Sense and decide
Say What You'd Check
Explain your own checks so you remember them at review time
Each person writes down, in their own words, what they plan to check in the AI output. Put that list in front of them again when they review it.
Borrowed from a 2026 preprint by Fu, Ramasubbu, and Galletta reporting two randomized experiments on keeping people alert to AI errors. Read the original source.
One session · Minutes to start · Anyone on the team can start it
Try this first
Ask each person to write three sentences on what the AI tool tends to get wrong in their work and how they would catch it. Put the most useful points where the output gets reviewed.
Use it when
People were trained on what to watch for in AI output, but errors are slipping through as the tool becomes routine.
Skip it when
Skip it when AI output is always checked by an automated test that does not depend on a person paying attention.
How to introduce it
When the team starts using an AI tool, or when you introduce this move to a team already using one, ask each person to write a short paragraph. It should say what kinds of mistakes this tool is likely to make in their work and what they would check to catch them. Have people read each other's paragraphs and add anything they missed. Then turn the most important checks into a short reminder that appears where people review the AI's output. That could be a line at the top of a review template, or a note pinned in the channel where the output arrives.
How to show up
Ask for people's own words, not a copy of a list. The benefit comes from the act of explaining. Keep the reminders short, and change them from time to time, because people stop reading a reminder that never changes.
How long it takes
Fifteen minutes for the writing and comparing. A few minutes to set up the reminders.
What makes it hard
People tend to write general answers, such as "check for accuracy." Ask for specific ones, such as "check that the dates match the source document." Reminders also lose their effect if they appear too often.
What it looks like when it's working
People can name specific checks without looking them up. The number of errors that get through stays steady instead of slowly rising as the tool becomes routine.
How long until it sticks
One session to write the checks, plus a refresh every few months or whenever the tool changes.
How you know it stuck
New team members write their own checks as part of starting on a tool, and the reminders get updated when new kinds of errors appear.
The idea behind it
Training people once is not enough, because attention fades as a tool becomes familiar. Explaining something in your own words makes it easier to recall later. A reminder at the moment of review brings that knowledge back when it is needed.
Where it comes from
Human-AI oversight research. Two randomized experiments with 640 customer-facing employees tested how to keep people catching AI errors over time. The paper was posted in 2026 and has not yet been peer reviewed. People caught more errors when the knowledge they needed for checking was easy to recall at the moment they reviewed the AI's output. Asking people to explain the checks in their own words during onboarding improved how many errors they caught. Brief reminders kept their error detection from declining as they used the tool again and again. The study's main point is that knowing what to check is not enough. People also have to remember it at the moment of review. When people start using an AI tool, each person writes, in their own words, what they would check in its output and why. Later, short reminders of those checks appear at the moment people review the AI's work.