AI data privacy risk checker
Check if your current AI workflow violates data privacy standards or risks corporate leakage.
A guide from published rules and terms, checked on 10 October 2026. Not legal advice.
On this page
What the result means
The risk level comes from the kind of data and the kind of service, by the decision tree under How it works.
- Critical: sending sensitive data to a consumer model usually violates compliance laws. The model provider may train on this data and expose it to others.
- High: your data is not secure. You are risking data leakage by sending information to a platform without explicit zero-retention guarantees.
- Medium: you are relying on a third-party vendor's promises. Ensure your legal agreements specifically cover this data class.
- Low: your data architecture is secure for this workload.
- With a plan picked, the tool shows that plan's own terms and anything that rules it out, and with a country, what the law there asks before the data goes.
How it works
The tool evaluates the intersection of your data payload and the AI vendor's architecture. Most consumer plans use your inputs as training data unless you turn it off. Enterprise APIs offer opt-outs but still process data on external servers. Private models retain the data entirely within your own firewall.
Whether a service trains on your data, how long it keeps it and whether it signs a business associate or data processing agreement differ by plan more than by vendor, so the tool reads each plan's own terms.
- A private self-hosted model: Low Risk.
- A public consumer chat with personal data, health data or financial data: Critical Risk.
- A public consumer chat with internal company data: High Risk.
- An enterprise API not opted out of training: High Risk.
- An enterprise API opted out, with health or financial data: Medium Risk (Requires BAA/Compliance checks).
- An enterprise API opted out, with internal company data or personal data: Low Risk.
- Anything else: Medium Risk.
Assumptions
- The tree measures the service's design, not your contract: a Low Risk enterprise API still needs the agreements the law asks for, listed by country.
- Opted out means the service doesn't train on your data, whether by its own terms or a setting you've changed. Picking a plan fills it in from the plan's terms where they're clear.
- Self-hosted means on machines you control. A model rented on someone else's cloud is a processor again for the law, if not for training.
- Plan terms are as each vendor's own pages state them on the date checked; vendors change them, so check the current terms before relying on one.
The plans, by vendor
What each plan does with your data, from the vendor's own pages.
OpenAI
OpenAI ChatGPT Free, Go, Plus or Pro (ChatGPT data controls; ChatGPT Temporary Chat; OpenAI enterprise privacy; OpenAI BAA)
- Trains on your data
- Yes, unless turned off. Conversations may train OpenAI's models unless you turn off "Improve the model for everyone" in Settings > Data controls; Temporary Chats never train.
- Kept for
- Kept until you delete them; deleted and Temporary Chats are removed within 30 days, longer under a legal hold.
- Zero data retention
- Not offered
- HIPAA business associate agreement
- Not offered
- Data processing agreement
- Not offered
OpenAI ChatGPT Business (OpenAI enterprise privacy; OpenAI BAA)
- Trains on your data
- No. Not used for training unless the workspace opts in.
- Kept for
- Admins control retention; deleted chats are removed within 30 days unless the law requires longer.
- Zero data retention
- Not offered
- HIPAA business associate agreement
- Not offered
- Data processing agreement
- Offered
OpenAI ChatGPT Enterprise or Edu (OpenAI enterprise privacy; OpenAI BAA; ChatGPT data residency)
- Trains on your data
- No. Not used for training by default.
- Kept for
- Workspace admins set retention; deleted chats are removed within 30 days unless legally required.
- Zero data retention
- Not offered
- HIPAA business associate agreement
- Offered
- Data processing agreement
- Offered
- Where it's kept
- Storage in Australia, Canada, Europe, India, Japan, Singapore, South Korea, the UAE, the UK or the US for new customers; in-region inference in Europe, the US and the UAE.
OpenAI API (OpenAI; OpenAI enterprise privacy; OpenAI BAA)
- Trains on your data
- No. Not used for training since March 2023 unless you opt in.
- Kept for
- Abuse-monitoring logs kept for up to 30 days.
- Zero data retention
- Zero data retention or modified abuse monitoring, with OpenAI's prior approval.
- HIPAA business associate agreement
- On request
- Data processing agreement
- Offered
- Where it's kept
- Storage by project in the US, the EU, Australia, Canada, Japan, India, Singapore, South Korea, the UK or the UAE; in-region processing in the US, the EU and the UAE.
