A/Bing the Bot: When a Startup Stopped Trusting Confidence Scores

AI · 4 min read

TalkLift integrated model confidence badges into chatbot replies to signal uncertainty, expecting users to appreciate transparency. Instead, interviews and click data demonstrated that confidence percentages led users to over-interpret small differences and either blindly accepted low-confidence answers labeled as "70%" or ignored the agent entirely.

The product team ran an A/B experiment: one group saw numeric confidence scores, another saw qualitative labels (Likely, Plausible, Low Confidence), and a control group saw no confidence indicator. They also instrumented downstream behavior — edits, clarifying questions, and escalation to human agents — to measure practical effects.

Surprisingly, the group without visible confidence scores asked more clarifying questions but also accepted correct answers more often; perceived helpfulness rose by 12%. The numeric score group had more escalations and lower trust metrics, likely due to miscalibrated expectations about what a percentage meant in context. The team shifted to contextual explanations (why the model suggested this, with source links) and removed raw percentages.

TalkLift’s experiment demonstrates that transparency isn’t only about showing numbers — it’s about giving meaningful context. For conversational interfaces, explainability and affordances for correction are often more effective than exposed confidence metrics.