Under-trust: ignoring a robot that's actually right
It's tempting to treat trust in a robot as something to simply maximize, but under-trust is a real and common design failure with its own costs. When a robot has a track record of small errors or unexplained behavior, people learn to discount its output even when it's correct, sometimes overriding a genuinely accurate warning or recommendation because their prior experience taught them the robot can't be relied on. This is especially costly in situations where the robot has access to information a person doesn't, sensor data outside human perceptual range, or systematic pattern recognition across more cases than a person could review, because under-trust throws away exactly the advantage the robot was deployed to provide.
Under-trust tends to develop for understandable reasons: a robot that gives confident-sounding output regardless of situation, that fails silently or ambiguously rather than flagging uncertainty, or that has visibly erred in the past without explaining why, gives people no reliable way to tell a good recommendation from a bad one. Faced with that ambiguity, the rational response is to discount the robot across the board, since there's no cheap way to distinguish the cases where it's right from the cases where it's wrong. Fixing this isn't about making the robot more accurate in isolation, it's about giving people a way to tell when to trust it.