Why Leaders Over-Delegate to AI: New Memory Research
A new paper from University College London finds that people set more reminders than is actually optimal. For working leaders delegating to AI agents, the noticing prompt sits in the half-second before the click.
A new paper from University College London finds that people set more reminders than is actually optimal. The bias is upward, not downward, and a short training pass can reduce it. Translated to working leaders: the reach for an AI agent is often a reach against forgetting, not a judgment that the agent is genuinely better at the task. The noticing prompt sits in the half-second before the click.
A head of operations described her Tuesday afternoon to me last week. A draft she could write in thirty minutes went to an AI agent. A follow-up she could send in two minutes went to another. By six, she had four threads delegated and a much shorter list of things she remembered to check. The smallest of the four, a follow-up she did not finally see again until Wednesday morning, was the one she had assumed would be easiest to keep track of.
What the research shows
A new paper from the Institute of Cognitive Neuroscience at University College London puts a number on something many working leaders are starting to feel. Ceri Ngai and Sam Gilbert ran two careful experiments, one with 164 people and one with 416, in a controlled task where participants could either hold a future intention in their head or set an external reminder to do it for them. The researchers then compared how often people chose the reminder against the level that would have actually maximized their accuracy given their real, measured memory performance.
People set reminders more than they needed to. In both groups of the first study, the bias toward reminders was statistically significant. The group that received no feedback about its own memory performance offloaded about twice as much as the group that did receive feedback. Even the feedback group was still biased upward. They just biased less.
The training did something. A short intervention asked participants to predict their own performance before each task and then to see how they actually did. That nudge significantly reduced the over-reliance on reminders. The effect size is what cognitive psychologists call small to moderate, meaning real but modest. Not a fix. A recalibration.
The construct this literature has been sharpening for a decade is reminder-setting, not AI-agent delegation. The underlying mechanism is the same: an external aid takes over the holding of a future intention. The new finding is the direction. We tend to offload too much, not too little. This sits upstream of the same vigilance gap researchers found in leaders supervising AI agents earlier this month, the gap that opens inside the first fifteen minutes of the watch. The watch happens after the handoff. The bias the new paper is naming happens before it.
"The magnitude of this reminder bias was significantly reduced in the feedback compared with the no-feedback group, t(162) = 2.22, p = .01, d = 0.347."
The bias is upward, not downward. A leader who feels they are using AI exactly the right amount is statistically more likely to be over-relying than under-using. The question is not whether to delegate. It is which delegations were judgments about the work and which were quiet hedges against forgetting.
What it doesn’t tell us yet
Two studies. One lab. One country. The researchers measured reminder-setting in a delayed-intentions task, not delegation to an AI agent across a working day. The bridge from one to the other is editorial and worth flagging plainly: the study did not measure AI use. It studied online adults setting reminders in a memory task. The recalibration effect was real but small to moderate, and the training reduced the bias without eliminating it. The underlying over-offloading direction is supported by a decade of work from this same lab. The translation to AI-agent delegation will need fieldwork before anyone treats it as settled.
One thing to notice in your work today
When you delegate to an AI agent this afternoon, notice the half-second before the click. Ask whether the delegation is a judgment that the agent is genuinely better at this task, or a quiet hedge against assumed forgetting. The research suggests the second reason is the one most working leaders under-notice. It is one shape of the AI brain fry that comes from oversight, not delegation, only earlier, at the moment of the handoff rather than the moment of the check. The morning review queue that fills overnight gets longer for the same reason. The noticing is the lever.
Sources
- Metacognitive training facilitates optimal cognitive offloading - Cognitive Research: Principles and Implications, 2026-03-12