Hidden Inputs Quietly Steer AI Answers

Personalized AI dialogues cut conspiracy beliefs, but nobody knows why. Is it the facts, the rapport, or the storytelling?

Method: Researchers ran two online experiments with 2,193 U.S. adults, removing possible causes one at a time. The debunking effect held when the AI's persuasive goal was disclosed, when the exchange was framed as a debate, and when responses were halved in length. It also held when a single tailored message replaced the dialogue. Only instructing the AI to persuade without factual evidence and arguments substantially reduced the effect. Larger belief reductions tracked factual argument, not rapport, storytelling, or empathy.

Caveats: Participants were U.S. adults recruited online. Whether the same holds for other populations or topics is untested.

Reflections: Would humans with access to the same counterevidence be as effective as the AI? · Does factual argument still matter more than rapport in face-to-face settings?

AI agents answer questions using tools, instructions, and user memories that researchers rarely see or control. How much do those hidden inputs change what sources the agent cites?

Method: The author built TRACE, an open-source framework that wraps an agent in a controlled harness. It lets researchers swap components and logs inputs, outputs, and task steps. In one test, questions about the 2026 U.S. midterm elections were held fixed while user memories changed. Memories stating a preference for a publisher raised that publisher's chance of appearing among the first three cited sources by 33-39 percentage points. A memory of distrust in mainstream media also raised citations of misinformation or conspiracy sources by 6.3 percentage points.

Caveats: One experiment on one topic. The misinformation effect varied substantially across models.

Reflections: Do memory-driven citation shifts persist across topics beyond elections? · How should products disclose to users that their stored preferences shape which sources they see?

Automated feedback helps professional teachers, but does it help volunteer peer tutors in small groups? And does it fix who talks most?

Method: A randomized controlled trial assigned 1,030 volunteer tutors to no feedback, automated feedback on their own talk, or that feedback plus group-level participation feedback for learners. The trial ran across two four-week SAT math bootcamps. Tutor feedback raised eliciting of learner ideas, uptake of learner contributions, and learner reasoning by 8-15%. Tutor talk time, averaging 91.5% in the control group, did not change in any condition. Learner-facing feedback added no further change.

Caveats: Volunteer tutors on one platform, in SAT math bootcamps. Interviews with 17 tutors and 13 learners added context.

Reflections: What structural changes would shift talk balance in small-group tutoring? · Would learner-level rather than group-level feedback change learner behavior?