Written by Rahim Hirji, author of SuperSkills (Kogan Page, 2026), founder of The SuperSkills Intelligence Company. The research is published openly with every source graded. Found an error? Tell me and it is corrected on the page.
Home / Keynotes / Human oversight and accountability
For boards, risk, audit and regulated industriesHuman oversight is now a legal requirement in the European Union and a control recorded as satisfied in most organisations. Those two facts are not the same fact, and the gap between them is where this keynote sits.
Watch the showreel · 2 minutes
A feel for the room before you put me in front of yours.
This is not an AI policy keynote. What the law should say is properly the ground of people who have written it, and the site's own directory names them. This is the operational question that follows: given the obligation exists, what does an organisation actually have to be able to do, and how would it know it could?
Article 14 of the EU AI Act names automation bias directly, which is unusual in legislation. The five things an overseer must be enabled to do, and why most arrangements fail, are at meaningful human oversight.
Not through negligence. Through reliability. A joint safety study by the UK Marine Accident Investigation Branch and its Danish counterpart, after a run of groundings involving electronic chart systems, found that distrust of the instrument was challenged because discrepancies were rarely encountered. The officers had been trained. What they lacked was reasons to doubt.
And awareness training does not fix it. Dzindolet and colleagues found that explaining why an automated aid might err increased reliance on it. The argument is at human in the loop is not a safeguard, and the ownership question at who owns verification, where in most organisations the answer is nobody.
A test they can apply to their own controls, and the uncomfortable finding that output quality no longer tells you whether the human in the chain is capable. The board version is at what a board should ask about AI, and the audit version at auditing an AI-assisted decision.
Article 14 of the EU AI Act sets out what an overseer must be enabled to do, and names automation bias in the legislation itself. The practical test this research applies is narrower and harder: whether the person exercising oversight could perform the work being supervised, and how the organisation would know. Most arrangements record oversight as satisfied without ever testing that.
The evidence says no. Dzindolet et al. (2003) found that explaining why an automated aid might err increased reliance on it, restoring trust even where that trust was unwarranted. Parasuraman and Manzey (2010), reviewing decades of work across aviation, medicine and the military, found automation bias resists training and worsens under workload; that review predates generative AI. The implication is to design the conditions of the decision rather than the disposition of the person.
No, and the distinction is deliberate. What the law should say is the ground of people who have written it. This is what oversight means operationally once the obligation exists: what an organisation has to be able to do, who owns it, and how it would know the control is real.
Written by Rahim Hirji, author of SuperSkills (Kogan Page, 2026), founder of The SuperSkills Intelligence Company. The research is published openly with every source graded. Found an error? Tell me and it is corrected on the page.
Tell me the audience, the date and the decision. A reply within 24 hours.
Also for boards and leadership offsites, or see ongoing advisory. Browse every topic, audience and region.
No enquiry today, or not yet the right time? The weekly letter is the low-commitment version of the same argument. Weekly essays on AI, capability and the future of work. Read by 25,000 people, every week since 2017. Free, and one click to stop.
Opens Substack to confirm. No pitch in it, unsubscribe in one click, and nobody follows up because you read something.
