Can ChatGPT Be Used Against You? How AI Chat Logs Become Evidence

Yes. Using ChatGPT logs as evidence is now ordinary litigation practice, and a ChatGPT conversation can be used against you in civil and criminal matters alike. A prompt and the response to it are electronically stored information (ESI), which makes a ChatGPT log discoverable on the same relevance and proportionality grounds as an email thread, and no privilege attaches to a conversation held on a consumer platform. For in-house legal teams, the useful question is not whether the log is reachable but which of three procedural routes reaches it, and what the log proves once it arrives.

Below: how a ChatGPT log is obtained, what a court requires before admitting it, what the log proves as evidence, what the OpenAI litigation has already put into discovery, and what a litigation hold has to name. Every section names the rule that governs it, because using ChatGPT logs as evidence is a procedural exercise before it is a technical one.

[cta-1]

Key Takeaways

Are ChatGPT chats discoverable as evidence? They are, on the same terms as an email thread. A ChatGPT chat is discoverable evidence, and the procedural route a litigant picks changes the cost, the speed and who gets to object.

1. A prompt and its output are ESI. They are discoverable when relevant and proportional, and no AI-specific privilege shields them.
2. Three routes reach a chat log: a Rule 34 request to the opposing party, a Rule 45 subpoena to the provider, and a device seizure or forensic image. The Rule 34 route to the party is the fastest and the cheapest.
3. Discoverable is not admissible. A log still has to be authenticated under the Federal Rules of Evidence, and the hearsay analysis differs for the prompt and for the output.
4. A preservation duty reaches ESI rather than a list of named systems, so a hold notice that names email and chat platforms but not AI accounts leaves the newest category of ESI unpreserved.
5. A log is strong evidence of state of mind and weak evidence of what the underlying transaction actually said.
6. Provider-held logs are reachable at scale. A court has already ordered OpenAI to produce 20 million de-identified ChatGPT output logs, under conditions rather than into the open record.

Are ChatGPT Conversations Discoverable in Court?

ChatGPT conversations are discoverable in court when they are relevant to a claim or defense and sit within a party's possession, custody or control. Rule 34(a)(1) of the Federal Rules of Civil Procedure reaches "data or data compilations stored in any medium from which information can be obtained," which describes a chat log precisely. An account holder who can open a conversation and export it holds that conversation for discovery purposes, whether the account is a personal one or a company-provisioned one.

Nothing about the technology changes the analysis. Courts have treated AI prompts and outputs as another category of ESI rather than as a new kind of record with its own rules, which means the familiar objections are the only objections available.

Proportionality Is the Limit, Not Privilege

The real limit on a request for a counterparty's chat history is proportionality. Rule 26(b)(1) confines discovery to matter that is relevant and proportional to the needs of the case, and a court weighing a chat-log request applies the same factors it applies to a request for a mailbox:

1. The importance of the issue the logs go to, and whether those facts are available more cheaply elsewhere.
2. The amount in controversy measured against the burden of collecting and reviewing the conversations.
3. The period the request covers, and whether it is tied to the matter rather than to the account's whole life.
4. The specificity of the search, including whether the request names topics, counterparties or date ranges.
5. Whether the party has already produced the underlying documents, which often makes the conversations duplicative.

A request narrowed to the matter, the counterparty and the relevant months tends to survive; a request for an entire conversation history draws a proportionality objection and usually deserves one.

What Confidential Data Exposure in ChatGPT Means for Lawyers

Confidential data exposure in ChatGPT matters to a lawyer for one discovery reason above the others: material typed into a consumer account is held by a third party on its systems, so it exists somewhere the account holder does not control and cannot unilaterally make disappear. Retention schedules, plan-level carve-outs, encryption posture and the list of things never to enter are the subject of the confidentiality guide for ChatGPT, and none of that material is repeated here, because the mechanics of storage are a different question from the reach of a document request.

The consequence for this page is narrower and harder. Because the conversation is a disclosure to a third party, it does not carry the protection a note to counsel would. The waiver analysis and the case law on it belong to the privilege treatment of ChatGPT conversations; the point here is that the log is producible.

Can ChatGPT Chat History Be Subpoenaed?

ChatGPT chat history can be subpoenaed, and a subpoena is usually the worst of the three routes available. Litigants reach a chat log in one of three ways, and they differ in speed, in cost and in who gets to object:

1. A Rule 34 request to the opposing party: the log is in that party's possession, custody or control, so it is produced like any other ESI. No third party, no motion practice, no standing fight.
2. A Rule 45 subpoena to the provider: slower, open to a motion to quash, and it reaches a copy the party may not control.
3. A device seizure or forensic image: reaches the stored conversations without involving the provider at all.

