
| t1-ms-consumer-old | Microsoft Copilot privacy controls (Microsoft Support, older Copilot app only) | T1 | 2026-11-04 | https://support.microsoft.com/microsoft-copilot/microsoft-copilot-privacy-controls |
| t1-an-dpa | How do I view and sign your Data Processing Addendum (DPA)? (Anthropic Privacy Center) | T1 | 2026-11-04 | https://privacy.claude.com/en/articles/7996862-how-do-i-view-and-sign-your-data-processing-addendum-dpa |
| t1-an-covered | Data retention practices for Covered Models (Anthropic Privacy Center) | T1 | 2026-11-04 | https://privacy.claude.com/en/articles/15425996-data-retention-practices-for-covered-models |
| t1-oa-nyt-20m | OpenAI’s Rule 72(a) objections to the orders compelling production of 20 million ChatGPT logs, In re OpenAI Copyright Infringement Litigation, 25-md-03143, ECF 935 (S.D.N.Y., filed 10 December 2025) | T1 | 2026-11-04 | https://cdn.openai.com/pdf/2025-12-10-memorandum-of-law-dckt-935-0.pdf |
| t1-uk-us-bridge | UK-US data bridge: supporting documents (UK Government, gov.uk) | T1 | 2026-11-04 | https://www.gov.uk/government/publications/uk-us-data-bridge-supporting-documents |
| t1-uk-gdpr-46 | UK GDPR Article 46: transfers subject to appropriate safeguards (legislation.gov.uk) | T1 | 2026-11-04 | https://www.legislation.gov.uk/eur/2016/679/article/46 |
| t1-uae-pdpl | Federal Decree-Law No. 45 of 2021 on the Protection of Personal Data, Article 2 (official Arabic text, linked from u.ae) | T1 | 2026-11-04 | https://assets.u.ae/api/public/content/954629a06cc64fd887771ce3273d6154?v=21d88839 |
| t1-uae-mr51 | Ministerial Resolution No. 51 of 2021 on the cases in which health data may be stored or transferred outside the UAE (Ministry of Health and Prevention, last updated 7 February 2026) | T1 | 2026-11-04 | https://uaephl.mohap.gov.ae/en/health-policies-and-legislations-advocacy/health-legislations?itemId=726efdef-580e-4817-bf5a-fbf8e53fddfc |
| t1-court-nyt-20m-order | Order denying OpenAI’s objections and affirming the orders to produce the 20 million log sample, In re OpenAI, Inc. Copyright Infringement Litigation, 25-md-3143, ECF 1021 (S.D.N.Y., 5 January 2026; RECAP copy on CourtListener) | T1 | 2026-11-05 | https://storage.courtlistener.com/recap/gov.uscourts.nysd.640396/gov.uscourts.nysd.640396.1021.0.pdf |
| t1-eu-c703-25-oj | Case C-703/25 P: appeal brought on 31 October 2025 by Philippe Latombe against the judgment of the General Court of 3 September 2025 in Case T-553/23 on the EU-US Data Privacy Framework adequacy decision (Official Journal C/2025/6610, EUR-Lex) | T1 | 2026-11-05 | https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:62025CN0703 |
| t1-curia-c703-25 | Case C-703/25 P, Latombe v Commission: case details, appeal from Case T-553/23 (Court of Justice of the European Union, InfoCuria) | T1 | 2026-11-05 | https://curia.europa.eu/juris/liste.jsf?num=C-703/25 |
| t1-us-cfr-7216 | 26 CFR 301.7216-1: penalty for disclosure or use of tax return information (eCFR, up to date as of 3 November 2026) | T1 | 2026-11-05 | https://www.ecfr.gov/current/title-26/chapter-I/subchapter-F/part-301/subpart-ZZ/section-301.7216-1 |

---
Deepstory methods are process guidance for AI tools. Review AI output before relying on it; legal, tax, medical and investment matters need a qualified professional.
