When should I override AI?

The question is usually asked backwards. "When should I override AI?" assumes the decision gets made in the moment, by whoever happens to be looking at the output, using judgement they will have to justify afterwards. That is the worst possible time and the worst possible basis. The override rule should be written before the system is deployed, not improvised while something is going wrong.

Aviation and medicine worked this out decades ago and wrote it into procedure. Most organisations deploying AI in 2026 have not, which is why their staff are making these calls alone, under time pressure, with no institutional cover.

Six conditions that should trigger an override

These are drawn from the human-factors literature and from what the AI evidence supports. They are conditions, not instincts, which is the point.

1 · You hold context the system does not. The strongest and most common ground. History, politics, a prior attempt, something a client said last week. The model is not wrong about the facts it has; it is answering a different question from the one that actually applies.

2 · The output conflicts with something you can independently verify. A primary document, a person who knows, the original study. Not another model: the South African High Court judgment in Mavundla records a judge testing a fabricated citation in ChatGPT, which confirmed the non-existent case was real.

3 · The task is at the edge of the domain rather than its centre. The competence boundary is jagged, and consultants working just outside it in the BCG experiment were 19 percentage points less likely to reach a correct answer than consultants using no AI at all.

4 · The output is unusually confident on an unusually hard question. Not proof of error, but the moment to slow down, because confidence is a property of the writing rather than the knowledge.

5 · The consequence is irreversible. Reversible decisions can tolerate a wrong answer. Irreversible ones cannot, and the override threshold should scale with that rather than with how confident anyone feels.

6 · You cannot explain why the answer is right. If you could not reconstruct the reasoning for someone who challenged it, you are not in a position to accept it, whatever the output says.

Three conditions where you should defer, and this is the harder half

A page that only lists reasons to override is not a rule, it is encouragement. The evidence supports deferring in specific circumstances, and saying so is what makes the rest credible.

Where the system demonstrably outperforms you on this task class, and you have measured it. The Vaccaro meta-analysis found that where the AI alone outperformed the human alone, combining them dragged results down towards the human's level. If you have evidence the system is better here, overriding on instinct is likely to make the outcome worse.

Where your objection is aesthetic rather than substantive. Disliking the phrasing, the framing or the approach is not the same as identifying an error, and it is the most common reason people reject correct output.

Where you are overriding because you produced a different answer first. Anchoring runs both ways. Having a prior position is essential for noticing divergence, and it is also a reason to be suspicious of your own resistance.

The precondition nobody checks

All of the above presumes something that is frequently untrue: that the person can detect the error at all. If nobody in the chain could have produced the work themselves, the override right is theoretical. That is the capability test, and it governs everything else. See who owns verification.

There is also an organisational precondition. If overriding costs the individual something, a missed target, a conversation with a manager, being the only one who did, the right exists on paper and not in practice. Article 14 of the EU AI Act, in force since 2 August 2026, requires that overseers of high-risk systems can decide not to use them or disregard their output. That is a capacity, not a permission, and the distinction is where most arrangements fail.

Where this is uncertain

No study has tested whether organisations with written override rules outperform those without. The six conditions are drawn from human-factors research and from the AI evidence cited here, but the specific list is a reasoned construction rather than a validated instrument. Treat it as a decision-forcing device.

There is also a genuine tension the page cannot resolve. Overriding a system that is better than you costs accuracy; not being able to override costs accountability and the capacity to catch the cases the system gets badly wrong. Both are real, and the balance depends on consequence and reversibility rather than on principle.

The measurement that tells you the truth

Count the overrides. If nobody has disregarded the system this quarter, that is not evidence of a good system. It is evidence of an untested right, and possibly of a workload that makes real scrutiny impossible or a chain in which nobody is competent to object. It is the cheapest diagnostic available and almost nobody runs it.

Related SuperSkills research

On the framework that holds this, the Delegation Boundary Map. On the legal duty, meaningful human oversight. On why review at the end is the weakest position, human in the loop is not a safeguard. On detecting error, how do I know when AI is wrong and automation bias.

Key sources

• Vaccaro, M., Almaatouq, A. and Malone, T. (2024). When combinations of humans and AI are useful. Nature Human Behaviour, 8.
• Dell'Acqua, F. et al. (2023). Navigating the Jagged Technological Frontier.
• Parasuraman, R. and Manzey, D. H. (2010). Complacency and Bias in Human Use of Automation. Human Factors, 52(3).
• Dietvorst, B. J. et al. (2015). Algorithm aversion. Journal of Experimental Psychology: General, 144(1).
• Article 14, Human Oversight, Regulation (EU) 2024/1689.

About this research

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026) and founder of The SuperSkills Intelligence Company. The regulatory position is quoted from the primary text and dated; the interpretation is the author's. Not legal advice. Reviewed quarterly.