The inference from absence problem: accountability when AI agents act on what is not there

AI accountability frameworks are built around positive evidence. A decision was made; it was based on these inputs; it produced this output. The audit trail traces what was there — what the agent observed, weighed, and acted on. This structure fits the most visible class of agent decisions well. It does not fit a class of decisions that is equally common and often more consequential: decisions made because something expected did not arrive.

AI agents regularly infer from silence. A care agent expects a vital-sign report every two minutes and receives none for twenty; it infers deterioration or sensor failure and escalates. A hardware agent expects attestation heartbeats from a device fleet and notices a machine has gone quiet; it infers potential compromise and quarantines the device. A security agent expects a scheduled cryptographic key rotation to complete by midnight; it receives no confirmation and triggers an alert. In each case, the agent's action is reasonable. In each case, the chain of reasoning that led to it is not what the accountability record shows.

The record shows an escalation, a quarantine, an alert. It does not show "no signal arrived when one was expected." The absence that drove the decision is invisible in the output — not because anyone chose to hide it, but because standard audit architectures are built to record what was present, not what was absent.

Why absence-inferences matter for accountability

An absence-inference contains three components that are all accountability-relevant and all at risk of being lost. First, the expectation: some prior configuration or learned behavior established that a signal should arrive. The nature of that expectation — who set it, when, under what conditions — determines whether the inference was warranted. Second, the window: absence is evaluated over time. The same missing signal means different things after five minutes, five hours, or five days. The window used to declare an expected signal absent is a parameter of the decision, and it belongs in the record. Third, the alternative hypotheses: absence is ambiguous. No vital-sign report could mean the sensor failed, the network dropped, the patient was moved, or the patient deteriorated. The agent's inference favored one hypothesis over others, and the basis for that weighting should be auditable.

None of these three components naturally surfaces in a standard action log. The log records the output: escalation triggered at 14:23. The expectation, the window, and the alternative-hypothesis weighting are pre-decision context that most logging infrastructure does not capture. A reviewer examining the audit trail after an adverse event sees what the agent did. They cannot see why the absence led to the specific inference it did, whether the inference was reasonable given the configured expectation, or whether the window was appropriate for the clinical or security context.

At the care crossing

Physical-world care is saturated with absence-inferences. A care agent monitoring an elderly person in an assisted living setting is continuously evaluating expected patterns against observed ones: expected sleep duration, expected movement, expected medication acknowledgement, expected response to a check-in prompt. When the expected pattern is not observed, the inference varies by context — routine variation, early deterioration, acute event, device failure. The agent must distinguish between these options using whatever signals it has, and it will often get the weighting wrong in ways that depend entirely on the expectation that was configured.

The accountability problem is sharp when the inference was wrong. A care agent that escalated based on an absence that turned out to be a routine sensor dropout triggered unnecessary intervention. The audit review begins: why was the escalation triggered? The log shows the time of escalation and the rule that fired. It does not show what absence triggered the rule, what the expected signal was, how long it had been absent, or whether any positive signals suggesting the person was fine were weighed against the absence. The absent record of the absent signal is where the accountability falls.

At the hardware crossing

Hardware agents managing device fleets depend on continuous attestation and heartbeat signals to maintain visibility into device integrity. When a device stops attesting, the inference is security-relevant: is this a hardware failure, a network interruption, a legitimate maintenance window, or a compromise indicator? The correct inference depends on context — device type, prior behavior, deployment environment — and the consequences of getting it wrong differ across error modes. A false-positive quarantine removes a device from service unnecessarily. A false-negative delays detection of a genuine compromise.

When a quarantine decision is later reviewed, the audit record should show: what attestation was expected, over what interval it was absent, what contextual signals were available, and what alternative hypotheses were evaluated. Most hardware agent deployments do not produce records at this level of detail. The quarantine event is logged; the absence reasoning that preceded it is not. Post-incident review of hardware agent decisions involving absence-inferences is consequently shallow — reviewers can confirm that the policy triggered, but not whether the policy was appropriate to the specific absence pattern observed.

At the post-quantum security crossing

Cryptographic operations have strong expectations about timing. Key rotation schedules, certificate renewal windows, challenge-response latencies — these are all temporal contracts whose violation is itself a security signal. An AI agent managing post-quantum key infrastructure watches for these violations and acts on them. A rotation that did not complete on schedule, a renewal confirmation that never arrived, a challenge that received no reply: each of these is an absence-inference, and each is an input to a security decision that must be accountable to auditors, regulators, and incident responders.

The inference from absence in post-quantum contexts carries additional weight because the transition to quantum-resistant algorithms introduces new expected signals — attestations in new formats, renewals using new algorithms — whose absence patterns are not yet well-understood. An agent that triggers a security response because a post-quantum attestation confirmation did not arrive on the expected schedule may be correctly identifying a migration failure, or may be responding to a timing difference between old and new algorithm formats. The accountability record for that decision must include the expectation that was violated and the basis for treating its violation as a security event.

What absence-aware accountability requires

Three changes to standard audit architecture would make absence-inferences visible and reviewable. First, expectation logging: the configuration that defines what signal an agent expects, over what interval, should be logged as a first-class audit object. It is the precondition for evaluating every absence-inference the agent makes. Second, absence events: when a configured expectation is violated and an inference is drawn, the absence event itself should be logged explicitly — "expected signal X not received in window W; inference: Y" — before the action that follows. This makes the reasoning reconstructable. Third, hypothesis logging: where an absence could support multiple inferences, the agent's basis for choosing among them should be recorded. This does not require logging every alternative possibility; it requires that the weighting used to select the inference is not lost between evaluation and action.

The inference from absence problem is not exotic. It arises in every domain where agents monitor for expected conditions — which is most of the domains where AI agents are deployed for high-stakes work. The accountability gap it creates is structural and silent: the records that exist are accurate, but they are systematically incomplete in a way that makes the most consequential parts of the reasoning unreviewable. Closing that gap requires treating the absent signal as a first-class input, not as the background against which the real decision happens.