Is a human approval step sufficient?

Not necessarily. The reviewer needs useful information, enough time, suitable authority, and a control that affects the actual workflow.

Does intervention always mean stopping the whole system?

No. It may involve a temporary hold, a targeted correction, a changed instruction, or transferring a particular case to another process.

Does correcting an output teach the model permanently?

Not automatically. A correction may affect only the current case unless an explicit learning or update process uses it.

AI Glossary / Governance, ethics and safety / Intermediate

Intervenability

Definition

Intervenability is the practical ability to step into an AI-supported process and change its course when needed. In an operational setting, it involves usable ways for authorized people to pause, correct, redirect, or override relevant behavior in time to matter. The term also has narrower research meanings, including the effectiveness of interventions on a model's internal concepts.

What it means

Observing a system is different from being able to influence it. A dashboard may reveal a problem without giving anyone a workable way to stop the next action. Intervenability concerns that ability to act, including the surrounding workflow rather than only the model.

It requires more than a nominal button. Someone must recognize the situation, have the authority and information to respond, and reach a control that changes what the system actually does.

Different sources emphasize different kinds of intervention. This entry mainly addresses operational human intervention. It distinguishes that broad use from concept-based machine-learning research, where an intervention changes an intermediate representation and its effect on the prediction is studied.

How it works

Designers identify points where automated behavior can be interrupted or adjusted. They specify who may intervene, what actions are available, what state follows an intervention, and how normal operation can resume appropriately.

Useful controls may include holding a pending action, changing a case-specific instruction, transferring control, or sending the case to a review process. Their effectiveness depends on timing and scope. Stopping one component may leave already-issued actions running in another system.

Tests should establish that intervention changes the relevant behavior and that users understand its consequences. They should include delayed responses, unavailable operators, and unsuccessful control actions. A person listed as responsible for oversight needs a realistic opportunity to exercise that responsibility; an intervention mechanism that acts after an irreversible consequence cannot prevent that consequence.

Three practical examples

1. An AI-assisted publishing workflow holds a draft before release, allowing an authorized editor to correct or reject it before it reaches an external audience.

2. A scheduling system lets a supervisor temporarily adjust a specific allocation when local information makes the automated recommendation unsuitable, while showing when the exception expires.

3. A concept-based research model allows a user to correct an intermediate attribute and then observe how the final prediction changes, testing whether the correction meaningfully affects the output.

Think of it this way

Think of a train signal operator who can see an approaching conflict and change the relevant signal before the train reaches it. Observation, authority, communication, and timely action must work together.

The analogy is limited. AI interventions may concern a document or decision rather than a moving vehicle, and the appropriate response may be a correction instead of a stop. A visible control does not establish that every downstream action is reversible or that a human can always respond quickly enough.

Worked example

Consider a fictional document-distribution process. At 10:00:00 an AI system flags a batch as ready, but release is scheduled for 10:00:30. A reviewer notices a wrong recipient group at 10:00:08, decides to hold the batch at 10:00:14, and the hold becomes effective at 10:00:18.

The effective intervention arrives 12 seconds before release. The control has enough time in this particular trial, provided it truly prevents the distribution action.

In a second trial, the same sequence takes 35 seconds from readiness to effective hold. A 30-second release window is insufficient. Displaying a hold button has not solved the timing problem.

These are illustrative timings, not recommended operating limits. A real assessment must consider variable delays, missed alerts, operator workload, and whether all distribution channels respect the hold. It must also define how a held batch is safely reviewed and released later.

Illustrative case study

Fictional case: Linden Events uses AI to propose attendee allocations across sessions. Organizers can edit a proposal, but an integration automatically copies the original allocation into a separate notification queue before their edits take effect.

The team discovers that the apparent override does not cover the full workflow. It changes the process so that an authorized hold applies before queue creation and makes the status of pending notifications visible. It then tests edits, holds, and controlled resumption with realistic cases.

This invented example claims no measured improvement. It illustrates why intervenability must be checked across connected components. The model may accept a correction while the organization's surrounding automation continues acting on the superseded output.

Common misconceptions and mistakes

Intervenability is not identical to interpretability: understanding a result does not guarantee an effective control. It is also broader than merely having an emergency shutdown, although shutdown can be one relevant mechanism.

An intervention does not necessarily retrain the model or permanently change its behavior. A case-specific correction may affect only that case. Human intervention is not automatically better than automation either; users may make mistakes or lack relevant information. Its design and effectiveness require evaluation alongside the system's other safeguards.

Business use

Businesses should map where AI recommendations become consequential actions. At each relevant point, identify the available intervention, responsible role, time window, and affected downstream systems.

Test the process with the people expected to use it. Record failed or late interventions and examine whether workload, unclear status, or permissions contributed. For research systems that permit concept corrections, evaluate whether those corrections have the intended effect. Operational control and model-level intervention are related capabilities, but evidence for one does not automatically establish the other.

Check your understanding

A reviewer changes an AI recommendation after another system has already executed it. Does that demonstrate effective prevention through intervention?

Answer: No. The change came too late to prevent that action. The intervention point, timing, and downstream control need review; later correction may serve a different purpose.

Related terms

Controllability
Human-in-the-loop
Interpretability
Incident response
Human oversight
override
concept bottleneck model
workflow

Learn next

Next, trace one automated action from recommendation to execution. Mark the last point at which it can be held, identify who can act, and test what happens when intervention is delayed or fails.

Sources

Thomas Herrmann, Intervenability as a Design Requirement: https://arxiv.org/abs/2607.10322 Laguna and colleagues, Beyond Concept Bottleneck Models: https://arxiv.org/abs/2401.13544 The first author's intervention-interface passages and the second paper's abstract were read. Operational and model-concept senses are distinguished; examples are original. Exact correspondence with the supplied NIST glossary entry needs editorial scope review.

Previous: Interval tree Next: Interview survey