A person and an AI system can produce excellent joint results while the person's unaided capability quietly declines. The dashboard sees the result. The failure arrives through what the dashboard stopped measuring.

Three quantities need separate names. Assisted performance is the quality of the joint output while the aid works. Unaided competence is what the person can still generate, judge or control without that aid. Residual responsibility is what the person or organisation must still detect, decide or repair when the aid is wrong. Automation becomes dangerous when the first rises, the second falls and the third remains fixed.

Human-factors researchers have described versions of this pattern for decades. Lisanne Bainbridge's analysis of industrial automation noted the irony of removing routine control while leaving the operator responsible for abnormal conditions.1 In a laboratory navigation task, Endsley and Kiris linked higher automation to lower situation awareness and slower manual response after failure.2 Endsley and Kaber later found that automating implementation improved normal performance in a dynamic control simulation, yet made recovery worse when automation failed.3 A widely used automation model separates information acquisition, analysis, decision and action, helping designers see where human work remains after another stage is delegated.4 Decision-aid experiments have also found omission and commission errors when people defer to an automated recommendation.5 These are task-specific studies, not a law that all automation causes decline. Their shared warning is narrower: the role left to the human may require skills that the normal workflow no longer rehearses.

The distinction also prevents an overreaction. Cognitive offloading is not synonymous with cognitive damage. In four studies, Sparrow, Liu and Wegner found that expecting future access to information reduced recall of the information itself while improving recall of where to retrieve it.6 External reminders can improve prospective-memory performance.8 Yet Internet search can inflate estimates of one's own explanatory knowledge,7 and reminder use can become excessive when confidence or effort costs are misjudged.9 A checklist, map or search engine may still create a better joint cognitive system than unaided memory ever could.

The design question is therefore not, “Should people remember everything?” It is, “Which internal capability must remain because the external path may be wrong, unavailable or unauthorised when it matters?” Cognitive integrity is not self-sufficiency. It is the preservation of an independent route to warranted disagreement and safe action.

flowchart LR
  A["The number that looked right"] --> B["When success conceals erosion"]
  B --> C["What work keeps a capability alive"]
  C --> D["The retained competence kernel"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#087A98,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

A skill is not maintained by contact with its finished answer. It is maintained by performing the transformations that the future failure will demand.

Consider learning a route. Watching a navigation trace teaches landmarks, but choosing turns requires a different operation. Reading a proof can create familiarity, but generating the next step requires retrieval and constraint search. Reviewing an AI diagnosis can teach vocabulary, but isolating the cause of a novel failure requires intervention and counterfactual comparison. A workflow may expose a person to more information while removing the cognitive act that turns information into executable capability.

Four kinds of work matter repeatedly across domains:

Retrieval research gives one clear example of why passive exposure can mislead. Karpicke and Roediger found that repeated testing after successful recall improved delayed retention, whereas repeated study did not; learners' predictions did not track this difference.10 The result concerns vocabulary learning, not complex professional judgement. Its design lesson transfers more cautiously: practice should sample the operation we expect to remain available, rather than only the material surrounding it.

A small model of decay and rehearsal

Organisations often ask for a single practice interval: run a drill every quarter, recertify annually, or rotate staff monthly. No universal interval follows from the literature. Skills differ in complexity, initial mastery, interference, frequency and consequence. Still, an explicit toy model is useful because it exposes the assumptions hidden inside a policy.

• C is task-specific unaided competence from 0 to 1.
• Cmin is a residual floor, not zero knowledge.
• Δ is days since operation-matched practice.
• h is a user-supplied half-life assumption.
• q is drill quality from generation, variation, feedback and recovery.
• b is the maximum rehearsal boost.

Suppose competence is 0.90, the residual floor is 0.25 and the assumed half-life is 90 days. After 60 days without unaided practice, the model gives about 0.66. A drill with quality 0.8 and boost 0.35 raises it to about 0.75. Those numbers are not empirical predictions. They show that a policy depends on a decay assumption, a target floor and the quality of rehearsal. Remove the half-life assumption and the equation cannot set a cadence. Replace a recovery drill with explanation-only review and its quality term should fall.

Recent generative-AI studies make the same distinction more urgent, while remaining context-bound. In a field experiment involving nearly one thousand pupils at one Turkish high school, unrestricted GPT assistance improved practice performance but was followed by worse unaided examination performance than the control condition. A tutor configuration that provided safeguards and hints largely mitigated the loss.12 A survey of 319 knowledge workers found that greater confidence in generative AI was associated with less self-reported critical-thinking effort, while work shifted towards verification, integration and stewardship.13 The first result is short-term and maths-specific. The second is correlational and self-reported. Together they justify experiments on workflow design, not a declaration that AI inevitably makes people less intelligent.

The answer is not to preserve the entire pre-automation job. Preserve the smallest set of human operations that can still interrupt a bad trajectory before its consequences become irreversible.

Call this set the retained competence kernel. It has four functions: frame, orient, challenge and recover. The names describe what the person must be able to do, not what the interface displays.

Each capability also needs a required depth. A useful five-level scale runs from awareness to degraded-mode control:

The minimum depth is the highest operation needed before reliable independent help can arrive. This makes recovery time a first-class design variable. An analyst with a four-hour escalation window may need to audit and frame, but not rebuild an entire valuation model. A control-room operator facing a ninety-second instability may need rehearsed degraded control. A consumer using a recommendation system for a reversible, low-cost choice may need only provenance and an easy way to undo.

A configured system, not “the AI”

Return to Anika's credit memorandum. The relevant system is not a language model in isolation. It includes identity, authorised data, temporal cutoff, document extraction, calculations, retrieval, model proposals, policy rules, human authority, typed actions and evidence receipts. Each component can fail differently. Calling all of this “AI” obscures what the human should retain.

