Is it safe to use ChatGPT in a law firm?

Yes, you can use ChatGPT in a law firm. No rule prohibits it, and the Spanish General Council of Lawyers (Consejo General de la Abogacía Española) has itself recognised that using generative AI to draft legal texts is a lawful practice. The risk is not in the tool: it is in what information you hand over and under what safeguards. A general-purpose AI is designed for general use, not for documents covered by professional secrecy. And that difference isn’t resolved by unticking a box in the settings: it depends on the contract behind it. With anonymised data or made-up scenarios, it is a useful tool. With a client’s case file inside, the question is no longer a technical one.

What really happens when you upload a client document to ChatGPT

When you hit send, the document leaves your computer. It travels to servers that aren’t yours, usually located outside Spain and often outside the European Union.

That is where the part almost nobody looks at begins: the chain has more than one link. It isn’t only the provider whose name you know. Behind it there are infrastructure providers, moderation services and support subcontractors. Each with its own terms and its own jurisdictions. When you press send, you are not handing the document to one company. You are handing it to a chain you cannot audit.

The difference between consumer and business versions is contractual, not technical. The system that answers is practically the same. What changes is what the company commits to doing with what it receives: whether it signs a data processing agreement, whether it commits not to use your data for training, where the data is hosted and how long it is kept. Those commitments exist in corporate plans and do not exist in the same way in an individual account, however much you pay for it. The “don’t use my data for training” toggle is a setting. The contract is something else.

What does it mean for a model to train on what it receives? It means the text you send may be used to adjust the system going forward. It isn’t that your contract will appear word for word in another user’s answer, but it doesn’t disappear either. And here is the underlying problem: not even the companies that build these systems can explain precisely what remains inside them, or how. If nobody can answer that question, you can’t answer it for your client either.

Turning off chat history doesn’t remove the risk either. There is a case that illustrates it without any need for theory. In May 2025, a New York court ordered OpenAI to retain indefinitely all output logs from ChatGPT and its API, overriding its usual deletion policy. The order covered Free, Plus, Pro and Team accounts, and included conversations that users had already deleted. Enterprise and Edu customers, and those using the API under a zero data retention agreement, were excluded. The order was lifted in October 2025, but what was captured during those months remains stored.

The lesson isn’t that ChatGPT is dangerous. It is that the retention terms of a consumer tool are not under your control, and litigation between third parties that has nothing to do with you can override your delete button. Once again, the contract made the difference.

What happened to Samsung in 2023 is also a warning. Several engineers pasted internal source code into ChatGPT to debug it, and the company ended up banning generative AI tools on its corporate devices. A huge security department, and it still happened. A 6-lawyer firm has far less margin.

What the rules require of a law firm that uses AI

4 layers, and it is worth not mixing them up, because they don’t say the same thing.

Data protection. Sending a client’s documents to an AI system is processing of personal data, with everything that entails: a legal basis, a data processing agreement with whoever receives them (Article 28 GDPR), security measures and, if the data leaves the European Economic Area, the safeguards of Chapter V. Of all the principles in play, the most useful in day-to-day practice is data minimisation, in Article 5(1)(c): do I really need to send all of this to get what I’m after? It is a question you can answer on your own, without a data protection adviser, and it solves a good part of the problem before it exists.

Professional secrecy. It doesn’t depend on personal data being involved. Article 542.3 of Spain’s Organic Law of the Judiciary requires lawyers to keep secret all facts or information they learn by reason of any form of their professional practice, and Article 21 of the General Statute of the Spanish Legal Profession sets out the same duty. It is your obligation. You don’t delegate it just because the one who compromises it is a provider.

Professional ethics. This is the new development, and the one that has been least well explained. On 10 April 2026, the Plenary of the Spanish General Council of Lawyers approved Interpretative Circular 3/2026, on the drafting, signing and filing of documents prepared with the assistance of generative AI. It prohibits nothing: it expressly recognises that using it is lawful. What it does is put responsibility back where it always was, with whoever signs. And it goes further than many expected: filing a document with AI-derived errors, where this reveals a lack of diligence and an absence of oversight, may fall under Article 125.u) of the General Statute of the Spanish Legal Profession, which classifies it as a serious infringement. The circular was preceded by the White Paper on Artificial Intelligence and the Legal Profession, presented in January 2026 by the Council itself together with the Valencia Bar Association.

