It takes about four seconds. A reviewer opens a 60-page supply agreement in the data room, copies the text, pastes it into a chatbot and asks for a summary of the termination rights. The summary is good. The reviewer has also just moved a confidential document out of a controlled environment into a system the seller never approved, on terms nobody on the deal team has read.
This is now one of the most common confidentiality questions on deals, from both sides. Sellers want to know whether bidders are doing it. Buyers’ advisers want to know whether they may. This guide sets out the issues and the safer alternatives.
Is it allowed to paste data room documents into ChatGPT?
Start with the documents that bind you. Three are relevant.
The NDA. Most confidentiality agreements limit use of the information to evaluating the transaction and limit disclosure to named representatives who are bound by confidentiality. An AI provider is not a representative. Some newer NDAs address AI explicitly, either banning it or permitting it only in tools with no-training terms. Older ones do not mention it, which leaves the question to how a court would read “disclose” and “use”.
The data room terms. Rooms usually require users to accept terms at first login. Those terms often prohibit copying, extracting or transferring content outside the room. Our guide to click-through NDAs in a data room explains how they work and why they are enforceable evidence of what each user agreed.
Your own policies. Many firms, banks and law firms in particular, have AI policies that limit which tools may be used on client data. Breaching them is a professional problem even when the NDA is silent.
If any of the three says no, the answer is no. If all are silent, the answer is still “only in an approved tool”, for the reasons below.
Where does the text go once you paste it?
Where a pasted deal document goes
- 1Paste or upload
Text leaves the data room and its watermark, permissions and audit trail.
High risk - 2Provider servers
Processed by the AI provider, possibly in another country.
Medium risk - 3Chat history
Stored in the account until deleted; visible to anyone with the login.
Medium risk - 4Logs and review
Retained for abuse monitoring; may be seen by provider staff in limited cases.
Medium risk - 5Model training
Consumer accounts may allow training on content unless switched off.
High risk - 6Output reused
A summary or answer is copied into emails, memos or bids without a source check.
Medium risk
Walk the path step by step. The text leaves the data room, so watermarking, permissions, view-only settings and the audit trail no longer apply to it. It travels to the AI provider’s servers, which may be in a different country. It is stored in the chat history of the account. It may be retained in logs for safety monitoring. Depending on the account type and settings, it may be used to improve the provider’s models. Finally, the output is copied into a memo, an email or a negotiating position, often without anyone checking it against the source.
What is the difference between consumer and enterprise AI accounts?
This is the distinction that matters most, and it changes as providers update their terms, so check the current version before relying on it.
| Question | Typical consumer or personal account | Typical enterprise or API account |
|---|---|---|
| Is content used to train models? | May be, unless the user switches it off | Generally not by default, set in the contract |
| Who controls retention? | The user, through settings | The organization, through admin settings and contract |
| Is there a data processing agreement? | No | Usually available |
| Can the organization audit use? | No | Admin logs in most enterprise products |
| Single sign-on and access control | No | Usually |
| Where is data processed? | Provider’s choice | Region options in some products |
OpenAI, for example, states that it does not train on business data from its enterprise products and API by default; its enterprise privacy page sets out the commitments. Other major providers make similar statements for their business tiers. That is a meaningful improvement, but it does not answer the NDA question, and it does not bring back the data room’s controls.
The personal account problem
The highest-risk setup is a deal team member using a personal AI account on a work laptop. The organization has no contract, no logs, no retention control and no way to delete what was pasted. If your firm has not provided an approved tool, assume people are using personal ones.
What makes deal documents more sensitive than ordinary files?
Three categories of information in a data room carry their own legal rules, on top of the NDA.
Material non-public information. If the target or the buyer is listed, deal documents can contain inside information. Insider dealing and selective disclosure rules apply however the information moves. A record of who received it, which the data room’s audit trail provides, is part of how companies manage that risk.
