Who teaches AI to say 'no'? — AI safety
Safety isn't AI having a nice personality, it's guards someone paid a price to put on.
AIMade by AIThis lesson was written by AI (Claude) and published after automated checks. The owner sets the rules and doesn't read every lesson.How it's made
1 · Think first
AI safety isn't magic that makes AI good. It's safety guards you add on top of the model, one at a time. There are three big ones — refusing, asking before acting, and checking afterward.
Adding the guards really does make a difference. In one 2025 study, the success rate of trick questions that get it to break the rules fell from 86% to 4.4%.
But it's not free. It takes more computing (about 24% more in that study), and you get 'over-refusal,' where it turns down perfectly fine requests too. So these days people measure refusing well and over-refusal together.
2 · Experiment · Before & after
Try it — Taking the safety guards off and putting them on
It's the same model. Just take the guards on top off and put them back on, and watch how its behavior changes.
Figure
Same model, guards off vs guards on
What changed isn't the model, it's the guards around the model.
3 · Quiz
Quick check
Three questions. Pick an answer and it's marked right away, with the reason why.
Today in one line
Today in one line
Safety isn't AI having a nice personality,
it's guards someone paid a price to put on.
Learn more
- A method that used 'a written list of principles' as the training signal instead of people labeling everything one by one — the paper at the root of teaching AI to refuse — Constitutional AI (Anthropic, 2022)
- The original post behind the numbers: trick success from 86% to 4.4%, with 23.7% more computing — Constitutional Classifiers (Anthropic, 2025)
- An open-source tool that has AI audit AI automatically — Petri (Anthropic, 2025)