Is it safe to use the free version of ChatGPT for legal work?

Under OpenAI's current terms, the consumer free and Plus tiers can use your conversations to improve future models unless you turn that setting off, and they come with no data-processing agreement suited to holding client data. Enterprise and team plans change the data-handling terms, but the account settings and the actual contract are what you should read, not the marketing copy. Either way, mask client identifiers before the text reaches any provider. Cedreon screens identifiers with pattern masking and layered best-effort detection before the text leaves, so the provider never sees the real values.

Do I have to tell clients that I used AI on their matter?

For routine low-risk uses like formatting or a first research pass, many bar authorities do not require disclosure, but sending confidential material to an outside tool can trigger a duty to get informed consent. The engagement letter is the usual place to set expectations, and some clients (banks, insurers, government bodies) impose their own AI terms that override your default. Check the rule in your own jurisdiction rather than assuming a blanket yes or no. When it is unclear, get consent in writing before any client data leaves your systems.

Does pasting client details into ChatGPT waive attorney-client privilege?

On some readings, sharing confidential facts with a third-party tool counts as disclosure outside the privileged relationship, which is why careful firms avoid pasting raw client details into a public chatbot. The safer practice is to strip or mask identifiers first, so the provider receives placeholders instead of names, account numbers, or addresses. Cedreon does this before the text is sent, holding the originals in an encrypted vault and filling the reply back in for the signed-in user. Masking is pseudonymization that lowers exposure but does not guarantee that nothing sensitive ever leaves.

Answers

Can lawyers use ChatGPT?

Yes, and many already do. No professional conduct rule prohibits AI as such, any more than one ever prohibited email or cloud storage. The duties that decide the question are the familiar ones: confidentiality, competence and supervision, applied to a new tool. In practice the answer splits by task, and by what you type into the box.

What the professional rules actually say

Conduct rules are technology neutral. The duty of confidentiality covers client information however it is processed, and it doesn't lapse because the recipient is a model rather than a person. Competence includes understanding the tools you rely on well enough to know what they do with your input and where they fail. Supervision covers the work of assistants, and regulators increasingly read software the same way.

Regulators have begun saying this explicitly. The American Bar Association issued its first formal opinion on generative AI in 2024 (Formal Opinion 512). Its themes track the rules:

understand the tool before relying on it
protect client confidences when prompts leave the firm
consider whether informed consent is needed before case facts go into a tool that trains on inputs
verify what comes back before using it

State bars and European regulators have published guidance along the same lines. None of this guidance bans AI; the responsibility for using it safely stays with the lawyer.

Split the tasks honestly

Some uses barely touch the confidentiality duty:

research questions on public law, with no client facts in the prompt
drafting standard language from facts that are already public
summarizing published decisions, legislation or commentary

High risk is just as easy to state: anything that identifies a client, entered into a consumer tool. A consumer chatbot account typically offers no contract with the firm, may use conversations for model training depending on settings, and stores prompts on infrastructure nobody at the firm has reviewed. A prompt naming the client, the counterparty and the dispute is a disclosure to a third party. If it is ever questioned, the lawyer is the one who has to justify it.

Most real work sits between the two: a research question that only makes sense with some case facts attached. That middle band is where your setup starts to matter.

What makes use defensible

Three things, taken together, turn "I used AI" into a position you can defend.

First, the right terms. Use business or enterprise terms that include a data processing agreement and a commitment not to train on your inputs. Where EU or Swiss clients are involved, data protection law runs alongside the professional duty; Is ChatGPT GDPR compliant? covers that side.

Second, minimisation. The less identifying detail in the prompt, the less weight the confidentiality analysis has to carry. Strip or mask names, addresses, account numbers and dates that don't change the legal question. How to anonymize data before using ChatGPT walks through doing this by hand, and where hand methods fall short.

Third, verification. Language models produce fluent, confident text, and sometimes they produce citations to decisions that don't exist. Courts in several jurisdictions have sanctioned lawyers for filing briefs with AI-invented case citations nobody checked. Every authority goes into a citator and every factual claim gets traced to a source before anything reaches a court or a client. The competence duty requires it, and the sanctions decisions show what happens without it.

Write it down

A one-page office policy turns these habits into something you can point to. The policy should cover:

approved tools and account types
what may never be entered anywhere (client identifiers in consumer accounts, privileged analysis)
masking for the middle band
citation checks before filing
who approves a new tool

Date it and have everyone read it. If a client or a court asks how the firm uses AI, you can hand them the policy.

How Cedreon approaches this

Cedreon is an AI workspace for sensitive data, developed in Switzerland, built around the two halves of this page: keep identifiers away from the provider, and be able to prove afterwards what happened. Before any text leaves for a model provider, Cedreon screens it. Names, IBANs, addresses and other identifiers are found by pattern matching plus layered best-effort machine detection and replaced with placeholders. The originals stay in an encrypted vault on Cedreon's side, and the reply is filled back in for the signed-in user, so the provider never sees the real values. Because detection is best-effort rather than certain, it reduces exposure but does not remove it, so do not paste material you could not defend disclosing.

The other half is the record. Every exchange lands in an append-only, hash-chained audit log, so a firm can show afterwards exactly what left the house and what was replaced first. That log is the record a lawyer needs to answer for a disclosure after the fact. Teams can also write guardrails in plain language (block, warn or flag) that run on every chat, which turns the one-page policy into something the software helps enforce. Uploaded Word, Excel and PDF files work the same way, with answers pointing to the source passage. Pricing starts at CHF 29.90 per user per month. Try Cedreon free.

Common questions

Is it safe to use the free version of ChatGPT for legal work?: Under OpenAI's current terms, the consumer free and Plus tiers can use your conversations to improve future models unless you turn that setting off, and they come with no data-processing agreement suited to holding client data. Enterprise and team plans change the data-handling terms, but the account settings and the actual contract are what you should read, not the marketing copy. Either way, mask client identifiers before the text reaches any provider. Cedreon screens identifiers with pattern masking and layered best-effort detection before the text leaves, so the provider never sees the real values.
Do I have to tell clients that I used AI on their matter?: For routine low-risk uses like formatting or a first research pass, many bar authorities do not require disclosure, but sending confidential material to an outside tool can trigger a duty to get informed consent. The engagement letter is the usual place to set expectations, and some clients (banks, insurers, government bodies) impose their own AI terms that override your default. Check the rule in your own jurisdiction rather than assuming a blanket yes or no. When it is unclear, get consent in writing before any client data leaves your systems.
Does pasting client details into ChatGPT waive attorney-client privilege?: On some readings, sharing confidential facts with a third-party tool counts as disclosure outside the privileged relationship, which is why careful firms avoid pasting raw client details into a public chatbot. The safer practice is to strip or mask identifiers first, so the provider receives placeholders instead of names, account numbers, or addresses. Cedreon does this before the text is sent, holding the originals in an encrypted vault and filling the reply back in for the signed-in user. Masking is pseudonymization that lowers exposure but does not guarantee that nothing sensitive ever leaves.