Anthropic
Anthropic Claude Free, Pro or Max (Anthropic consumer terms, 2025; Claude consumer retention; Anthropic BAA; Anthropic DPA; Anthropic server locations; Anthropic data protection)
- Trains on your data
- Only if you allow it. Used to improve the models only if you allow it, a choice made at sign-up and changed in Settings > Privacy; Anthropic doesn't say which way it starts.
- Kept for
- With model improvement off, deleted chats go within 30 days; with it on, up to 5 years; flagged chats up to 2 years.
- Zero data retention
- Not offered
- HIPAA business associate agreement
- Not offered
- Data processing agreement
- Not offered
- Where it's kept
- Stored in the US.
- Who reads it
- Staff can't access conversations by default, except feedback you send or a usage policy review.
Anthropic Claude Team (Anthropic Privacy Center; Anthropic BAA; Anthropic DPA; Anthropic server locations; Anthropic commercial retention)
- Trains on your data
- No. Not used for training unless you send feedback or opt in.
- Kept for
- Kept while the organisation keeps the chats; deleted chats go within 30 days.
- Zero data retention
- Not offered
- HIPAA business associate agreement
- Not offered
- Data processing agreement
- Offered
- Where it's kept
- Stored in the US; traffic may be routed through the US, Europe, Asia and Australia.
Anthropic Claude Enterprise (Anthropic Privacy Center; Anthropic BAA; Anthropic DPA; Anthropic server locations; Claude Enterprise retention)
- Trains on your data
- No. Not used for training unless you send feedback or opt in.
- Kept for
- The organisation can set custom retention periods.
- Zero data retention
- Not offered
- HIPAA business associate agreement
- Offered
- Data processing agreement
- Offered
- Where it's kept
- Stored in the US; traffic may be routed through the US, Europe, Asia and Australia.
Anthropic Claude API (Anthropic; Anthropic; Anthropic Privacy Center; Anthropic BAA; Anthropic DPA; Anthropic covered models)
- Trains on your data
- No. Not used for training.
- Kept for
- Inputs and outputs deleted within 30 days.
- Zero data retention
- For eligible models by agreement; the Covered Models (Fable and Mythos) keep 30-day retention whatever the agreement.
- HIPAA business associate agreement
- Offered
- Data processing agreement
- Offered
- Where it's kept
- US-only inference available at 1.1 times the price.
Google Gemini app, personal account (Gemini Apps Privacy Hub)
- Trains on your data
- Yes, unless turned off. Used to improve Google's products while Keep Activity is on; turn it off, or use Temporary Chats.
- Kept for
- Activity auto-deletes after 18 months by default; with Keep Activity off, 72 hours; reviewed chats are kept up to 3 years.
- Zero data retention
- Not offered
- HIPAA business associate agreement
- Not offered
- Data processing agreement
- Not offered
- Who reads it
- A subset of chats is read by human reviewers, including service providers.
Google Gemini in Google Workspace (Workspace generative AI privacy; Google HIPAA; Workspace data regions)
- Trains on your data
- No. Not human reviewed or used to train models outside your domain without permission.
- Kept for
- Admins set conversation history from 3 months to indefinite.
- Zero data retention
- Not offered
- HIPAA business associate agreement
- Offered
- Data processing agreement
- Offered
- Where it's kept
- Workspace data regions cover Gemini prompts and responses (US or Europe).
- Who reads it
- No human review without permission.
Google Gemini API, free tier (Google)
- Trains on your data
- Yes, unless turned off. Prompts and responses are used to improve Google's products and human reviewers may read them, except in the EEA, Switzerland and the UK, where the paid terms apply.
- Kept for
- Not stated beyond product improvement.
- Zero data retention
- Not offered
- HIPAA business associate agreement
- Not offered
- Data processing agreement
- Not offered
- Who reads it
- Human reviewers may read prompts and responses.
Google Gemini API, paid tier (Google; Google)
- Trains on your data
- No. Not used to improve Google's products.
- Kept for
- Abuse-monitoring logs kept for 55 days.
- Zero data retention
- Not offered
- HIPAA business associate agreement
- Not offered
- Data processing agreement
- Not published
Google Gemini on Google Cloud (Agent Platform, formerly Vertex AI) (Google HIPAA; Agent Platform zero retention)
- Trains on your data
- No. Not used to train Google's models.
- Kept for
- Logged for abuse monitoring unless zero retention is configured.
- Zero data retention
- By configuration, with an exception process for abuse logging.
- HIPAA business associate agreement
- Offered
- Data processing agreement
- Not published
Microsoft
Microsoft Copilot, personal account (app from 18 August 2026) (Copilot for individuals)
- Trains on your data
- No. Prompts, responses and files aren't used to train foundation models.