Why the Provider Route Costs the Most

The Rule 34 route is the one to start with. A party served with a document request has a duty to search its own records and limited grounds to refuse, while a non-party served under Rule 45 has every incentive to object. A litigant who opens with the subpoena has selected the route with the most friction and the least control over scope.

When the Subpoena Route Is Still Right

The provider route is correct when the party's own copy is gone or its completeness is contested. OpenAI states that it discloses user content when compelled by subpoena, court order or search warrant, and its published law enforcement policy sets out the process it applies. What the rules permit a litigant to request is a separate question from how a provider answers a particular demand.

How a Device Seizure Produced the First Federal Ruling

The seizure route is not hypothetical. In United States v. Heppner, No. 1:25-cr-00503-JSR (S.D.N.Y.), thirty-one documents the defendant had generated with a consumer AI platform were seized from his home on November 4, 2025, and Judge Jed S. Rakoff ruled from the bench on February 10, 2026 that they carried neither attorney-client privilege nor work-product protection. As the analysis of the ruling records, the court treated what the defendant had typed as independent research later shared with counsel, and noted that a public-facing platform owes its users no duty of loyalty or confidentiality. Which platform it was is not stated in the reports of the ruling.

Are ChatGPT Chat Logs Admissible in Court?

ChatGPT chat logs are admissible in court only after the proponent clears two gates that discovery does not touch: authentication and the hearsay rules. Discoverable and admissible are different questions. A log can be produced in discovery, used in a deposition and still be kept out at trial.

How a Chat Log Is Authenticated

Authentication comes first, and there are three practical paths to it. Rule 901(b)(4) of the Federal Rules of Evidence allows a record to be authenticated by its distinctive characteristics, which for a chat log means the account identifier, the timestamps, the conversation title and the writing itself:

1. Testimony from the account holder, or from a custodian who exported the conversation, that the exhibit is what it purports to be.
2. A written certification from a qualified person, which removes the need for live testimony.
3. A hash value or export manifest generated at collection (a fingerprint of the file recorded when it was captured), which ties the exhibit back to the file that was actually collected.

Self-Authentication Under Rule 902

Certification is the path worth planning for. Rule 902(13) makes a record self-authenticating when it is "generated by an electronic process or system that produces an accurate result, as shown by a certification of a qualified person," and Rule 902(14) does the same for data copied from a device and authenticated by digital identification. A collection done with a certification in mind is cheaper at trial.

Why a Prompt and an Output Differ Under the Hearsay Rules

The hearsay posture is where the analysis becomes genuinely unfamiliar, and it splits along the prompt and the output. Rule 801(d)(2) is the provision that does the work:

1. The prompt: a party's own prompt is that party's own statement offered against it, which Rule 801(d)(2) excludes from the definition of hearsay altogether.
2. The output: the model's response is not a person's statement, so it is not hearsay either, and the questions that remain are relevance and what the exhibit is offered to prove.

A response is therefore evidence that the account holder was told something. It is not evidence that the thing is true.

Privilege and Work Product After Heppner

Privilege and work product are the one place a chat log differs from an email, and Heppner is the first federal answer. The ruling, which one analysis frames as a discovery and confidentiality question rather than a technology one, treated documents created outside counsel's direction on a third-party platform as unprotected, and Judge Rakoff called it a question of first impression. The doctrinal reasoning sits with the privilege treatment linked above.

Proposed Rule 707 and What It Does Not Reach

Proposed Federal Rule of Evidence 707 would subject machine-generated evidence to a reliability gate modeled on Rule 702. As of the most recent published account, the proposal had been published for public comment through February 16, 2026, with the Advisory Committee on Evidence Rules due to deliver its final report at the June 2026 Standing Committee meeting. Rule 707 addresses output offered for the truth of what a machine concluded, while a chat log offered as a party's own statement travels the ordinary ESI route above.

What a ChatGPT Log Actually Proves

What ChatGPT logs prove in litigation is narrow and specific. A chat log is strong evidence of what a person was thinking and weak evidence of what actually happened. It proves three things reliably:

1. State of mind: a prompt records what someone was worried about, what they were trying to accomplish and what they already knew, which is why logs surface in intent, scienter and good-faith disputes rather than in proving the terms of an agreement.
2. Timeline: a sequence of prompts with server-side timestamps places a person's attention on a subject at a time, which can corroborate or contradict a declaration about when an issue was recognized.
3. Impeachment material: a prompt that describes the facts one way is useful against a later pleaded position that describes them another way, and the inconsistency does not have to be material to be damaging.

What a Log Does Not Prove

A log does not establish the content of the underlying transaction. A model's answer about a termination-for-convenience clause establishes nothing about the clause, and a prompt describing a counterparty's conduct is characterization rather than proof.

