All learning paths Learn See Try Make Run a review round with others What this lesson is about Give three reviewers the same cases, compare where they split, and fix the instruction rather than the people.
Five cases went out. On the last one, two people called a message a threat and one called it a joke. Now what?
A Give three reviewers the same cases, then compare where they split. B Go with the two who agreed, and ask the third to look at it again. C Send two more people the same message and see whether they agree.
Check my recall Why this matters A round is not about catching a bad reviewer. In one published safety study, three reviewers judged the same five hundred cases. Counting the author, agreement across the four scored 0.32 once chance was taken out. It rose to 0.49 when the rating was cut to a yes or a no. The authors put the blame on their own task wording, not on the people.
Run a round on shared cases, find where reviewers split, and rewrite the instruction that split them.
A round is not about catching a bad reviewer. In one published safety study, three reviewers judged the same five hundred cases. Counting the author, agreement across the four scored 0.32 once chance was taken out. It rose to 0.49 when the rating was cut to a yes or a no. The authors put the blame on their own task wording, not on the people.
Give three reviewers the same cases, then compare where they split. When people disagree, change the instruction before you blame the reviewer. Publish the round's result to everyone who took part, including what you got wrong. Task Five safety cases went to three reviewers. Four matched. On the fifth, two called a message a threat and one called it a joke.
Weak approach Goes with the two who agreed and asks the third to look again until she agrees. The round produced a decision, and the sentence that split them is untouched in the guideline.
Stronger approach Writes both readings and shows that the rule never says what counts as a threat. Rewrites that sentence, and sends the outcome to all three.
Why the stronger approach works The split was the finding, and the strong route acts on it. Voting hides it, and the same argument comes back next week.
Try a changed situation Two shopkeepers in your street read the same municipal notice in opposite ways. Which rule still applies?
Make something yourself Run a round on five shared cases with two other people. Record who decided what, the case they split on, and the sentence of the instruction you rewrote afterwards.
Lesson 8 of 8 on Expert, safety & agents next Test and review generated code Part 2 ends in a work sample you can send
Read the primary or official source: Ganguli et al., Red Teaming Language Models to Reduce Harms (arXiv)