- Kept for
- Chat history kept until you delete it.
- Zero data retention
- Not offered
- HIPAA business associate agreement
- Not offered
- Data processing agreement
- Not offered
Microsoft Copilot, personal account (older app) (Copilot privacy FAQ (older app))
- Trains on your data
- Yes, unless turned off. Used for training by default in most countries; you can opt out.
- Kept for
- Activity kept for 18 months by default.
- Zero data retention
- Not offered
- HIPAA business associate agreement
- Not offered
- Data processing agreement
- Not offered
- Who reads it
- Some conversations get automated and human review.
Microsoft Copilot with a work account (Copilot enterprise data protection)
- Trains on your data
- No. Prompts, responses and your organisation's data aren't used to train foundation models.
- Kept for
- Not stated on the page cited.
- Zero data retention
- Not offered
- HIPAA business associate agreement
- Offered
- Data processing agreement
- Offered
Microsoft Azure Foundry Models (formerly Azure OpenAI) (Foundry data privacy; Azure HIPAA)
- Trains on your data
- No. Not used to train models, and not available to OpenAI.
- Kept for
- Content isn't stored for abuse review once modified abuse monitoring is approved.
- Zero data retention
- Modified abuse monitoring for approved customers.
- HIPAA business associate agreement
- Offered
- Data processing agreement
- Not published
Amazon
Amazon Bedrock (Amazon Web Services; Amazon Web Services; AWS HIPAA eligible services; Bedrock abuse detection)
- Trains on your data
- No. Neither AWS nor the model providers train on your inputs or outputs.
- Kept for
- Not stored by default, except 30 days for models whose providers require abuse review.
- Zero data retention
- A no-retention mode for an account or project; models that require retention are then blocked.
- HIPAA business associate agreement
- Offered
- Data processing agreement
- Offered
- Where it's kept
- Stored in the Region you use; geographic routing stays in that geography.
- Who reads it
- Only flagged traffic on models that require it, by AWS, never the model provider.
Mistral AI
Mistral AI Vibe (formerly Le Chat) Free or Pro (Mistral AI; Mistral training; Mistral storage; Mistral zero retention)
- Trains on your data
- Yes, unless turned off. Used by default to train models unless you opt out in settings.
- Kept for
- Kept until you delete the conversation or account.
- Zero data retention
- Not offered
- HIPAA business associate agreement
- Not published
- Data processing agreement
- Not offered
- Where it's kept
- Hosted in the EU by default.
Mistral AI Vibe Enterprise (Mistral AI; Mistral training; Mistral storage; Mistral zero retention)
- Trains on your data
- No. Opted out of training by default; an admin can opt in.
- Kept for
- Kept until deleted.
- Zero data retention
- Not offered
- HIPAA business associate agreement
- Not published
- Data processing agreement
- Offered
- Where it's kept
- Hosted in the EU by default.
Mistral AI Studio API, pay as you go (Mistral AI; Mistral AI; Mistral training; Mistral zero retention)
- Trains on your data
- Can be turned off. Can be switched off at any time with the "Anonymous improvement data" setting; Mistral doesn't say it starts off.
- Kept for
- Kept 30 rolling days to monitor abuse, unless zero retention applies.
- Zero data retention
- For stateless endpoints, on request.
- HIPAA business associate agreement
- Not published
- Data processing agreement
- Offered
- Where it's kept
- EU by default; processing kept in the EU or the US on the regional endpoints.
xAI
xAI Grok, personal account (Grok consumer FAQ)
- Trains on your data
- Yes, unless turned off. May train models unless you turn off "Improve the model"; Private Chat never trains.
- Kept for
- Deleted chats go within 30 days unless needed for legal, compliance or safety reasons.
- Zero data retention
- Not offered
- HIPAA business associate agreement
- Not offered
- Data processing agreement
- Not offered
- Who reads it
- A limited number of authorised staff may review conversations.
xAI API (xAI; xAI)
- Trains on your data
- No. Not used for training without explicit permission.
- Kept for
- Requests and responses stored encrypted for 30 days, then deleted.
- Zero data retention
- Switched on for a team in the console.
- HIPAA business associate agreement
- On request
- Data processing agreement
- Not published
- Where it's kept
- A US-only endpoint for some models.
DeepSeek
DeepSeek App or API (DeepSeek)
- Trains on your data
- Can be turned off. Used to train its models; its privacy policy gives a right to opt out without naming the setting.