Personal data. Employee records, customer contacts and benefits data are personal data. Sending them to an AI provider is processing by a new processor, possibly with a cross-border transfer. Under the GDPR and similar laws, that needs a basis and a processing agreement. Our guide to privacy laws that touch deal data covers the main regimes.
Competitively sensitive information. Where the buyer is a competitor, some information may only be seen by a clean team. Running it through a tool used by the wider buyer organization can undermine the clean team arrangement.
What can go wrong besides confidentiality?
Accuracy. Large language models summarize well and sometimes invent. A summary that misses an amendment, misstates a cap or adds a clause that is not there can move a negotiation in the wrong direction. Security researchers also track risks specific to language model applications, from prompt injection hidden in documents to leakage of data between sessions; the OWASP Top 10 for LLM applications is a good overview.
There is also a records problem. If an AI-generated summary shaped a decision, there may be no trace of what was asked, what came back and who checked it.
Which options are safer?
| Option | Confidentiality | Controls that still apply | When it fits |
|---|---|---|---|
| Consumer chatbot, personal account | Poor | None | Never for deal documents |
| Enterprise AI tool approved by your firm | Better, with no-training and retention terms | Your firm’s logging and access control | Your own work product; seller documents only if the NDA allows |
| AI features inside the data room | Best of the three | Room permissions, audit trail, residency | Reviewing seller documents where the seller has enabled AI |
| No AI, manual review | Full | All room controls | Highly sensitive folders, clean team material |
AI inside the room keeps documents in place and lets the seller decide whether it is switched on. In our directory, five of sixteen providers list AI features: Ellty, Datasite, Ansarada, Drooms and Box. Our guide to AI in virtual data rooms covers what those features do and the questions to ask before enabling them.
How should sellers address AI in the process?
Sellers can remove the ambiguity rather than wait to find out what bidders did.
- Say it in the NDA. State whether AI tools may be used on the information and, if so, which kinds: for example, only tools under contract with no training on inputs and deletion on request.
- Repeat it in the process letter and the room terms. Users accept the room terms personally, which reaches the individual reviewer.
- Offer an approved route. If the room has AI features, enabling them for reviewers reduces the temptation to export.
- Restrict downloads on sensitive folders. View-only access with screenshot protection makes bulk copying harder, though not impossible.
- Ask for confirmation at the end. Return-or-destroy confirmations from losing bidders can cover AI tools explicitly.
How to set an AI rule for a deal team
Read the binding terms
Check the NDA, the room's click-through terms and your firm's AI policy for anything that restricts AI tools.
Classify the folders
Mark folders that hold personal data, inside information or clean team material as no-AI unless counsel agrees otherwise.
Name the approved tools
List which AI tools may be used, on which material, with enterprise accounts only.
Prefer in-room AI
Where the seller has enabled AI in the data room, use it instead of exporting documents.
Log what was processed
Keep a simple record of documents processed and outputs relied on, with the reviewer who checked each one.
Check outputs against sources
Never rely on a summary for a negotiating point without reading the clause it describes.
Compare data rooms that list AI features alongside permissions, audit trail and two-factor login.
Compare providersFAQ
Can I paste a confidential contract into ChatGPT?
Not on a personal or consumer account, and only in an approved enterprise tool if the NDA, the data room terms and your firm's policy allow it. Pasting moves the document outside the data room's controls.
Does ChatGPT train on what I paste?
It depends on the account type and settings. Consumer accounts may allow training unless switched off; OpenAI states that business and API data are not used for training by default. Check the current terms.
Is using AI on deal documents a breach of the NDA?
It can be. Most NDAs limit use and disclosure to evaluating the deal and to named representatives. Unless the NDA permits AI tools, assume uploading to an outside AI service needs the seller's consent.
What is the safest way to use AI in due diligence?
Use AI features inside the data room where the seller has enabled them, or an enterprise tool approved by your firm with no-training and retention terms, and check every output against the source.
How can sellers stop bidders using AI on the data room?
Address AI in the NDA, the process letter and the room's click-through terms, restrict downloads on sensitive folders, and ask losing bidders to confirm deletion from any AI tools used.