The EU Artificial Intelligence Act. There is a widespread misunderstanding here that is worth clearing up. Since 2 August 2026, most of the transparency obligations in Article 50 of Regulation (EU) 2024/1689 have applied. The so-called Digital Omnibus postponed the full regime for high-risk systems to December 2027 and August 2028, and many people have read that postponement as if the whole regulation had been put back. That is not the case. For a law firm, moreover, the role matters more than the dates: you are normally not the one developing the system but the one using it, and the obligations of each role are not the same.

What you can and can’t do with a general-purpose AI

This is the practical part, and it works today: without switching tools, without signing up for anything and without breaking the confidentiality you owe your client.

Reasonably safe, with no identifiable data:

• Explaining a legal concept or going over general legal doctrine.
• Structuring an argument in the abstract, based on a made-up scenario.
• Rewording, shortening or polishing a text of yours from which you have removed every reference to the case.
• Preparing the outline of a first meeting with a client.
• Translating terminology or reviewing the wording of a generic text.
• Summarising a public piece of legislation, which has nothing confidential about it.

Should not leave the firm:

• Contracts, pleadings and communications involving identifiable parties.
• Documents from ongoing proceedings.
• Data on health, criminal matters or beliefs, or any other special category of data.
• Correspondence with the client.
• Any document for which you can’t answer the question of where it has ended up.

And the trap almost nobody explains: removing the name is not anonymising. The bar is much higher than it seems: as long as the document makes it possible to single out a person, link it to another source or infer who it is from what remains written, you are still processing personal data. Striking out one word doesn’t change that.

What sets an AI designed for law firms apart

A tool built to work with confidential information isn’t distinguished by its brand. It is distinguished by specific conditions, and you can demand them from any provider who sits across the table from you:

• Real isolation between clients. One firm’s documents should not share space with another’s.
• A contractual commitment not to train on what you upload, in writing and not on a help page.
• Hosting under European law, with a signed data processing agreement.
• Identifying data replaced before anything goes out to the model, and by a deterministic procedure, not by another AI. If the anonymisation is done by a model, the document has already left the firm in plain text: the shield arrives too late. And that replacement must be consistent across documents, because if the same person gets a different code in each file, the relationships between them are lost, and those relationships are exactly what a case file is made of.
• Source traceability, so that every statement comes with the document and the exact passage it comes from, and you can check it yourself.

None of this makes a tool infallible, and none detects 100% of the personal data in a text. A name on its own in running text, with nothing around it that points to it, will slip past any system. You need to know that before you start, not find it out afterwards.

In short

The useful question isn’t whether artificial intelligence is safe. It is what information you are handing over and what the recipient has signed.

That is a question you can answer today, with any tool, without changing anything about the way you work. And answering it doesn’t mean giving up on technology. It means doing what you already do with the rest of the firm’s documents: knowing where they are, who can see them and what happens if something goes wrong.

If you’d rather we talk it through with your own documents in front of you, write to us.

Sources

• Interpretative Circular 3/2026 on the Code of Conduct of the Spanish Legal Profession, approved by the Plenary on 10 April 2026 — Consejo General de la Abogacía Española
• White Paper on Artificial Intelligence and the Legal Profession — Consejo General de la Abogacía Española and Ilustre Colegio de Abogados de Valencia, January 2026
• Organic Law 6/1985 of 1 July on the Judiciary, Article 542.3
• Royal Decree 135/2021, General Statute of the Spanish Legal Profession, Articles 21 and 125.u)
• Regulation (EU) 2016/679 (GDPR), Articles 4(5), 5(1)(c) and 28, Chapter V and Recital 26
• Regulation (EU) 2024/1689 on Artificial Intelligence and Regulation (EU) 2026/1744 (Digital Omnibus)
• Opinion 05/2014 on Anonymisation Techniques — Article 29 Working Party
• Guidance and safeguards in procedures for the anonymisation of personal data (Orientaciones y garantías en los procedimientos de anonimización de datos personales) — Spanish Data Protection Agency (AEPD)