- Kept for
- Kept for as long as you have an account.
- Zero data retention
- Not offered
- HIPAA business associate agreement
- Not offered
- Data processing agreement
- Not offered
- Where it's kept
- Collected, processed and stored in the People's Republic of China.
Meta
Meta AI (Meta on AI training)
- Trains on your data
- Yes, unless turned off. Your questions and queries train Meta's models; only people in the EU can object, by form.
- Kept for
- Not stated.
- Zero data retention
- Not offered
- HIPAA business associate agreement
- Not offered
- Data processing agreement
- Not offered
Perplexity
Perplexity Free, Pro or Max (Perplexity data collection)
- Trains on your data
- Yes, unless turned off. AI Data Retention is on by default; turn it off in settings.
- Kept for
- Not stated.
- Zero data retention
- Not offered
- HIPAA business associate agreement
- Not published
- Data processing agreement
- Not published
Self-hosted
Open-weight model on your own machines (DSK guidance on AI)
- Trains on your data
- No. Nothing leaves your control unless you send it somewhere.
- Kept for
- Whatever your own systems keep.
- Zero data retention
- Not offered
- HIPAA business associate agreement
- Not offered
- Data processing agreement
- Not offered
- Where it's kept
- Wherever your machines are.
Example
Example: patient records in a consumer chat
Protected health information pasted into a standard consumer AI chat:
Sending Protected Health Information (PHI) to a standard consumer AI chat interface is a critical risk. Even if the output is helpful, the input has crossed your compliance boundary and may have entered a third-party training set.
Example: customer emails through an API in Germany
Customers' names and emails sent to the OpenAI API, which doesn't train on API data by default, from a business in Germany:
The tree reads Low Risk, and 3 of Germany's rules still apply first: a processor contract; sending data to the US; breaches and fines.
When this is wrong
This is a high-level architectural checker, not legal advice. A Low Risk score on an Enterprise API still requires your legal team to verify Business Associate Agreements (BAA) and vendor compliance certifications.
- Sector rules. Banks, insurers and health systems have their own rules on where data may go, beyond privacy law; the country rules here name the main ones, not all.
- Contracts with your own clients. A client agreement can bar sending its data to any third party, whatever the law allows.
- Plan changes. Vendors rename plans and change terms; Microsoft, Google, Mistral and Anthropic all did in 2025 and 2026.
- Data in the output. What a model writes back can carry personal data too, and needs the same care.
- Employees on personal accounts. The plan your company buys doesn't protect data staff paste into their own accounts.
To price running a model yourself, use the private vs public AI cost estimator.
What we couldn't confirm
- Official source unavailable. Quebec's private sector act as consolidated on LégisQuébec, which is down; the regulator's consolidation of March 2023 is used.
- Not published. How long Meta AI keeps chats, and whether people review them.
- Not published. Whether Gemini's Keep Activity setting starts on or off.
- Not published. How long Azure keeps abuse-monitoring logs for its Foundry models.
- Official source unavailable. Perplexity's BAA and DPA terms: its trust centre doesn't load.
- Not found in official sources. xAI's data processing agreement.
- Official source unavailable. The English text of the UAE's health ICT law on storing health data in the country.
- Not final yet. The GDPR changes in the EU's Digital Omnibus, and the appeal against the EU-US Data Privacy Framework.
Rules by country
For each country in the tool: the contract a processor needs, the extra rules for health and financial data, sending data abroad, breach deadlines, fines, and what the regulator has said about AI tools.
