We Audited Our Own Chatbot Detector. On Our Own Headline's Question, It Was Right Once In Eight.

We Audited Our Own Chatbot Detector. On Our Own Headline's Question, It Was Right Once In Eight.

We Audited Our Own Chatbot Detector. On Our Own Headline's Question, It Was Right Once In Eight.Uldis Zalcmanis
Published on: 03/09/2026

Every chat-coverage number we have published rests on one rule, and nobody had ever checked whether it works. We pre-registered the test, published the result whatever it said, and it says our rule finds a visible chat control about six times in ten and detects an AI assistant almost never.

Field Notes
Read more
Why Does A Chatbot That Worked Six Months Ago Start Giving Wrong Answers?

Why Does A Chatbot That Worked Six Months Ago Start Giving Wrong Answers?

Why Does A Chatbot That Worked Six Months Ago Start Giving Wrong Answers?Uldis Zalcmanis
Published on: 03/09/2026

Three causes of bot decay, none of which appear in a change log: the model underneath is retired and silently swapped, a fact that was typed in stops being true, and a link in the instructions stops resolving. Plus the benchmark of ours we are withdrawing.

Answers
Read more
Do You Have To Tell People They Are Talking To An AI?

Do You Have To Tell People They Are Talking To An AI?

Do You Have To Tell People They Are Talking To An AI?Uldis Zalcmanis
Published on: 03/09/2026

The EU duty became applicable on 2026-08-02 and attaches to the function, not to a risk tier. Malaysia has no equivalent statute. But the measured research on what disclosure costs, and what being caught costs, is the part that should change what you build.

Answers
Read more