Easing nervous first-timers into a live argument with an AI coworker
Cornell Future of Learning Lab, 2026
When your coworker jumps in mid-thought. Participants practice handling conflict at work with AI coworkers who talk back. I redesigned the onboarding, built the researchers’ replay view and measured the coworkers in live voice tests.
162 issues fixed before data collection
One was that session recordings could be opened without a login.
- Role
- Undergraduate researcher and full-stack developer, AI-Ready Workforce Initiative
- Team
- The lab’s principal investigator and research staff, with Anthropic Research
- Timeline
- August 2026 to now
- What changed
- First-timers used to stall before the call. Now they get through.
In short
- The study runs with Anthropic Research. I brought in five new company partners.
- The coworkers speak through Gemini Live but had only been measured in text. Out loud, they jumped in mid-thought.

The numbers
My part
- Redesignedthe onboarding, from consent to the live call, and changed it wherever a first-timer got stuck.12 participants
- Builtthe Evidence Trace, which shows each turn of a session as one row.Researcher dashboard
- Measuredthe coworkers live and fixed the turn-taking and audio problems that came up.False cut-offs 22% to 0%
Sam couldn’t tell your voice from keyboard noise.
- Cut off with nobody speaking
- 22% → 0%
- Real interruptions that got through
- 25% → 56%
Sam is talking, and a keyboard is clicking beside the participant’s microphone. In the first version, the voice detector had a fixed threshold and took the clicks for the participant. Sam stopped mid-sentence with nobody speaking, and the rest of the reply was never heard. On Sam’s next line, the participant did cut in, with “Sorry, can I stop you there?” The detector was still busy with the typing and missed it, so Sam talked over them. With room-adaptive voice detection, the threshold is set from the room’s own noise. Sam finished the reply and stopped as soon as the participant spoke.
Sample lines from the study’s Taken credit scenario. The numbers come from the live tests. Turns cut off with nobody speaking went from 7 of 32 to 0 of 35. Real interruptions that got through went from 5 of 20 to 10 of 18.
How long should Sam wait before answering?
- 0.9 s. Sam answered before you’d finished your thought, then got cut off when you kept going. With the first settings, every hesitant line in the live tests split this way: 18 of 18.
- 1.5 s. Sam waited through both pauses and answered 1.5 seconds after you finished. That’s longer than every pause the live tests measured (0.7 to 1.3 seconds), and it’s the longest wait that kept a reply from feeling slow. It’s the setting on Gemini Live now.
- 2.5 s. Sam waited through both pauses too, but then left 2.5 seconds of silence before every answer.
A participant who hesitates says, “So, the deck this morning… the analysis was mine… and I’d like my name on it,” with a 0.8-second pause and then a 1.2-second pause. Sam answers once the silence lasts as long as the wait you pick.
Sample lines from the study’s Taken credit scenario. The pauses are as long as the ones measured in the live tests.
Clips



The problem
The platform has two kinds of users, and they never meet.
One is a nervous participant in a headset who’s about to argue with an AI coworker named Sam. Each coworker has two parts: an actor the participant hears, and a director that tells the actor how to play each turn.
The other is the researcher. They need to know what happened in every turn, down to whether a director’s note reached the actor late.
The first version failed both of them. Participants stalled between consent and the call, and researchers couldn’t line up the audio with the logs.
Decision log
to measure the coworkers live, with a scripted participant who hesitates and gets lost.
Rejectedtext-only tests.
Whythe coworker jumped in during mid-thought pauses. Room noise, like a keyboard by the microphone, also cut off its own replies. The text tests hadn’t shown either problem.
Resultthe coworker now waits through a pause. With room-adaptive voice detection, false cut-offs fell from 22 percent to zero.
to pick the voice model by talking to the candidates.
Rejectedtheir spec sheets and published benchmarks.
Whywhat matters is whether a coworker sounds like it’s listening and whether the director’s notes reach it. No benchmark measures either.
Resultthe live trials found that the director’s notes never reached the Gemini coworkers. The same setup also caught bugs that were dropping participant audio.
the Evidence Trace. It gives each turn one row, with the participant’s words, the director’s note and the actor’s reply on one timeline. It also has fidelity checks, and flags for notes that arrived late or leaked into the conversation.
Rejecteda transcript plus a separate log download.
Whyif a researcher scoring a session can’t see that a note arrived late, they’ll read the delay as the character’s personality. That error would carry into the study’s conclusions.
Resultresearchers can replay any session against its logs in one view. The build marks each note as steered or unsteered. The Late flag needs the platform to time its notes, which it doesn’t do yet. The leak flag isn’t built.

to strip the call screen down to what a participant needs: who they’re talking to, that it’s recording, and a way to leave.
Rejectedthe usual video-call controls.
Whya self-view, a mute button and a timer all gave nervous participants something to fiddle with instead of talking.
Resultfirst-time participants now get through the flow. The revisions came out of the 12 think-aloud sessions. The build kept a self-view and a running timer on the call.







Not in the first version
These were designed but left for later: a live view of several sessions at once, analytics across sessions, a prompt editor, and rater assignment and export.
What I would do next
I’d run the researcher side through the same 12-person usability protocol. So far, the dashboard’s only tester is me.