United States
Under HIPAA for health data and the GLBA Safeguards Rule for financial data; there's no general federal privacy law:
- A business associate agreement for PHI. A vendor may receive protected health information only under a written business associate agreement that sets its permitted uses and bars any other, model training included; a cloud provider needs one even if it only stores encrypted data without the key. (45 CFR 164.502(e), 164.504(e); HHS, HIPAA and cloud computing)
- A contract for financial customer data. A financial institution under the FTC must vet any service provider given customer information and bind it by contract to keep safeguards. (16 CFR 314.4)
- Breach deadlines. A breach of unsecured PHI is notified to the people affected within 60 days of its discovery; an FTC-regulated financial institution reports one involving 500 or more consumers to the FTC within 30 days. (45 CFR 164.404; 16 CFR 314.4)
- Penalties. HIPAA civil penalties run to $73,011 per violation, and $2,190,294 for uncorrected wilful neglect. (45 CFR 102.3)
- A promise not to train is enforceable. The FTC treats an AI provider's promise not to train on customer data as binding: breaking it, or quietly changing the terms, can be unfair or deceptive. (FTC on AI privacy commitments)
Canada
Under PIPEDA, and Quebec's private sector act for Quebec:
- You stay accountable. An organisation stays responsible for personal information it sends to a third party for processing, and must secure comparable protection by contract. (PIPEDA)
- Health and income records. Medical and income records are almost always sensitive, needing stronger consent and safeguards. (PIPEDA)
- Leaving Quebec. A Quebec business completes a privacy impact assessment and signs a written agreement before personal information goes to a provider outside Quebec; penal fines reach $25 million or 4% of worldwide turnover. (Québec private sector act)
- Breaches and fines. A breach with a real risk of significant harm is reported to the Privacy Commissioner and the people affected as soon as feasible; PIPEDA fines reach $100,000. (PIPEDA)
- What the regulators expect of generative AI. Enter sensitive information only where you're authorised to, and make sure prompts aren't kept or reused for purposes such as training. (OPC generative AI principles)
United Kingdom
Under the UK GDPR and the Data Protection Act 2018, as amended by the Data (Use and Access) Act 2025:
- A processor contract. Using any processor, an AI vendor included, needs a written contract with the Article 28 terms: instructions, confidentiality, security, sub-processors, deletion and audits. (ICO processor contracts)
- Training makes the vendor a controller. The ICO's view is that a provider that uses your prompts to improve its own models is a controller for that, not your processor, so the disclosure needs its own lawful basis. (ICO on generative AI controllership)
- Health data. Health data needs an Article 9 condition, for care by or under a professional bound by confidentiality; financial data isn't special category data. (ICO on special category data)
- Sending data abroad. A provider outside the UK means a restricted transfer: UK adequacy (US companies certified to the UK Extension of the Data Privacy Framework count), an IDTA or the Addendum. (ICO on transfers)
- Breaches and fines. A notifiable breach is reported to the ICO within 72 hours; fines reach £17.5 million or 4% of worldwide turnover. (ICO breach guide)
Germany
Under the GDPR and the Bundesdatenschutzgesetz:
- A processor contract. A cloud AI provider needs an Article 28 contract, and the data protection authorities advise choosing services that don't train on your inputs. (GDPR; DSK guidance on AI)
- Health data. Health data is a special category: processing needs an Article 9 condition and, for care, people bound by secrecy with safeguards such as access limits, logging and encryption. (GDPR; § 22 BDSG)
- Professional secrecy. Doctors, lawyers, tax advisers and insurers may share secrets with a service provider only as far as its service needs, and commit an offence if they don't oblige it to secrecy. (§ 203 StGB)
- Health data in the cloud. In the statutory health system, health data may be processed in the cloud only in Germany, the EU or an adequate country, by a provider with a German establishment and a current BSI C5 attestation. (§ 393 SGB V)
- Sending data to the US. The EU-US Data Privacy Framework stands after the General Court upheld it on 3 September 2025, so certified US providers can receive data without other safeguards. (Latombe judgment)
- Breaches and fines. A breach is reported to the authority within 72 hours; fines reach €20 million or 4% of worldwide turnover. (GDPR)
India
Under the Digital Personal Data Protection Act 2023, whose core duties apply from 13 May 2027:
- A processor contract. From 13 May 2027 a Data Fiduciary may use a processor only under a valid contract with security safeguards, and stays responsible for what it does. (DPDP Act 2023; DPDP commencement; DPDP Rules 2025)
