No, and the reason is narrower than the usual one. Jakob Nielsen's October 30, 2026 roundup says fake users produce fake data, reports Yuanzi Li and co-authors finding aligned language models too agreeable to stand in for users, and says five real users with the traits that break designs find more problems than 50 imagined ones. My onboarding agent watched real first sessions, found every real stuck point, and offered help at those moments; a twin built from that behavior would have had the stuck points right and waved it through. Users read it as surveillance, activation among the users who got the help sat below the control group, and it cost a quarter plus the build and the removal. The twin's data was true about where and empty about how it felt, and the real users' behavior data was empty there too. What held the reaction was asking. So a synthetic user may fill the where column of the evidence log, and only a real person, asked, may fill the felt column; a decision needs both.
No. A twin may fill the where column. Only a real person, asked, may fill the felt column.
The Nielsen section, the onboarding agent, and the two-column rewrite are in Five Real Users Beat Fifty Synthetic Ones, If You Ask Them How It Felt. The long version of the onboarding story is The Twin Would Have Shipped It, and where synthetic users sit in the weekly research cadence is in Research at Generation Speed(coming Dec 16).
Last reviewed 2026-10-30 · 1 min read