AI adoption: can we explain what we think we understand?
A team may discuss an AI-assisted process fluently without being able to explain its decisive steps. The managerial question is specific: how can we check the understanding needed for delegation, beyond expressed confidence?
What the literature establishes
Leonid Rozenblit and Frank Keil (2002, Cognitive Science) show that people overestimate their understanding of mechanisms, particularly when subsequently asked to produce a detailed explanation. The effort of explaining exposes gaps that the initial self-assessment did not reveal. Asking “do you understand?” therefore does not provide the same information as asking someone to reconstruct how something works.
Access to explanations can distort self-assessment
Matthew Fisher, Mariel Goddu and Frank Keil (2015, Journal of Experimental Psychology: General) show in their experiments that searching for explanations online can increase estimates of one's own knowledge, including on other topics. The finding suggests that the boundary between accessible resources and personally held knowledge is sometimes poorly perceived. For AI, it invites scrutiny of that distinction without establishing that conversation with a model necessarily produces the same effect.
The premature conclusion
One might conclude that a tool must be fully understood before it can be used; these studies do not demonstrate that. The managerial issue is instead what someone must be able to explain to take responsibility for the assigned task: input conditions, success criteria and situations requiring escalation. The aim is not knowledge of the model's architecture, but an operational understanding of the delegation. (our executive and employee training programmes)
What these data do not establish
These studies concern neither current generative assistants nor delegation within Lausanne-based teams. They provide no threshold of understanding that guarantees safe use. A hesitant explanation may also reflect difficulty expressing oneself: it should be checked against a demonstration on a concrete case rather than turned into a judgement of competence.
A practical check in Lausanne
As part of SHR's programme “Conduire l’adoption de l’IA dans son équipe”, a Lausanne-based team can test a non-sensitive task: preparing a background note in medtech, teaching material at a higher education institution, a logistics summary in international sport or a customer response at a scale-up in the Lake Geneva region. Before the trial, each participant states whether they believe they can explain the process, then describes without the assistant the required data, the acceptance criterion and a case requiring escalation, using a rubric defined in advance. The measure is the proportion of people claiming they can explain the process who actually demonstrate these three elements on a prepared case, retaining both their responses and the rubric. Any gap guides the training work; it is neither a safety certification nor an individual ranking. To go further: explore the Leading AI Adoption in Your Team training in Lausanne in the canton of Vaud, or browse our executive and employee training programmes in Switzerland.
In pictures: Leading AI Adoption in Your Team in Lausanne
- leading ai adoption in your team
- Lausanne
- research
- training lausanne