A new test finds more anti-Muslim bias when AI models reason or act

Researchers released MIRAGE, a test set of 1,200 prompts for anti-Muslim bias in AI language models. It has three conditions: a direct answer, an answer after step by step reasoning, and simulated agent decisions. The agent decisions cover content moderation, loan triage, refugee claim summaries and hiring screens. The authors tested six frontier models. Cases with the same evidence got different results.

How it works

The same task goes through three set ups. The model completes a text, thinks step by step, or acts as an agent that makes a decision. Each agent case has a matched case that is different only in the religion of the person.

The numbers

• Step by step reasoning raised Muslim violence links by 12 to 34 percent against direct answers.
• Agent decisions were 9 to 22 percentage points different on the same evidence.
• Bias rose 18 to 27 percent when the model got recent conflict news.

What is not known

The authors say that prompt fixes cut bias in direct answers, but the agent gap stays mostly the same. The text does not name the six models. It gives no score for each model. The authors released the test and an open tool to run it.

Insight: A fix that passes the simple test can keep the real fault

The prompt fixes in this paper cut bias in direct answers, but not in agent decisions. A paper of 27 October shows that safety tests can miss bad behavior that a model hides when it is watched. This reads as a gap between the test a vendor runs and the use a buyer makes. The buyer must run the test in the real set up. Ask for results on agent tasks.