The AI Evidence Landscape

Traditional digital forensics collects file system artifacts, network logs, memory dumps, and database records. These evidence types are deterministic — the same input to the same system produces the same output, and the forensic record proves what happened. AI systems introduce probabilistic evidence. The same input to the same model may produce different outputs on different runs. The context window is ephemeral — it exists only during the session and is not preserved by default. The model's reasoning process is opaque — you can observe the output but not the internal decision path. AI forensics must account for evidence that is non-deterministic, transient, and partially observable.

1. Ephemeral Evidence Context window contents, in-flight model state, and session-specific configurations exist only during the active session. If they are not captured contemporaneously — at the time they exist — they are gone permanently. Forensic readiness for AI systems means logging these ephemeral artifacts as a routine operational practice, not as an incident-triggered activity.
2. Probabilistic Evidence A model that produced a harmful output during the incident may not produce the same output when the investigation team replays the same input. This does not mean the incident did not happen. It means the evidence of the model's behavior must be captured from the original session logs, not from reproduction attempts. Original logs are evidence. Reproduction is demonstration.
3. Distributed Evidence In multi-agent systems, the evidence for a single incident may span multiple agents, multiple services, and multiple log sources. The injection entered through Agent A, the exploitation occurred through Agent B's tool access, and the exfiltration happened through Agent C's output channel. Reconstructing this chain requires log correlation across agent boundaries.