Preservation Is the Elephant in the Room for AI Evidence
When an AI system becomes relevant to litigation, the conversation usually moves quickly toward reliability, accuracy, authentication, or expert testimony. Those questions matter, but they assume something more basic has already gone right: that the evidence needed to understand what the system did still exists.
That assumption may not hold.
The Fragmentation of AI Records
AI-mediated activity can leave behind a fragmented record. A final report might survive while the source material, retrieval history, prompts, tool calls, intermediate outputs, system version, or execution logs do not. Some of those records may sit with the organization using the system. Others may be held by an AI vendor, cloud provider, or third-party application. Some may never have been retained at all.
That changes the nature of the evidentiary problem.
Consider an AI system used to review a large document population and identify material that appears relevant to an investigation. The organization may preserve the final report and the documents the system selected. Months later, a dispute develops over whether important records were omitted.
At that point, the final report tells only part of the story. To understand the selection process, investigators may need to know what documents were available to the system, what instructions it received, what filters or retrieval methods were used, what version of the system performed the task, and whether intermediate results were preserved.
If those records are gone, a later expert cannot recreate them simply by examining the final report.
Authentication vs. Preservation
This is what makes preservation different from authentication. Authentication can help establish that a particular file or output is what someone claims it is. Preservation determines whether enough of the underlying record survives to evaluate what actually happened.
The distinction becomes more important as AI systems move from producing isolated text to performing sequences of work. Agentic systems may retrieve information, call tools, make intermediate decisions, update files, communicate with other systems, and take actions before a human ever sees the result. The evidentiary object is no longer just an output. It may be an entire execution path.
A missing portion of that path does not necessarily make the event impossible to investigate. Other evidence may still establish who authorized the system, what task was assigned, what output appeared, or who acted on it. But those facts do not automatically establish what the system actually did during the execution that produced the result.
Upstream Considerations for Organizations
That is why preservation needs to be considered upstream.
Organizations adopting AI systems should understand what records those systems create, where those records are stored, how long they remain available, whether they can be exported, and who controls access to them. They should also understand what disappears when a session ends, a model changes, a vendor updates its service, or a retention period expires.
Those questions are not merely technical housekeeping. They determine what lawyers, investigators, experts, and courts may eventually be able to establish.
The most difficult AI evidence problem may therefore arise long before anyone argues about admissibility.
By the time litigation begins, the question may no longer be, “What evidence should we preserve?”
It may be, “What evidence was ever preserved at all?”