ChatGPT or Claude to review a contract: what to know before relying on them

A general-purpose assistant can read a text it is given, summarize it, answer questions and rephrase. On a contract, this can help identify its structure, list dates and obligations, or prepare the questions to ask the lawyer. These uses remain helpful because their output can be checked by going back to the document.

Things change as soon as the tool has to judge what is acceptable for your company. Without further information, it knows neither your contract policy, nor your negotiating positions, nor the amendments and schedules it was not given. It then answers based only on the text it received and on general knowledge.

Not at this time, based on the independent sources we were able to read. Several comparisons are circulating, but those we found are published by providers that themselves appear in the ranking. We set them aside for that reason.

A study published in August 2026 offers a general perspective. The ContractScrub preprint, submitted to a research workshop, evaluates the final proofreading of contracts drafted by experienced lawyers: misused defined terms, incorrect cross-references, inconsistent wording. Its authors consider the results of the most advanced models weak, with only one reaching an average recall of 0.75. In other words, even the best one misses a significant share of the errors being looked for.

In the abstract we read, this result does not target any particular provider. It also concerns a specific task, and models that will have changed. Above all, it is a reminder that a well-written answer does not show what was missed.

The first distinction to make is not between ChatGPT and Claude, but between individual plans and business plans. Both providers follow a similar logic, subject to their terms in force:

• Individual plans. Conversations may be used for training, depending on a setting left to the user. This setting has to be checked account by account, and it does not replace a company decision.
• Business plans and API. Content is not used for training by default. Anthropic’s commercial terms state this expressly, and OpenAI states it for its Business, Enterprise and API offerings, according to the sources we consulted.
• Exceptions. Voluntary feedback or an enabled sharing option can change the rule. These exceptions therefore need to be read, not just the principle.

No training does not settle everything. You still need to know the retention period, the processing location, the subprocessors and the access controls. These elements vary by plan and can change: they are checked in the terms that apply to your contract.

The CNIL (the French data protection authority) offers a useful method. In its July 2024 guidance on deploying generative AI, it recommends starting from a concrete need, defining authorized and prohibited uses, assessing whether the provider may reuse the data, training users and involving the DPO and the CISO.

These questions do not prohibit use. They allow an informed decision, and asking them once is often enough to set a team rule:

• The contract itself. A non-disclosure agreement or a dedicated clause may limit disclosure to third parties. The AI provider may be one, depending on the case.
• Personal data. Names, contact details, salaries or health data fall under the GDPR. Masking these elements reduces exposure without always removing it.
• Sensitive information. Prices, margins, negotiation strategy and ongoing transactions deserve an explicit rule.
• The account used. A personal account generally falls outside the company’s settings, logs and contract.
• Internal legal opinions. The French law of February 23, 2026 creates confidentiality for the opinions of in-house lawyers, under conditions, including specific training, labeling and filing. It is to enter into force no later than February 1, 2027, by decree. How it fits with these tools deserves to be reviewed with your counsel.

Some errors are easy to check, provided you look for them. They apply to any assistant, general-purpose or specialized:

• The clause that does not exist. The tool describes a provision missing from the document or mixes two versions. Check each point in the text.
• The invented reference. An article or a court decision can be wrong while looking plausible. Open the source on Légifrance or Judilibre. The French court decisions on hallucinations show this risk.
• The wrong law. An answer may rely on another legal system, or on an amended text. Always ask which law and which date are used.
• The omission. This is the hardest error to see, because nothing flags it. A checklist prepared in advance helps spot it.
• Hidden content. A document you receive may contain invisible text meant to influence the AI. Morgan Lewis noted in September 2026 that no product should be presumed to detect this type of injection.

The providers themselves call for this review. OpenAI’s usage policy and the repository of Anthropic’s legal plugins both require the involvement of a professional, each in its own wording.

According to the 2025 barometer by Oxygen+, the AFJE and the Cercle Montesquieu, 87% of the 387 respondents had already used generative AI. Use therefore probably already exists in your team, with or without a written rule. A short framework is better than a ban that no one follows:

• Choose the authorized accounts. State which plans may be used, and with which settings.
• Classify documents. Specify what can be shared as is, what must be masked and what remains excluded.
• Define the uses. Distinguish reading assistance from risk analysis, and reserve approval for a named person.
• Record the essentials. For an important contract, note who checked what. This marker is often enough to prevent an AI answer from becoming a decision by default.
• Review the framework. Terms change: set a review date, at least once a year.

Docform publishes contract software that includes AI: keep this context in mind when reading this article. The assistant works on the open contract, applies your review playbook, flags points of attention and cites the source passage so it can be checked. The lawyer decides, and the AI can make mistakes. Contracts are hosted in France, and the Trust page details the subprocessors, including Azure OpenAI for the assistant.