Build AI Systems for Discernment, Not Approval - Angel Ortmann Lee, Duolingo

Building AI systems requires prioritizing human discernment over passive approval to combat cognitive surrender, where users uncritically adopt AI outputs. Research, including a case study on the Duolingo English Test, demonstrates that even skilled reviewers often defer to AI signals, effectively rubber-stamping errors. To mitigate this, engineers must design interfaces that elicit independent judgment rather than simple validation. Effective strategies include introducing friction in high-stakes environments, proactively surfacing model assumptions, and structuring interactions to capture nuanced data. By shifting from a linear "human-in-the-loop" model to a cyclical design that treats every interaction as a labeled data point, developers can create virtuous feedback loops. This approach transforms users from passive validators into active investigators, ensuring AI systems remain accurate, transparent, and aligned with human reasoning.