The five-star review that was a warning

The five-star review that was a warning

It was a confident quarter when a software brand's sentiment dashboard glowed green — until an analyst manually read a sample and felt the floor tilt. The model was reading sarcasm as praise.

Lines like 'love waiting 40 minutes for support, truly five stars' scored as glowing positives. The automated read said 76%; the human spot-check of 400 mentions found roughly one in five 'positive' posts was bitterly sarcastic, concentrated entirely around a support-wait complaint the dashboard had buried under good vibes.

Corrected, sentiment was 61% and falling, with a clear, single cause they'd been blind to for weeks. They staffed up support; the sarcasm cluster shrank by half in a month.

The takeaway: sentiment models are confident precisely where they're wrong — sarcasm reads as enthusiasm. Sample by hand regularly. A green dashboard that's never been spot-checked is a guess wearing a lab coat.