One copy change on screen raised reviewers' rejection rate by 21 points
Build AI Systems for Discernment, Not Approval - Angel Ortmann Lee, Duolingo · Angel Ortmann Lee
26 min total·Actually worth watching closely: ~11 min·3 must-watch clips
- 0:00 – 3:10Listen
Opening: how we trust AI has changed
The speaker opens with the proctoring system behind the Duolingo English Test and sets out the premise of the talk: as AI is embedded in more and more everyday decisions, what the human supplies is no longer judgment but approval.
The target isn't whether the model is accurate; it's whether the human is still willing to think in front of the AI
This stretch is essentially the speaker making an argument over a title slide — put headphones on and listen while walking, you lose nothing▶ Jump to 0:00Speaker · Angel Ortmann Lee - 3:10 – 7:04Listen
Cognitive surrender: numbers that sting
Research puts numbers on cognitive surrender: 80% of participants accepted the AI's answers even when they were wrong, lowering their barrier to entry and trusting the AI without critically examining whether the result was correct.
Rising trust in AI and falling caution are two sides of one thing — when the AI is wrong, people don't just fail to stop it, they get carried along
The numbers are all spoken out loud and there's nothing on screen to look at, so audio is the most efficient way to take this in▶ Jump to 3:10Speaker · Angel Ortmann Lee - 7:04 – 9:30Watch
The fake-alert experiment: 21 points from one piece of copy
The team slipped fabricated AI cheating signals into the real proctoring flow, and highly trained reviewers waved large numbers of them through. The investigation cleared the model (a 1% false positive rate) and cleared the people; the problem was the interface. Changing the guideline copy to say the AI signal is only a preliminary alert and that upholding a flag requires independent evidence in the video took the rejection rate from 50% to 71%.
No model change, no layout change — one line of guidance changed the quality of human judgment, and this is the hardest evidence in the talk
The before and after copy is put on screen side by side; listening alone makes it hard to appreciate how small the wording difference is against how large the effect is, so it's worth pausing to read both▶ Jump to 7:04Speaker · Angel Ortmann Lee - 9:30 – 12:08Skim
The human in the loop is actually a circle
A correction to a common misreading: human in the loop isn't a straight line from model output to human review to decision, it's a closed loop — the interface shapes human behavior, that behavior becomes eval and training data, and it feeds the next generation of the model.
An interface that doesn't force people to think rolls the loop the wrong way: rubber-stamped output becomes false positives in the data set, and that data feeds the next model
The argument here rides on one loop diagram; read the arrows and their direction and you can speed through the narration▶ Jump to 9:30Speaker · Angel Ortmann Lee - 12:08 – 14:07Watch
Headphones: two questions hidden in one button
A teardown of the headphones-detected confirmation button in the proctoring interface: it asks both whether the model got it right and whether this should count as a violation. A test taker wearing a hearing aid makes the two answers come out opposite, so the team split it into two questions.
Cramming a perception judgment and a policy judgment into one button forces a choice, and the price is a falsely accused test taker plus a bad label fed to the model
The before and after versions of the button are placed side by side on screen; one look shows what the extra question looks like, and this is the easiest screen in the talk to copy into your own product▶ Jump to 12:08Speaker · Angel Ortmann Lee - 14:07 – 16:05Watch
The writing tutor: pin feedback to the actual sentence
Using the analogy of asking a friend to look at something you wrote, the speaker shows the writing tutor's approach: no block of commentary, but green, yellow and red markup laid directly on the text, with concise actionable feedback on hover that you can accept on the spot.
The same model output, anchored to a specific piece of text, triggers sentence-by-sentence thinking; piled into a block, it just gets skimmed
The colored markup, the hover feedback and the inline accept all live in the visuals; listening alone gives you no picture of what that annotation experience feels like▶ Jump to 14:07Speaker · Angel Ortmann Lee - 16:05 – 21:12Skim
Two bad patterns in coding agents
Two anti-patterns get named: dumping a giant set of changes in one go so you go back and page through them on a code hosting platform, or pinging you for confirmation on every single change. Either way you become a rubber stamp, and the system only gets low-information clicks biased toward yes. What you want instead is something like a good junior developer: it plans, asks good questions, documents design trade-offs, and breaks changes into reviewable chunks.
The measure of collaboration quality isn't how many times you confirmed, it's whether the system learned anything from your follow-ups and your edits
The visuals are a few diagrams of code review flow with low information density; the argument moves in the narration, so glance at the slides and keep listening▶ Jump to 16:05Speaker · Angel Ortmann Lee - 21:12 – 25:52Watch
Closing: every interaction is already a label
It lands on the most practical point: an approved plan or accepted suggestion is a positive sample, a rewritten or overturned one is a negative sample, and you never have to hire annotators for it. But most systems miss the signal that matters most — the user clicks yes, then goes in and manually changes the output, and that diff is never captured, so a data-polluting false positive walks right into the set. It closes on method: work out what kind of thinking you want the interaction to elicit before you design it.
Measure what the user changed after saying yes; that diff is the most honest and most overlooked training signal you have
The ending opens a screen of real user questions and the speaker walks through what the tone and slant of each one reveals; the words on screen are part of the argument, and skipping them leaves you with only the conclusion▶ Jump to 21:12Speaker · Angel Ortmann Lee