- Payment data stays in India. All payment system data must be stored only in India; it can be processed abroad if it's deleted there within one business day or 24 hours. (RBI payment data storage)
- Sending data abroad. Transfers abroad are allowed unless the government restricts a country, and stricter sector rules still apply. (DPDP Act 2023)
- Breaches and fines. From 13 May 2027 a breach is reported to the Data Protection Board in detail within 72 hours; the top penalty is ₹250 crore. (DPDP Act 2023; DPDP Rules 2025)
Singapore
Under the Personal Data Protection Act 2012:
- Sending data abroad. Personal data sent overseas must get protection comparable to the PDPA's, by law, contract or binding rules. (PDPA 2012)
- Banks and the cloud. MAS treats cloud services as outsourcing: a bank assesses the risk before customer information goes to an overseas provider. (MAS outsourcing guidelines)
- Breaches and fines. A notifiable breach is reported to the PDPC within 3 calendar days of assessing it; fines reach 10% of Singapore turnover above S$10 million, otherwise S$1 million. (PDPA 2012)
Philippines
Under the Data Privacy Act of 2012:
- Sensitive personal information. Health, genetic and government-issued ID information is sensitive; processing it is prohibited unless an exception such as specific consent applies. (Data Privacy Act 2012)
- You stay accountable. The controller stays accountable for personal information it transfers, at home or abroad, and must protect it by contract. (Data Privacy Act 2012)
- Banks and the public cloud. Banks may use public cloud only for non-core work that doesn't directly involve sensitive customer data. (BSP Circular 808)
- Breaches. A personal data breach is reported to the National Privacy Commission and the people affected within 72 hours, with the full report to the Commission within 5 days. (NPC Circular 16-03)
Malaysia
Under the Personal Data Protection Act 2010, as amended in 2024:
- Processors and transfers. Since 1 April 2025 processors must meet the Security Principle themselves, and data may go abroad to places with substantially similar law or adequate protection. (PDPA amendments 2024)
- Banks and the cloud. A financial institution consults Bank Negara Malaysia before its first use of public cloud for critical systems. (BNM RMiT)
- Breaches and fines. Since 1 June 2025 a breach is notified to the Commissioner within 72 hours and to the people affected within 7 days; the general penalty is up to RM1 million. (PDPA amendments 2024; PDP breach notification guideline)
United Arab Emirates
Under the federal data protection law, with the DIFC's and ADGM's own laws in those free zones:
- Sending data abroad. Under the federal law, personal data may go to a country with data protection law in cases the UAE Data Office approves, or elsewhere under a contract binding the recipient to the law's standards. Controllers have six months from the law's executive regulations to comply, and the federal legislation portal lists none so far. (UAE PDPL (English))
- Bank data stays in the UAE. Banks keep customer data in the UAE, and sharing it abroad needs the central bank's approval and the customer's written consent. (CBUAE outsourcing regulation)
- AI in the DIFC. DIFC firms deploying AI give a clear notice at first use, and high-risk processing through AI waits for the Commissioner's certification requirements. (DIFC Regulation 10)
- ADGM. ADGM firms report breaches within 72 hours where feasible, send data out only under an adequacy decision or other safeguards, and face fines up to USD 28 million. (ADGM Data Protection Regulations)
Australia
Under the Privacy Act 1988 and its Australian Privacy Principles:
- Don't use public AI tools for personal information. As a matter of best practice, the OAIC recommends not entering personal information, and particularly sensitive information, into publicly available generative AI tools. (OAIC guidance on AI products)
- You answer for the overseas provider. An entity that discloses personal information to an overseas recipient is accountable for that recipient's breaches of the Privacy Principles. (Privacy Act 1988)
- My Health Record data stays in Australia. Records held for the My Health Record system must not be held, processed or handled outside Australia. (My Health Records Act, s. 77)
- Regulated financial firms. Banks, insurers and super trustees cover information held by third parties and report material security incidents to APRA within 72 hours. (APRA CPS 234)
- Penalties. A serious interference with privacy can draw the greatest of $50 million, three times the benefit or 30% of adjusted turnover. (Privacy Act 1988)
Anywhere else
Under your country's privacy law:
- Check four things. Whether a vendor processing personal data needs a written contract, whether health or financial data may leave the country, how fast a breach must be reported, and what the regulator has said about AI tools.
Sources
Every source this page cites (85)
Each one was opened and checked on the date beside it.
