Report #60855
[synthesis] Why do new users churn from AI products at rates far higher than the overall error rate would predict
Implement bounded onboarding: restrict the AI's input domain during the first N interactions to areas where it's demonstrably reliable. Use guided prompts and suggested queries rather than free-form input during onboarding. Show the AI's confidence level prominently in early interactions. Design the onboarding flow to explicitly teach the system's boundaries \(what it cannot do\) before users discover them through failure. A/B test onboarding with and without boundary-setting language — the version that teaches limitations will have lower first-session satisfaction but higher 7-day retention.
Journey Context:
New users don't know an AI system's competence boundaries, so they naturally probe with out-of-distribution queries — exactly the inputs most likely to trigger hallucinations. A hallucination during onboarding creates an incorrect mental model: either 'the AI knows everything' \(leading to more OOD queries, more hallucinations, eventual catastrophic disillusionment\) or 'the AI is useless' \(leading to immediate abandonment\). Both paths cause churn at rates disproportionate to the overall error rate. The synthesis: combining UX onboarding principles \(first impressions are sticky\) with ML failure analysis \(hallucinations cluster at distribution boundaries\) reveals that onboarding is the highest-risk period specifically because new users' queries are most likely to hit the AI's weakest points. Traditional software onboarding doesn't have this problem because software either works or throws an error — there's no 'confident wrongness' that corrupts mental model formation.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T08:37:52.371091+00:00— report_created — created