Does the person overseeing this actually have what they need to judge it?
Enhanced: 22 controls
Who it applies to
Enter for every AI system whose letter is E or A (this area has no Baseline controls). Go as deep as your system letter, and implement in your mode.
Why it exists
Give the overseeing human the information and the time to make a real judgement — because the most common way human oversight fails is not absence but futility: a person asked to approve, at speed, decisions presented with no basis on which to disagree. When the interface shows only the AI’s recommendation and an approve button, people approve; they trust the confident machine, especially under load. Design against that. An oversight interface that does not equip the human to say no is not oversight; it is a signature machine.
What to actually do
Show the human what they need to disagree, not only what to approve. Present the factors behind the AI’s output, its confidence and its known weaknesses for this kind of case, and — where it matters — what the alternatives were. The test is whether a competent person could, from what the interface shows, form an independent view and override the AI. If all they can do is agree, redesign it.
Design against the pull to defer to the machine. People over-trust confident automation — the tendency named in the Competence dimension — and a fast, fluent interface amplifies it. Build in friction where the stakes justify it: require the reviewer to record a reason, surface disconfirming information, slow the moment of decision for the highest-stakes actions. The friction is the point.
Controls
Enhanced2 controls
Control 194
Present oversight interfaces that equip the human to override, not only to approve.
EnhancedPreventiveall rolesper system
What it is
Basis, confidence, limitations, and alternatives — enough to form an independent view.
How to implement it · Minimum rules
Basis: the main factors and inputs behind the output, not only a score or label
Confidence for this case, clearly flagged when low
Known weaknesses for this kind of case, e.g. thin data or unusual profiles
Alternatives: next-best options or the outcome without the AI, where relevant
Override as easy as approval, with a field for the reason
How to start
Sit with a reviewer for an hour of real cases and note whether the screen would let them disagree; specify what is missing.
Evidence
The interface design showing basis, confidence, limitations and alternatives, and the override rate.
Mode: light / standard / full
light
With a provider's tool, turn on its reasons or sources display and give reviewers a one-page list of its known weaknesses.
standard
The system owner specifies the fields and they are built into the review screen.
full
Designed with reviewers, tested on whether they catch errors, and monitored through override rates.
Control 195
Build friction against automatic deference for high-stakes decisions.
EnhancedPreventiveall rolesper system
What it is
Recorded reasons, surfaced counter-evidence, or a deliberate pause where warranted.
How to implement it · Minimum rules
Scope: define which decisions are high-stakes, e.g. refusals, irreversible actions, decisions about people
Recorded reason required to approve as well as to override
Counter-evidence shown before approval: the factors pointing the other way
Deliberate pause: approval unavailable until the case details have been opened
No bulk approval of high-stakes cases
How to start
Identify the highest-stakes decisions of this system and require a written reason before any of them can be approved.
Evidence
The friction measures configured and a sample of the reasons recorded.
Mode: light / standard / full
light
A mandatory reason column in the decision record, checked by the owner-manager.
standard
Reason fields and blocked bulk approval configured in the workflow by the system owner.
full
Friction built into the interface, with reason quality and approval times reviewed by the oversight lead.
Evidence
The oversight interface designs; evidence they present basis and alternatives; override rates and the friction measures in place.