- Office of the Australian Information CommissionerGuidance on privacy and the use of commercially available AI products17 January 2025
- DatenschutzkonferenzOrientierungshilfe Künstliche Intelligenz und Datenschutz, Version 1.06 May 2024
- eCFR45 CFR 164.502(e) and 164.504(e): disclosures to business associates and business associate contracts7 October 2026
- OpenAIData controls in ChatGPT29 September 2026
- OpenAITemporary Chat FAQ20 September 2026
- OpenAIGetting a Business Associate Agreement for the OpenAI API11 October 2026
- OpenAIEnterprise privacy at OpenAI8 January 2026
- OpenAIData residency for ChatGPT4 October 2026
- OpenAIYour data10 October 2026
- AnthropicUpdates to consumer terms and privacy policy28 August 2025
- AnthropicHow long do you store my data?1 July 2026
- AnthropicBusiness Associate Agreements (BAA) for commercial customers10 October 2026
- AnthropicHow do I view and sign your Data Processing Addendum (DPA)?16 March 2026
- AnthropicWhere are your servers located? Do you host your models on EU servers?15 June 2026
- AnthropicHow does Anthropic protect the personal data of Claude users?16 March 2026
- Anthropic Privacy CenterIs my data used for model training?18 August 2026
- AnthropicHow long do you store my organisation's data?1 July 2026
- AnthropicConfigure custom data retention controls for Enterprise plans10 October 2026
- AnthropicAPI and data retention10 October 2026
- AnthropicData retention practices for covered models5 September 2026
- AnthropicClaude API pricing10 October 2026
- GoogleGemini Apps Privacy Hub24 September 2026
- GoogleGenerative AI in Google Workspace Privacy Hub6 October 2026
- Google CloudHIPAA compliance on Google Cloud and Google Workspace7 October 2026
- GoogleData covered by data regions9 October 2026
- GoogleGemini API additional terms of service28 April 2026
- GoogleZero data retention in the Gemini API14 September 2026
- Google CloudGemini Enterprise Agent Platform and zero data retention7 October 2026
- MicrosoftMicrosoft Copilot for individuals: your activity history10 October 2026
- MicrosoftPrivacy FAQ for Microsoft Copilot (older app)10 October 2026
- MicrosoftEnterprise data protection in Microsoft Copilot and Microsoft Copilot Chat18 August 2026
- MicrosoftData, privacy and security for Foundry Models sold by Azure5 June 2026
- MicrosoftHIPAA (Azure compliance offering)5 April 2023
- Amazon Web ServicesAmazon Bedrock FAQs10 October 2026
- Amazon Web ServicesData retention in Amazon Bedrock10 October 2026
- Amazon Web ServicesAmazon Bedrock abuse detection10 October 2026
- Amazon Web ServicesHIPAA eligible services reference3 September 2026
- Mistral AIDo you use my user data to train your artificial intelligence models?31 August 2026
- Mistral AIPrivacy policy3 September 2026
- Mistral AIWhere do you store my data or my organisation's data?12 August 2026
- Mistral AICan I activate Zero Data Retention (ZDR)?12 August 2026
- Mistral AIRegional inference10 October 2026
- xAIConsumer FAQs (Grok)11 October 2026
- xAISecurity FAQ10 October 2026
- xAIRegional endpoints10 October 2026
- DeepSeekPrivacy policy10 February 2026
- MetaMaking AI work harder for Europeans27 March 2026
- PerplexityData collection at Perplexity3 March 2026
- US Department of Health and Human Services, Office for Civil RightsGuidance on HIPAA and Cloud Computing23 December 2022
- eCFR16 CFR 314.4: elements of an information security program (GLBA Safeguards Rule)7 October 2026
- eCFR45 CFR 164.404: notification to individuals of a breach of unsecured PHI7 October 2026
- eCFR45 CFR 102.3: civil monetary penalty adjustments (HIPAA)7 October 2026
- Federal Trade CommissionAI companies: uphold your privacy and confidentiality commitments9 January 2024
- Justice Laws Website, Government of CanadaPersonal Information Protection and Electronic Documents Act21 September 2026
- Commission d'accès à l'information du QuébecLoi sur la protection des renseignements personnels dans le secteur privé (version administrative)14 March 2023
- Office of the Privacy Commissioner of CanadaPrinciples for responsible, trustworthy and privacy-protective generative AI technologies7 December 2023
- Information Commissioner's OfficeWhat needs to be included in the contract? (contracts between controllers and processors)10 October 2026
- Information Commissioner's OfficeResponse to the consultation series on generative AI: allocating controllership12 December 2024
- Information Commissioner's OfficeWhat are the conditions for processing? (special category data)10 October 2026
- Information Commissioner's OfficeA brief guide to international transfers15 January 2026
- Information Commissioner's OfficePersonal data breaches: a guide20 August 2025
- EUR-LexRegulation (EU) 2016/679 (General Data Protection Regulation)4 May 2016
- Bundesministerium der Justiz§ 22 BDSG: Verarbeitung besonderer Kategorien personenbezogener Daten10 October 2026
- Bundesministerium der Justiz§ 203 StGB: Verletzung von Privatgeheimnissen10 October 2026
