Fatigue Discounting

Fatigue discounting is when a system treats later tasks as less important after doing work. It makes choices that favor easy, quick actions instead of harder ones over time.

Fatigue discounting refers to a progressive reduction in the effective value assigned to options as cognitive or resource fatigue accumulates. This phenomenon leads decision policies to favor lower-cost, short-horizon actions, altering expected utility estimates during extended task sequences.

A judge who has been hearing cases all day is more likely to grant a quick parole denial (the easy, low-effort default) in the afternoon than in the morning, even for cases with similar merits—simply because extended decision-making has drained their capacity to carefully weigh complex factors.

In a large language model evaluated over extended inference chains, the decision evaluator's coupling weights to the resource-monitoring module shift as token-budget utilization climbs. Empirically, selection probability for multi-step reasoning paths (high-cost action representations) drops relative to single-hop shortcuts (low-cost actions) after approximately 60% of context window saturation—exhibiting a time-dependent utility discount consistent with fatigue discounting. Introducing a dynamic reweighting layer that applies a fatigue-corrective coefficient inversely proportional to remaining capacity restores near-baseline salience_weighting_matrix distributions for high-effort options, partially recovering decision quality without full resource replenishment.

When the system gets tired it lowers the value of hard options. That causes it to pick simpler, faster actions instead.

Resource depletion reduces the decision utility via a weighted discount applied to option values, implemented at the evaluator-module interface; the evaluator's selection probability becomes asymmetric toward lower-cost actions. Structural constraints in the monitoring-to-evaluator pathway impose a time-dependent weighting that scales down high-effort option salience.

Give the system breaks so it regains energy and chooses better. Add reminders to pick long-term good options when tired.

Implement scheduled resource replenishment and dynamic reweighting of option utilities to counteract time-dependent discounting. Introduce metacognitive triggers that temporarily boost high-effort action salience during critical tasks.

Premature abandonment of complex tasks; Over-selection of low-value shortcuts; Failure to recover after rest

An adversarial actor can deliberately schedule complex or consequential decisions late in a long session, knowing that fatigue discounting will cause evaluators to favor low-effort, status-quo-preserving choices and reduce scrutiny of high-cost options. In negotiation or audit contexts, a counterparty may intentionally extend proceedings to exhaust the opposing party's decision quality, then introduce unfavorable terms when fatigue-induced discounting is at its peak. Automated systems can be probed by submitting critical queries after inducing computational or token-budget depletion, exploiting the asymmetric weighting shift to extract lower-effort, lower-quality outputs.

Implement mandatory scheduled replenishment breaks and enforce hard task-sequencing policies that place high-stakes decisions early in sessions, before significant resource depletion occurs. Introduce metacognitive monitoring checkpoints that flag when a decision evaluator's output distribution has shifted toward low-cost actions, triggering a forced reassessment or escalation to a rested reviewer. In automated systems, dynamic reweighting of option utilities using a fatigue-corrective coefficient—calibrated against baseline performance benchmarks—can partially counteract the time-dependent discount accumulation.