Confirmation bias
People look for evidence that fits the belief they already hold.
Once a hypothesis is in mind, search, interpretation, and recall all tilt toward supporting it. The strongest and best-documented component is the search tilt: people test cases that would confirm the rule rather than cases that could break it. Interpretation effects, where the same evidence strengthens opposing beliefs, are more variable.
How it shows up in software
Confirmation bias mostly damages the team, not the user. Analytics dashboards get built to show the metric that justifies the roadmap. Interview guides ask leading questions. A launch that underperforms gets a post-mortem that finds a measurement problem. In user-facing surfaces it appears as personalised feeds that keep narrowing toward what someone already engaged with.
Using it well
- Write the disconfirming result before you run the test: name the number that would make you kill the feature, and log it.
- In discovery interviews, ask about the last time the user did the thing rather than whether they would like a solution.
- Give dashboards a counter-metric next to the headline metric, so a rise in engagement shows up next to whatever it might be costing.
- Assign someone in a launch review to argue the opposite reading of the data, and give them the same data access.
Where it turns manipulative
- Personalisation that optimises purely for engagement narrows what a person sees and hardens what they already believe. Recommender systems are a live public policy problem for exactly this reason.
- A/B test: Showing two variants to randomly assigned groups to see which performs better. dashboards that let anyone peek and stop early hand the team a confirmation machine dressed as rigour.
- Feedback widgets that only appear to users who completed a flow successfully collect a sample that cannot contradict you.
Where you have seen it
YouTube
Recommendations weigh watch history heavily, and the settings offer controls to pause or clear that history.
Optimizely
Experiment results display statistical significance and warn against calling results before the planned sample is reached.
X
The timeline offers a chronological following feed alongside the algorithmic one as a separate tab.
What the research says
- Wason, 1960Well evidenced
Given the sequence 2-4-6 and asked to discover the rule, participants proposed triples consistent with their guess and rarely tried a triple that would falsify it, so most announced a wrong rule confidently.
- Wason, 1968Well evidenced
In the four-card selection task, most participants turned the card that could confirm the conditional rule and neglected the card that could falsify it.
Performance improves dramatically when the same logical task is framed as a social rule about permission, so this is partly about abstraction, not only about bias.
- Lord, Ross and Lepper, 1979Contested
Participants with opposing views on capital punishment read the same mixed evidence and reported more extreme versions of their original positions afterwards.
This attitude polarization result has a patchy replication record, and later work often finds belief updating rather than polarization. Cite the search-tilt findings when you need a solid claim, and be careful with 'evidence makes people more extreme'.
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.