Your AI logs are now evidence

The Verge reported this weekend that prosecutors used the defendant’s ChatGPT logs as evidence in the Palisades fire criminal trial. Set aside the specific case for a moment. The procedural fact is the point. AI chat logs were treated, in open court, as discoverable, admissible evidence of intent and action. This is no longer hypothetical.

Three years ago the legal status of AI conversations was unsettled. As of this weekend, prosecutors and defense counsel have a working template for how to subpoena, authenticate, and present them. That template will not stay confined to criminal cases. Civil litigation, regulatory enforcement, employment disputes, and internal investigations all use the same evidentiary rules. AI conversation logs are a new category of business record, and the muscle to handle them is being built right now in courtrooms.

Three things worth being explicit about for any enterprise running AI in production.

• Conversation retention is now a governance decision, not an engineering default. Most enterprise AI deployments today log everything, indefinitely, because storage is cheap and “we might want to debug it.” That posture made sense when nobody was reading the logs but the on-call engineer. The same posture, in 2027, will be the discovery surface for any litigation that touches an AI-assisted process. Set retention deliberately. Document why. Make sure legal, security, and product are at the same table for that decision.
• Sensitive prompts and outputs are not the same as ordinary application data. A finance agent’s transcript can reveal pre-decision deliberations, candidate questions about employment terms, regulated-industry hypotheticals, customer health details, draft legal arguments. A reasonable lawyer will treat those transcripts as privileged where applicable and as discoverable where not. Your retention, redaction, and access-control posture should reflect the same distinctions, automatically, at write time.
• For SI and ISV partners on Microsoft Foundry, this is where the observability and audit story earns its keep. Customers in regulated industries are going to start asking, in procurement language, for documented retention policies, role-scoped access, automated redaction of named entities, and a clear chain of custody on prompt and response logs. Partners that ship a reference “agent records management” pattern (retention tiers, redaction templates, audit hand-offs, legal-hold workflows) will outrun partners still treating logs as engineering exhaust.

The deeper read is the one most uncomfortable for vendor decks. The “log everything” era of AI product development is going to look, in three years, the way “store every password in plaintext” looks today: technically simple, operationally cheap, legally indefensible. The teams that quietly do the records-management work now will be the ones who don’t have to do it under subpoena later. Worth one calendar invite this week.