← Demos
Demo · Detecting AI lies
Detecting AI lies
Assistant UIs render guesses and verified facts with the same weight. This replay shows how
inline marks — dashed for heuristic, solid for researched — would have flagged a wrong fork
slug before it burned three turns.
Essay
·
Docs
·
Literature seed
Replay scene Next line
Provenance on.
Real scene that seeded this idea — same story as the replay above.
What you are seeing Heuristic tokens pop first — dashed underline on
AMDphreak/clients before any GitHub check.
Researched tokens stay quiet — solid underline on
AMDphreak/bitwarden-clients after verification; hover for source.
Upgrade in place — when the agent corrects itself, class changes are visible
in the tokens, not only in the apology sentence.
People first — the mark answers “did it check?” without reading the tool log.
HCI Nerdz
