Anthropic tracks how often users swear at Claude
An engineer proposed it last fall after the team kept seeing frustration, and Fung still calls it the swear word dashboard.
Lenny's Podcast: Product | Growth | Career · What happens after coding is solved? | Fiona Fung (Claude Code & Cowork)Anthropic keeps a dashboard that counts how often users swear at its AI assistant out of frustration.
Fiona Fung, who runs Claude Code and Cowork at Anthropic, said the idea came from an engineer on her team last September. The team had been seeing frustration, and the engineer suggested tracking swear words. Fung, who had just joined, remembers it sparking a really fun conversation.
She still calls it the swear word dashboard. She said it shows why evals, the tests used to grade AI systems, are hard: they struggle to capture the user experience, and whether using the product feels delightful or frustrating.
Swearing, in her telling, is a fun way to look at that frustration.
Questions people ask
What does Fung mean by evals?
Evals are short for evaluations, the structured tests AI companies use to grade their models. They typically check accuracy, reasoning, coding skill, and safety on fixed sets of questions. Fung's point is that those scores can miss how the product actually feels to use.
Why would swearing be a useful product metric?
Swearing is spontaneous and emotional, so it can surface frustration that users never bother to report in feedback forms. Product teams often track indirect signals like retries, cancellations, or angry messages alongside star ratings. A spike in profanity can flag a bad release faster than a formal survey.
Do other software companies track frustration this way?
Yes, many track rough language, rage clicks, repeated refreshes, or deleted messages as proxies for frustration. Game studios, customer support teams, and app developers routinely mine chat logs and reviews for profanity. It is cheap, easy to count, and hard for users to fake.