The Second Exposure: Unchecked Output in a Filing

There is a second exposure that is not evidentiary at all. Where an AI output reaches a filing without being checked, the consequence arrives through the court's own supervisory powers rather than through the rules of evidence, and the sanctions record on fabricated AI citations documents that risk.

[cta-2]

What the OpenAI Litigation Has Already Put Into Discovery

A court has already ordered OpenAI to produce ChatGPT logs, which answers the question of whether provider-held conversations can be reached at scale. In the consolidated copyright litigation brought by The New York Times and other publishers, Magistrate Judge Ona T. Wang ordered OpenAI to produce 20 million de-identified ChatGPT output logs, and the order and its conditions were upheld on review. Judge Wang noted that the sample is only a small fraction of the "tens of billions" of consumer logs OpenAI retains.

The Conditions the Court Imposed

The order did not force a binary choice between a litigant's need and a user's privacy. The court imposed conditions instead:

1. De-identification of the logs before production.
2. Production under an existing protective order rather than into the open record.
3. An attorneys'-eyes-only designation restricting who on the receiving side may review them.
4. A sample rather than the full set, limited to output logs and bounded by the court's own order.

How the Order Was Affirmed

On January 5, 2026, Judge Sidney H. Stein affirmed, rejecting OpenAI's comparison of the logs to wiretapped telephone calls on the ground that ChatGPT users had voluntarily submitted their communications to the platform. Voluntary submission to a third-party service is the same premise that makes a log producible from the party who typed it.

What a Retention Policy Should Take From It

A vendor's deletion schedule is a commercial commitment, not a legal guarantee, and legal process can suspend it. Two consequences follow for a retention policy:

1. OpenAI publishes a thirty-day window for deleting removed conversations, and an earlier preservation order in the same litigation required it to keep logs that would otherwise have been deleted.
2. A policy cannot treat a lapsed preservation order as the end of the exposure, because the live question in this litigation is a production order rather than a preservation one.

Does a Litigation Hold Cover ChatGPT Accounts?

A litigation hold covers ChatGPT accounts only if it says so. The duty to preserve attaches when litigation becomes reasonably foreseeable and it reaches ESI as a category, not a list of named systems. A notice that enumerates email, Slack and Teams but stops there leaves the newest category of ESI outside the instruction, and the omission is the drafter's rather than the custodian's.

Which Accounts a Hold Notice Must Name

Account ownership is the distinction to put first, because it changes the notice this week, and the first two cases point in opposite directions on the same facts:

1. A company-provisioned workspace account sits within the company's control. An administrator can suspend deletion, preserve conversations and export them, so it belongs in the hold notice and the data map behind it.
2. A personal account an employee opened for work does not sit within the company's control for Rule 34 purposes. That is not protection: the company cannot preserve what it cannot reach, the employee can delete it, and the resulting gap is the company's to explain.
3. A former employee's account of either kind is why the exit checklist and the hold template need to reference each other.
4. Shared or team workspaces need a named custodian, or preservation falls to whoever happens to notice.
5. Accounts reached through a browser extension or a plugin can leave a second copy outside the primary workspace, so the inventory step comes before the notice step.
6. Accounts held by a contractor or an outside vendor sit outside the employee lifecycle entirely, so the preservation instruction travels through the engagement terms instead.
7. Accounts opened with a personal email but used on company devices are the hardest to inventory and the easiest to miss.

Why Personal Accounts Are the Harder Problem

That asymmetry is a discovery argument against unmanaged bring-your-own-AI, and it is an argument about preservation rather than confidentiality. A company with provisioned accounts has a problem it can solve; a company whose staff use personal accounts has one it can only describe.

The Rule 37(e) Consequence

Rule 37(e) supplies the consequence. Where ESI that should have been preserved is lost because reasonable steps were not taken and cannot be restored or replaced, a court may order measures to cure the prejudice, and on a finding that a party acted with intent to deprive another of the information it may instruct the jury that it may presume the information was unfavorable. The sanction turns on the steps taken, which is why the hold notice and the actions under it should be documented.

How to Avoid Creating a Discoverable ChatGPT Record

Lawyers can use ChatGPT safely only by controlling what record the use creates. Deleting the conversation is not the fix, because after a preservation duty attaches it converts a disclosure problem into a spoliation problem. Safe use is therefore a routing question settled before a record exists, and whether the tool is permitted at all is addressed in the guidance on lawyers using ChatGPT.

Six Controls That Reduce the Record

These six controls do most of the work, ordered by how much exposure each one removes:

1. Provision accounts centrally: a workspace account under administrative control can be preserved, exported and governed. A personal account can be none of those things.
2. Route matter work to tools under a contract: where the engagement terms include zero data retention, the record a consumer account would have created does not come into existence.
3. Name AI accounts in the hold template and the data map, so preservation does not depend on someone remembering the category.
4. Retain prompts and outputs in the firm's document management system rather than relying on a vendor's chat history as the system of record.
5. Start a new session for each matter instead of continuing a prior conversation, which keeps one matter's facts out of another matter's record.
6. Adopt a written AI use policy covering permitted uses, prohibited categories of information, incident handling and retention, and train staff against it.