Suppose the system computes ratios deterministically, uses a model to draft explanations, and requires human approval before the memorandum is released. The retained kernel might be:

• frame the decision and identify non-negotiable policy limits;
• estimate two key ratios from independently selected figures;
• challenge temporal alignment, source sufficiency and material exceptions;
• stop release, reconstruct a minimal memo and preserve evidence if the system is unavailable.

Anika need not remember every policy paragraph or reproduce the draft's prose. Those can remain external, provided provenance and entitlement are reliable. She does need enough semantic and numerical structure to know which sources must agree, which approximation should be in range and what safe fallback is possible.

Choosing depth from the failure surface

Four variables do most of the work. Consequence asks what happens if the output is wrong. Opacity asks whether a plausible failure is visible from the interface. Urgency asks how quickly harm becomes hard to reverse. Backup independence asks whether another qualified path can arrive in time and whether it shares the same sources, models or assumptions.

Verification must be differently caused

A second answer is not independent merely because it appears in a second window. Two models trained on overlapping data, prompted with the same retrieved context and asked to use the same calculation may reproduce one another's blind spot. Independence is about causal path, not visual separation. Useful alternatives include a deterministic invariant, a source chosen before seeing the proposal, a different measurement modality, a human-generated estimate or a reviewer with separate context and authority.

stateDiagram-v2
  [*] --> Claim
  state "The twin desks" as Claim
  state "Generate" as Test
  state "Predict" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

Good delegation removes unnecessary production work while preserving the smallest independent capability that the person may still need under failure.

A retained-competence policy should begin with the configured task, not with a nostalgic list of everything people once did manually. Map the normal assisted path. Mark the faults that can remain plausible at the interface. Identify who still has authority when one occurs, how long they have, and which genuinely independent backup can arrive. Only then choose what must be practised.

This sequence produces a practical rule: retain the deepest operation required before a reliable independent path reaches the problem. A reviewer who can safely pause a reversible decision may need recognition and hand-off. An operator facing an irreversible effect in seconds may need diagnosis and degraded-mode recovery. A person whose approval has no causal role should not be used as decorative oversight. Either give that person the information, time and practice needed to intervene, or move authority elsewhere.

Put rehearsal inside the operating system

Annual training is often too distant from the live workflow. A better design distributes small interventions across ordinary work and reserves full simulations for the rare operations that normal work cannot exercise.

• Predict before reveal. On a small risk-weighted sample, ask the human to record a range, rank hypotheses or name a likely exception before the proposal appears.
• Challenge by relation. Check a source cutoff, invariant, causal dependency or entitlement rule. Do not ask only whether the prose sounds convincing.
• Withhold selectively. Occasionally remove one convenience while preserving safety, so the person generates the missing step rather than merely recognises it.
• Inject blind faults. Seed plausible, novel failures in an operational twin and score detection, diagnosis, containment and evidence preservation.
• Rehearse recovery. Practise a minimal safe service, rollback, hand-off or stop condition under the real recovery window.

These interventions should be sparse. Requiring a prediction for every low-risk task will teach people to click through the ritual. Sampling should rise with consequence, opacity, change rate and elapsed time since successful unaided performance. It should fall when a deterministic control or independent specialist removes the person's residual responsibility.

Measure the capability, not attendance

Completion records show exposure to training. They do not show whether the needed operation remains available. A retention dashboard should use direct, role-specific measures:

• Blind-fault detection rate: the proportion of novel injected faults found before release.
• Diagnostic precision: whether the person identifies the causal region rather than naming a broad symptom.
• Time to safe state: elapsed time until effects are contained, reversed or validly handed off.
• Independent-path rate: the share of checks that differ from the proposal in source, representation, mechanism or authority.
• Practice coverage: whether frame, orient, challenge and recover have each been exercised within their required interval.

Do not combine these into a decorative “human readiness score” unless the decision rule is explicit. A high average can hide the one missing recovery operation that matters. Keep the measures typed, show missingness, and record the configured system version used in each exercise.

Use the retained-competence policy lab

The lab below turns the framework into a traceable decision aid. It estimates exposure from consequence, opacity, urgency, residual responsibility and weakness of backup. It then applies the paper's synthetic decay model to a measured unaided baseline. The output recommends a retention level, a practice interval and missing operations.

function assessRetention(input) {
  const exposure = 100 * (
    0.28 * scale(input.consequence) +
    0.20 * scale(input.opacity) +
    0.22 * scale(input.urgency) +
    0.15 * scale(input.responsibility) +
    0.15 * scale(input.backupWeakness)
  );

  const current = 0.20 +
    (input.baseline - 0.20) *
    Math.pow(2, -input.daysSincePractice / input.halfLifeDays);

  const depth = exposure >= 80 ? 4 :
                exposure >= 60 ? 3 :
                exposure >= 40 ? 2 :
                exposure >= 20 ? 1 : 0;

  return { exposure, current, depth };
}

Govern competence as a changing system property

Retention needs change whenever the system changes. A new data source may make a failure less visible. A deterministic guard may make recovery easier. A model update may alter error shape without changing benchmark accuracy. A reorganisation may remove the independent expert who previously supplied backup. Each material change should trigger a new boundary and recovery analysis.

The same rule applies to individuals. A failed drill is not proof of negligence or a stable personal deficit. It is evidence that the present configuration cannot rely on that capability. Respond first by changing the system: increase practice, narrow authority, improve evidence, lengthen the recovery window or add an independent control. Accountability is only credible when the organisation funds the competence, information and time that accountability requires.