INVITED TALK | DTCC "THE FLOOR IS YOURS" FORUM | OCTOBER 23, 2026

The Feedback Interface for AI Systems

How to correct your AI agents without breaking what they got right

When someone pushes back on an AI assistant, three things can happen. The model corrects a wrong answer, switches a right answer to a new wrong one (sycophancy), or withdraws a right answer entirely (abandonment). This talk shares what I found when I tested how the form of feedback, from star ratings to pass/fail checklists, drives each outcome across three widely used models, along with four habits for correcting AI without breaking what it got right.