Feedback loops
Behaviour changes when its consequence returns quickly, specifically, and in a usable unit.
A feedback loop connects an action to information about its effect. Loops work when the delay is short, the signal is attributable to something the person controls, and the unit means something to them. Most real-world decisions have broken loops, and closing one is often a larger intervention than any nudge: A change to how choices are presented that shifts behaviour without removing any option..
How it shows up in software
Loops appear as a password strength meter, a build that fails in seconds instead of twenty minutes, spend shown against budget as you type, and a lint error in the editor rather than in CI. The strongest product loops are the ones that move information earlier, not the ones that add a dashboard.
Using it well
- Shorten the delay before adding detail. Feedback in one second beats richer feedback in one hour.
- Attribute the signal to something the user can change on this screen, otherwise it is just a number.
- Give feedback about the action, not the person. Your query scanned 4 million rows helps; you are a heavy user does not.
- Check the third of cases where feedback hurts. If a metric makes people game it or disengage, remove it.
Where it turns manipulative
- Feedback tuned to produce anxiety, such as streak loss warnings or public activity counts, drives engagement through guilt and is a common source of quiet harm.
- Scores that grade the person instead of the work invite gaming and demoralise the people you most want to keep.
- Selective feedback that reports only the metrics flattering to the product misleads the user about the cost of their own behaviour, for example showing time saved but never money spent.
Where you have seen it
Vercel and similar preview deploys
Every pull request gets a live URL, which turns a review comment into an observation of the running change.
Google Docs collaborator cursors
Another person's edit appears as it happens, so the loop between action and consequence is under a second.
Apple Watch activity rings
Three closing rings give continuous progress in a unit the wearer set, visible without opening an app.
What the research says
- Karlin, Zinger and Ford, 2015 (Psychological Bulletin)Well evidenced
A meta-analysis of 42 energy feedback studies found a small but reliable reduction in consumption, larger when feedback was frequent and given in a meaningful unit.
Average effects are modest, around a few percent, and vary with how the feedback is delivered. Real-time in-home displays outperformed delayed bill inserts.
- Hattie and Timperley, 2007 (Review of Educational Research)Well evidenced
Synthesised feedback research in learning and found effects depend heavily on type: feedback about the task and the process helps, feedback about the person does not.
- Kluger and DeNisi, 1996 (Psychological Bulletin)Well evidenced
A meta-analysis of over 600 feedback effect sizes found feedback improved performance on average, but made it worse in roughly a third of cases.
The most important correction in this area. Feedback that directs attention to the self rather than the task frequently backfires.
Grades are a judgement about the evidence, not about how useful the idea is. Plenty of contested effects are still worth knowing, as long as you do not cite them as settled.