A court assessing reasonable steps looks at what the organization can show.

Where Routing Goes Wrong

Routing goes wrong in the other direction too, when contract work goes to a general assistant because it is open in another tab. What a consumer tool can and cannot do on an agreement is treated in the assessment of ChatGPT for contract review, and the gap between those capabilities and a reviewed contract is where the record gets created.

Where Spellbook Fits

Spellbook is safer than ChatGPT for confidential contract work in one specific sense: confidential contract work runs under commercial terms that address what happens to the material, rather than on a consumer account that disclaims confidentiality. It is an AI platform for commercial legal work that runs inside Microsoft Word, and the Review feature surfaces risks and proposes redlines as native Track Changes in Word, so a draft under review stays in the document and the matter file rather than in a chat history nobody has mapped.

The security posture is the part that bears on this article. The platform operates under SOC 2 Type II and zero data retention, so contract text submitted for review is not retained for model training and the engagement is governed by a contract rather than a consumer policy. Those are commitments about handling, and they are checkable.

Grounding review in a firm's own contract playbook changes the quality of the output and the location of the record. It does not change the reach of Rule 34, and that distinction is the one to hold on to when any tool in this category is weighed against a discovery problem that already exists.

What Spellbook Does Not Do

Three limitations matter more than the capabilities on a page about discovery:

1. Scope: Spellbook is a specialist tool for commercial contract work rather than a general assistant, so the legal research or litigation work a lawyer might otherwise take to a chatbot sits outside what it does.
2. Records that already exist: no AI tool removes a record from discovery once it exists. A tool can change where work is done and what is retained going forward; it does nothing about a conversation that has already happened.
3. Judgment: the platform supports review. Whether a term is acceptable on a given deal stays with the lawyer.

For a team weighing a general assistant against something built for legal work, the comparison of Harvey and ChatGPT covers the broader category question.

Frequently Asked Questions

Does Deleting a Chat Remove It From Discovery?

No. Deleting a chat does not remove it from discovery, and after a duty to preserve has attached the deletion is itself a problem. Provider-side copies, backups and preservation orders can outlive a user-initiated deletion, and the question a court asks is what reasonable steps were taken once litigation became reasonably foreseeable.

Can an Employer Produce an Employee's ChatGPT History?

An employer can produce an employee's ChatGPT history when the account is company-provisioned, because an administrator holds the access that makes it within the company's possession, custody or control. A personal account the employee opened independently is generally not within the company's control for Rule 34 purposes. However, the employee may still be subpoenaed and the company may still have to account for the gap.

Can You Request a Counterparty's ChatGPT Logs in a Contract Dispute?

Yes, a counterparty's ChatGPT logs can be requested in a contract dispute, subject to relevance and proportionality. A request tied to the negotiation period, the agreement at issue and named topics is far more likely to survive than one for the account's full history. The logs tend to be most useful on intent and knowledge questions rather than on the terms, which the documents establish.

Can What I Type into ChatGPT Be Subpoenaed in Court?

Yes. Depending on jurisdiction and facts, prompts and outputs can be treated as ESI and discovered on familiar relevance and proportionality grounds. There is no special AI privilege.

Is It Safe to Use ChatGPT for Legal Drafting or Contracts?

Public tools can help with brainstorming. However, they may misinterpret queries. Legal counsel should supervise and verify all outputs.

[cta-3]

What to Do Before the Next Matter Opens

What to do about discoverable ChatGPT records is a question to settle before the next matter opens. A chat log becomes an exhibit through ordinary procedure, and five things are worth settling now:

1. Decide which route the organization would use to obtain a counterparty's logs, and which route it should expect to be used against it.
2. Add AI accounts to the hold template and the data map, and name a custodian for every shared workspace.
3. Plan collection with a certification in mind, so authentication does not require a witness at trial.
4. Brief the team that deleting a conversation once a dispute is foreseeable creates a second problem rather than a solution.
5. Move matter work onto accounts that are provisioned, governed and exportable, so the next log is one the organization can produce rather than explain.

Try Spellbook free to see what contract review looks like when the record stays in the matter file.

> About this content: This article is provided for informational purposes only. It does not constitute legal advice, and reading it does not create an attorney-client relationship. For legal advice on a specific situation, consult a licensed attorney in your jurisdiction. Spellbook is a software product; it supports legal professionals' work but does not replace their judgment. References to AI model capabilities and to pending rules reflect publicly available information as of the publication date and may change.