Are AI chatbot conversations discoverable in litigation?

Generally yes. Discovery reaches relevant electronically stored information in a party's possession, custody, or control. A prompt an employee typed into a company-provisioned assistant, and the output it returned, is ordinarily ESI like any other message — the fact that one participant was a model does not change its status.

Can we just set AI logs to auto-delete after 30 days?

A short retention period applied consistently in the ordinary course of business is defensible. It stops being defensible the moment litigation is reasonably anticipated, because the duty to preserve overrides the routine schedule. The failure mode is almost never the policy itself — it is that the auto-delete timer keeps running after a hold is issued because nobody wired the AI tool into the hold process.

Do we control logs held by a third-party AI vendor?

Often, yes, in the sense that matters. Control is generally assessed by your practical or contractual ability to obtain the data, not by where the servers sit. If your enterprise agreement gives you export or admin access to conversation history, courts will usually treat that data as within your control — and your inability to retrieve it in time is a self-inflicted problem.

How does this conflict with privacy deletion requests?

It conflicts directly, and the resolution is that a live preservation obligation is a recognised basis for declining or deferring deletion of the specific data under hold. What you cannot do is treat that exception as a blanket excuse: it covers the records within the hold's scope, for its duration, and you should document the reasoning at the time rather than reconstruct it later.

AI Chatbot Logs, Litigation Holds, and E-Discovery: 2026 Guide

Why AI Logs Are Worse Than Email

Email discovery is a solved problem operationally: one system, one archive, one hold tool, custodians everyone can enumerate. AI usage is the opposite. It is spread across a sanctioned enterprise assistant, an IDE plugin, a support-desk bot, three agents wired into internal tooling, and whatever individual employees signed up for with a work address.

The result is that most organisations cannot answer the first question a court or opposing counsel asks: where does AI conversation data live, who can retrieve it, and how long does it survive by default? Not knowing is itself the finding.

What Counts as the Record

The prompt

What the employee asked, verbatim. In an employment or trade-secret dispute this is frequently the single most revealing artefact in the case, because people type things into an assistant they would never put in an email.

The output

What the model returned — including the version the employee rejected before regenerating. Where a business decision was made on model output, the output is the contemporaneous basis for the decision.

Retrieval and grounding traces

In a RAG system, which internal documents were pulled into context. This shows what information was actually in front of the decision-maker, which is often more probative than the answer itself.

Agent action logs

For agents that take actions — sending messages, updating records, executing transactions — the action log is the audit trail. If an agent did something contested, this is the only record of why.

Configuration and version history

System prompts, guardrail settings, model version, and temperature at the time. Reproducing behaviour later is impossible without them, and 'we can't reproduce it' is not a defence that lands well.

Feedback and override signals

Thumbs-down ratings, human overrides, and escalation flags are the closest thing to a contemporaneous admission that the system was getting something wrong — and they are routinely retained longer than anyone realises.

The Two Ways Teams Get This Wrong

Both failure modes are common, and they pull in opposite directions:

Delete-everything. A short vendor-default retention window feels privacy-forward and cheap. It works until a dispute arises and the logs that would have exonerated you evaporated on schedule — or worse, kept evaporating after the duty to preserve attached, which converts a records-management decision into a sanctions question.
Keep-everything. Retaining every prompt indefinitely because storage is cheap creates a permanent, searchable corpus of unguarded employee speculation, half-formed legal theories, and confidential material pasted in for summarisation. Every future matter gets to search it.
The hold that misses the AI tools. The most frequent real-world failure: legal issues a hold covering email, chat, and file shares, and nobody adds the AI assistant because it is not on the systems inventory. The timer keeps running.
Shadow AI with no admin access. Employees using personal accounts for work create records you may have a duty to preserve and no technical ability to reach.

Where Preservation and Privacy Deletion Collide

A consumer submits a deletion request under a state privacy law. Their conversation with your support bot is also within the scope of an active litigation hold. Privacy regimes generally recognise legal-obligation and legal-claim exceptions to deletion, so the hold usually wins for that specific data — but only if you can articulate the scope at the time and confine the exception to it. The organisations that get in trouble are the ones that either honour the deletion and destroy evidence, or invoke "legal hold" as a standing reason to ignore deletion requests wholesale. Both are avoidable with a written mapping between your hold scope and your deletion workflow.

A Retention Program That Survives Contact

The goal is a defensible, consistently applied schedule with a hold override that actually reaches every AI system — not the longest or shortest possible retention.

Inventory every AI system that stores prompts or outputs, including plugins and agents— Essential

For each, record the default retention window, admin export path, and who owns the account— Essential

Add AI systems to the litigation-hold checklist as named custodial sources— Essential

Verify you can technically suspend auto-deletion per system — test it before you need it— Essential

Set retention by use case, not one global timer: support transcripts and agent action logs are not the same record— Policy

Map the privacy-deletion workflow against active hold scopes so both are honoured— Privacy

Capture model version, system prompt, and configuration alongside the transcript— Evidence

Close off shadow AI with sanctioned tooling — you cannot preserve what you cannot reach— Governance

Document the retention rationale in writing, before any dispute exists— Defensibility

Your public AI disclosures are part of the same record

What your site tells users about AI chat, data retention, and processing is the first thing pulled when a dispute starts — and it is frequently out of date. RatedWithAI scans your public properties for those gaps. Start with a free scan.

Scan Your Site for Free →

Frequently Asked Questions

When does the duty to preserve AI logs actually attach?

When litigation is reasonably anticipated — not when a complaint is filed. A demand letter, a credible internal complaint, or a regulator's inquiry can start the clock. Because AI retention windows are often short, the gap between anticipation and issuing the hold matters far more here than it does for email.

Is an AI-generated summary privileged if legal asked for it?

Privilege turns on the same analysis as any other document: was it a confidential communication for the purpose of legal advice. Routing a request through a model does not create privilege where none existed, and doing it in a shared workspace can undermine confidentiality. Treat AI-assisted legal work with the same access controls as any privileged draft.

What if the vendor says they don't retain our prompts?

Get it in writing at the contract level, with the specific retention period and any exception for abuse monitoring or trust-and-safety review, and confirm whether a zero-retention setting is actually enabled on your tenant. Marketing pages and contractual commitments frequently differ, and the exception carve-outs are where the surviving copies live.

Do we have to preserve logs from an employee's personal ChatGPT account?

Possibly, if it was used for company business — the analysis parallels personal devices and personal email, where control is assessed practically. The realistic answer is that enforcement is hard and prevention is easy: sanctioned tooling with admin access is worth more than a policy telling people not to.

Can we sample rather than preserve every conversation?

Not once a hold attaches to that category of records. Sampling is a reasonable approach to routine retention design before any dispute, but preservation obligations run to relevant records, not to a representative subset of them.

What leaves the building when staff paste confidential material into an assistant.

CCPA Data Retention and Minimization for AI Training 2026

The privacy-side constraint that pulls against long retention windows.

AI Meeting Notetakers and Recording Consent 2026

Another system quietly generating records nobody classified.

AI Employee Monitoring Disclosure Laws 2026

What you must tell employees before you log their activity.