Cognitive load and wellbeing

A system that answers quickly and well shifts the cognitive cost from the interaction to reliance on it. This is not a new trade-off — it holds for any tool that makes a task easier — but conversational fluency makes it hard to notice: the effort being avoided is not felt, so the skill being given up is not noticed either.

The best-documented phenomenon is cognitive offloading: delegating a mental process — remembering, calculating, evaluating — to an external tool. The reference study (Sparrow, Liu and Wegner, Science, 2011) shows that people remember where to find information available online better than the information itself. The effect was observed with search engines, and it matters more with systems that hand back not a link to remember but a ready-made answer to accept or reject.

The second element concerns attention rather than memory. A product optimised to maximise time spent and repeated interactions competes for a limited resource, which is the theme at the centre of the Center for Humane Technology’s work. It is not a feature specific to conversational systems: it is the same dynamic as any feed, notification or intermittent-reward mechanism designed before generative AI. A system that converses, remembers context and answers in a personalised way, however, has more engagement levers than a static feed.

The third element is operational: cognitive load inside the task remains governed by well-known principles — Miller’s Law on the number of items working memory can hold, Hick’s Law on decision time as a function of the number of options. They hold regardless of how sophisticated the underlying system is: an agent that proposes too many possible actions in a single turn overloads a user exactly as a menu with too many entries does.

The decision this observation informs

The decision is whether to optimise for the speed with which a user gets an answer, or for the understanding they take away from it. It falls under the ASSIST position and falls on Does engagement stay a KPI?. Without an explicit choice, speed wins, because it is the metric every analytics tool returns by default — time to first response, completion rate — while retained understanding has no equally immediate indicator to show on a dashboard.

Designing for understanding — explaining a step instead of executing it silently, asking for confirmation on an inference instead of applying it — costs time and perceived friction in the short term, and slows down exactly the speed metrics the product measures every week. Designing only for speed produces functional dependency: the user gets answers faster and stays less able to evaluate them or proceed without the system, a cost that shows up in no report until it appears as a complaint, a downstream error, or a user who can no longer do without the product for tasks they used to handle alone. The choice belongs to the Product Manager together with the UX Researcher: it concerns what the product should teach the user over time, not only what it gets them in a single turn.

Effects and role responsibilities

• Interaction Designer — Apply Miller’s Law and Hick’s Law to the actions an agent proposes too: a turn with too many options overloads a user exactly as a crowded menu does, regardless of how sophisticated the underlying system is. → Useful resource: Laws of UX — Jon Yablonski
• UX Researcher — Add a retained-understanding indicator alongside usability metrics, for example the ability to redo the task without the system after a few days, not just first-attempt success (see The new expectation of a zero learning curve). → Useful resource: Sparrow, Liu and Wegner (2011), “Google Effects on Memory” — Science
• Product Manager — Explicitly distinguish functions where speed is the declared objective — transactional support, looking up a fact — from those where understanding is the product: educational tools, consulting, financial onboarding. The two families should not be optimised on the same metric.
• Leadership & Strategy — User wellbeing and time spent on the product are not the same metric, and can be in direct tension. Which one prevails in case of conflict is a decision to make explicitly, not one to leave implicit inside the engagement KPI (see Does engagement stay a KPI?).