Claude Code is scanning your messages for curse phrases

“WTF?” “Dammit!” “Now I’m actually irritated.” Cursing out a flailing AI helper is one thing we’ve all finished, however it seems probably the most in style Claude instruments is actively checking our messages for particular indicators of frustration—together with swear phrases.

The revelation is one in every of many startling particulars stemming from a large Claude Code leak that laid naked lots of Anthropic’s plans for its upcoming instruments and fashions. The greater than 500,000 strains of code, which Anthropic by chance revealed to a public software program registry on Tuesday, contains every kind of juice particulars, together with specs for brand new Claude fashions, an “undercover mode” for Claude that permits it to make “stealth” contributions to public code bases, an “always-on” agent for Claude Code, and even a Tamagotchi-style “Buddy” for Claude.

However one of many stranger bits found within the leak is that Claude Code is actively watching our chat messages for phrases and phrases—together with f-bombs and different curses—that function indicators of person frustration.

Particularly, Claude Code features a file referred to as “userPromptKeywords.ts” with a easy pattern-matching device referred to as regex, which sweeps each message submitted to Claude for sure textual content matches. On this specific case, the regex sample is looking forward to “wtf,” “wth,” “omfg,” “dumbass,” “horrible,” “terrible,” “piece of —-” (insert your favourite four-letter phrase for that one), “f— you,” “screw this,” “this sucks,” and several other different colourful metaphors. 

To be clear, this “swear phrase” search perform was noticed particularly in Claude Code through the Anthropic leak. The code for Claude’s desktop and internet apps wasn’t included within the leak, so we don’t know what’s happening underneath the hood of these Claude interfaces.

It’s additionally price noting that the regex perform used right here is nothing fancy. Certainly, regex is supported by all kinds of programming languages (every thing from Java to Python) and has been in use for many years, and its performance isn’t far faraway from a plain-old Ctrl-F.

Whereas the Claude Code leak revealed the existence of the “frustration phrases” regex, it doesn’t give any indication of why Claude Code is scouring messages for these phrases or what it’s doing with them.

I’ve reached out to Anthropic for remark.

In fact, we will guess. One believable motive is that Anthropic is gathering telemetry to assist it decide how particular Claude fashions and instruments are performing. A lift in detected “frustration” phrases can be a simple method to spot if a brand new Claude characteristic isn’t going over properly.

One other risk is {that a} spike in detected frustration indicators may set off behavioral modifications in Claude itself, maybe making it extra empathetic or apologetic. In fact, swearing at Claude or another AI will usually steer the chat in a unique route (because it did for Google Search outcomes), however a selected regex verify just like the one in Claude Code may assist make a behavioral pivot extra dependable.

Whereas the “frustration phrases” regex has solely been confirmed for Claude Code, it makes you ponder whether it exists in Claude’s desktop and internet apps—or whether or not ChatGPT, Gemini, and different large AI gamers harbor comparable features of their code bases.