Is it legal to put client data into ChatGPT or Claude?

Is it actually legal to paste client data into an AI tool?

There is no law that says “never use AI with client data”. But there are three separate layers of rules, and you need to pass all three. Passing one does not cover the others.

The first is privacy law, which protects information about identifiable people. The second is your professional duty of confidentiality, which protects everything about your client’s affairs, personal or not. The third is your contracts: engagement letters, NDAs and client terms that may restrict where data can go.

The good news is that the same few steps satisfy most of all three. Use the right kind of AI plan, share as little as possible, and check the output before it reaches anyone.

What does privacy law say about putting client data into ChatGPT?

In the EU and UK, the GDPR applies whenever you handle personal data, meaning any information about an identifiable person. You need a ‘lawful basis’, a recognised legal reason, such as performing your contract with the client or a legitimate interest. Health data, ethnicity and similar ‘special category’ data need an extra condition on top.

You also need to know the AI provider’s role. If it only processes data on your instructions, it is your ‘processor’, and Article 28 of the GDPR requires a written contract with it, usually called a data processing agreement or DPA. Consumer AI apps are usually offered on terms where the provider decides how data is used for itself, which is why they are a poor fit for client data. Sending data to a provider outside the EU or UK is also an ‘international transfer’ that needs a legal mechanism, such as standard contractual clauses.

India’s Digital Personal Data Protection Act 2023 follows a similar pattern. You are the ‘data fiduciary’ (the business deciding why data is used) and the AI provider is your ‘data processor’, engaged under a valid contract. The DPDP Rules were notified in November 2025, and most core duties are scheduled to apply from around May 2027.

The US has no general federal privacy law, but sector rules bite hard. Under HIPAA, a cloud service that stores or processes patient health information for a doctor or clinic is a ‘business associate’ and needs a business associate agreement, or BAA. State laws such as California’s CCPA expect contracts with service providers, and Washington’s My Health My Data Act covers health data outside HIPAA. Australia’s privacy regulator, the OAIC, has gone further and recommends not entering personal information, especially sensitive information, into publicly available generative AI tools at all.

What do professional bodies say about AI and client confidentiality?

Professional confidentiality is wider than privacy law. It covers a company client’s deal terms, a building’s security layout or a tax position, even when no individual is named. Most professional bodies now say the same thing in different words: you may use AI, but you stay responsible, and confidentiality comes first.

In the US, the American Bar Association’s Formal Opinion 512, issued on 29 July 2024, says lawyers must understand how a generative AI tool handles data. Before putting client information into a ‘self-learning’ tool that could expose it to others, a lawyer generally needs the client’s informed consent, and boilerplate wording in an engagement letter is not enough. In England and Wales, solicitors’ confidentiality duty sits in the SRA Code of Conduct, and the Bar Council updated its generative AI guidance for barristers in November 2025.

Other professions are catching up. Accountants are bound by the confidentiality principle in their ethics codes, and US tax preparers face specific restrictions on disclosing tax return information without consent. Doctors are bound by medical confidentiality and health data law, though formal regulator guidance on generative AI is still developing in several countries, including from the UK’s GMC. In India, the Bar Council of India had not issued AI-specific rules for advocates as of mid-2026, but the Supreme Court has asked it to frame norms.

Do NDAs and client contracts stop you using AI?

Sometimes. Many NDAs say confidential information may only be shared with your ‘representatives’, such as employees and advisers, who need it for the agreed purpose. An AI provider is a third party, and pasting the material into its system may count as a disclosure.

Other contracts are more specific. Large corporate clients, banks and public bodies increasingly add clauses that ban or limit AI use, require approved tools only, or require you to tell them. Breaking those clauses is a contract issue even if no privacy law is broken.

The fix is simple and boring. Read the confidentiality clause before you start. Where it is unclear, ask the client and record the answer, and update your own engagement letters to say how you use AI.

Consumer vs business AI plans: what actually changes?

The biggest difference between a free account and an enterprise account is not the model. It is the contract. Business plans typically switch training off by default, give admins control over retention, offer a DPA, and come with security certifications such as SOC 2 or ISO 27001. Some also let you choose the region where data is stored.

Paying for a personal plan, such as ChatGPT Plus or Claude Pro, does not change this. Those are still consumer plans, on consumer terms, aimed at individuals. For client data, the line that matters is usually personal versus business, not free versus paid.

Does ChatGPT, Claude, Gemini or Copilot train on what I type?

It depends on the plan and your settings. Here is the position as of October 2026, but providers change their terms often, so check the current version before relying on it.

Even with training switched off, your chats are usually stored for a period, and staff or contractors may review some content for safety. ‘Not trained on’ does not mean ‘not kept’.

• ChatGPT: on consumer plans, the “Improve the model for everyone” setting under Data controls decides whether chats can be used for training. OpenAI says ChatGPT Business (formerly Team), Enterprise, Edu and the API are not used for training by default.
• Claude: from late 2025, Anthropic asked Free, Pro and Max users to choose whether their chats could be used for training. If they allow it, data is kept for up to five years; if not, the 30-day period continues. Claude for Work (Team and Enterprise) and the API are outside that change.
• Gemini: personal accounts are governed by the “Keep Activity” setting, formerly Gemini Apps Activity. Google says some chats may be reviewed by people, and even with the setting off, chats can be kept for up to 72 hours. Gemini in Google Workspace business accounts comes with separate enterprise terms.
• Microsoft Copilot: Microsoft 365 Copilot and Copilot Chat used with a work account come with ‘enterprise data protection’, and Microsoft says prompts and responses are not used to train its foundation models. The consumer Copilot app is a different product with different terms.

Is anonymising client data enough?

It helps a lot, but only if it is real anonymisation. Under Recital 26 of the GDPR, data that has been ‘pseudonymised’, where names are replaced by codes or initials but could be linked back, is still personal data. Only data that can no longer identify anyone, taking into account all means reasonably likely to be used, falls outside the law.

In practice, removing a name is rarely enough. A job title, a town, a deal value, a rare medical condition or a date can identify a person or company, especially when combined. The UK ICO’s anonymisation guidance makes the same point.

Anonymising also does nothing for your NDA if the commercial information itself is confidential. Treat it as one layer of protection, not a licence to use any tool.

How regulators and courts reacted in 2024 to 2026

The first wake-up call came from business. In 2023 Samsung was widely reported to have restricted staff use of generative AI after engineers pasted internal source code into ChatGPT. Many large firms followed with their own bans, then moved to approved enterprise tools.

Courts focused on accuracy. In Mata v. Avianca (2023), a New York federal judge fined two lawyers and their firm $5,000 for filing fake cases invented by ChatGPT. Similar sanctions, from fines to referrals to regulators, followed in many later US cases. On 6 June 2025, the High Court of England and Wales warned in Ayinde v Haringey that freely available tools like ChatGPT are not capable of conducting reliable legal research, and that citing fake cases can lead to regulatory referral or even contempt proceedings. In July 2026, India’s Supreme Court called citing unverified AI-generated precedents misconduct and asked the Bar Council of India to frame norms.

Regulators, meanwhile, moved from warnings to guidance. Professional bodies for lawyers in the US, UK, Ireland and Australia issued or updated generative AI guidance in 2024 and 2025, almost all of it stressing confidentiality, verification and supervision. Privacy regulators such as Australia’s OAIC told organisations to keep personal information out of public AI tools. The direction of travel is clear: AI use is accepted, but careless AI use is now treated as a professional failing.