- Bundesministerium der Justiz§ 393 SGB V: Cloud-Einsatz im Gesundheitswesen10 October 2026
- Court of Justice of the European UnionJudgment of the General Court in Case T-553/23, Latombe v Commission (press release)3 September 2025
- Gazette of IndiaThe Digital Personal Data Protection Act, 202311 August 2023
- Gazette of IndiaNotification G.S.R. 843(E): commencement of the Digital Personal Data Protection Act, 202313 November 2025
- Gazette of IndiaDigital Personal Data Protection Rules, 2025 (G.S.R. 846(E))13 November 2025
- Reserve Bank of IndiaStorage of payment system data (RBI/2017-18/153) and its FAQ6 April 2018
- Singapore Statutes OnlinePersonal Data Protection Act 20122026
- Monetary Authority of SingaporeGuidelines on outsourcing (banks)11 December 2023
- Lawphil Project (Arellano Law Foundation)Republic Act No. 10173, Data Privacy Act of 201215 August 2012
- Bangko Sentral ng PilipinasCircular No. 808, series of 2013: guidelines on IT risk management for all banks22 August 2013
- National Privacy Commission, PhilippinesNPC Circular No. 16-03: Personal Data Breach Management15 December 2016
- Personal Data Protection Department, MalaysiaPersonal Data Protection (Amendment) Act 2024 (Act A1727) and its commencement order2024
- Bank Negara MalaysiaRisk Management in Technology (RMiT)25 September 2026
- Personal Data Protection Department, MalaysiaGuideline on data breach notification25 February 2025
- UAE Legislation (Ministry of Cabinet Affairs)Federal Decree-Law No. 45 of 2021 concerning the Protection of Personal Data (English text)20 September 2021
- Central Bank of the UAEOutsourcing regulation for banks15 July 2021
- Dubai International Financial CentreData Protection Regulations, Regulation 10: processing through autonomous and semi-autonomous systems1 September 2023
- Abu Dhabi Global MarketData Protection Regulations 20212025
- Federal Register of Legislation, AustraliaPrivacy Act 19882026
- Federal Register of Legislation, AustraliaMy Health Records Act 2012, section 771 July 2026
- Australian Prudential Regulation AuthorityPrudential Standard CPS 234 Information Security1 July 2019
Frequently asked questions
Is standard ChatGPT safe for customer data?
No. By default, consumer interfaces use your prompts to train future models: on ChatGPT's personal plans conversations may train OpenAI's models unless you turn it off, and those plans carry no data processing agreement or business associate agreement. ChatGPT Business, Enterprise and the API don't train on your data by default. Australia's regulator goes further and recommends not entering personal information into publicly available AI tools at all.
Does an API key guarantee privacy?
Not automatically. While enterprise API tiers often exclude your data from training by default, they still keep it for a while: up to 30 days at OpenAI, to monitor abuse, and at Anthropic, and 55 days on Google's paid Gemini API. Zero retention is something you ask for, and OpenAI grants it only with prior approval; Amazon Bedrock stores nothing by default. Google's free Gemini API tier does train on what you send.
What is the safest way to use AI on sensitive data?
Hosting an open-weight model on your own private infrastructure. If the model runs on your machines, the data never crosses the public internet to a third-party vendor. Germany's data protection authorities say closed, self-hosted systems are preferable for the same reason.
Do I need a BAA to use AI with patient data?
In the US, yes: a vendor that receives protected health information is a business associate, and needs a written agreement that bars any other use. Plans that offer one: OpenAI ChatGPT Enterprise or Edu, OpenAI API, Anthropic Claude Enterprise, Anthropic Claude API, Google Gemini in Google Workspace, Google Gemini on Google Cloud (Agent Platform, formerly Vertex AI), Microsoft Copilot with a work account, Microsoft Azure Foundry Models (formerly Azure OpenAI), Amazon Bedrock, xAI API. Plans that don't: OpenAI ChatGPT Business, Anthropic Claude Team, Google Gemini API, free tier, Google Gemini API, paid tier, and every personal plan.
Which AI plans train on your data by default?
These train on your data by default unless you turn it off: OpenAI ChatGPT Free, Go, Plus or Pro, Google Gemini app, personal account, Google Gemini API, free tier, Microsoft Copilot, personal account (older app), Mistral AI Vibe (formerly Le Chat) Free or Pro, xAI Grok, personal account, Meta AI, Perplexity Free, Pro or Max. Anthropic Claude Free, Pro or Max and Mistral AI Studio API, pay as you go and DeepSeek App or API let you choose or switch it off. The other business plans and APIs here don't by default, and nor does Microsoft's newer consumer Copilot.
Related
Keep your data behind your own firewall.
We deploy open-weight models directly into your environment so you never have to send sensitive data to a third party.
Get a readiness checkBuilt and maintained by Stafflo. Questions or a wrong number: contact. Last updated 11 October 2026.