Direct answer
Approval fatigue occurs when many repetitive or low-value decisions consume human attention. Automation bias occurs when a person gives undue weight to an AI recommendation. Rubber-stamping occurs when a recommendation is accepted almost automatically under the appearance of review. Detection requires considering workload, review time, runs of identical decisions, the rate at which AI errors are caught, reasoned amendments and refusals, later corrections and outcome quality together.
In plain language
If your phone gives an unimportant warning every minute, you will eventually press 'OK' without reading. You may then miss the warning that matters. Human approval filled with unnecessary repetition stops being a real safety gate in the same way.
Why this matters
A tired person, or one who trusts AI too readily, may approve a wrong recommendation. If the organisation treats a high approval rate as evidence of compliance, it may fail to notice that the control is not working.
Do not confuse
- A high approval rate is not proof of automation bias by itself; the AI recommendations may genuinely be correct.
- A fast decision is not proof of shallow review by itself; simple cases may take little time.
- The number of overrides is not a quality measure; reasoned and correct intervention matters.
- Approval fatigue concerns workload; automation bias concerns excessive weight given to an AI recommendation.
- Rubber-stamping means that a human name is present but no material review took place.
What should you do?
- Reduce repetitive, low-risk approvals and reserve human attention for material and high-impact decisions.
- Apply capacity limits and rotation based on workload, shift length, decision complexity and interruptions.
- Present the AI recommendation, evidence, uncertainty and counter-view separately and in a balanced way.
- For a sample of cases, ask the person to make an initial independent assessment before seeing the AI recommendation.
- Measure review time, error detection, amendments, refusals, later corrections and outcome quality together.
- Examine unusual patterns of agreement, speed and error across human groups and decision types.
- When quality falls, pause the flow, correct the root cause and refresh reviewer training with evidence.
How do you audit it?
- Do the volume and complexity of decisions exceed the person's genuine review capacity?
- Do approvals show unusual runs with the same duration and outcome?
- Can reviewers detect planted or known AI errors?
- Are amendments and refusals reasoned rather than random?
- Are the AI recommendation, evidence and alternatives presented neutrally?
- Has the high approval rate been compared with outcome quality and later corrections?
- When fatigue or bias signals appeared, were workload and interface design actually changed?
Limit
No single threshold for speed, approval or override suits every task. Behavioural signals do not prove a person's intention on their own; they must be interpreted with outcome, risk, task difficulty and observed error data.
Remember in one sentence
Control is not the human saying 'yes'; it is the human being able to say the right 'no' when necessary.
Sources for this record
- S26NIST SP 800-53 Rev. 5, *Security and Privacy Controls*Standard
- S38NIST AI 600-1, *Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile*Voluntary standards-oriented institutional guidance
- S44NIST, *AI Risk Management Framework Playbook* and AI RMF CoreVoluntary implementation guidance
- S45European Union, Regulation (EU) 2024/1689 — *Artificial Intelligence Act*Legislation
- S50NIST AI 100-1, *Artificial Intelligence Risk Management Framework (AI RMF 1.0)*Voluntary standards-oriented institutional framework