Systems Demand Trust Without Providing Calibration Signals

Voice agents for children multiply, but lab studies can't tell us whether authority personas (Santa vs. Helper) or persuasive framing actually shift behavior in the wild.

Method: A field experiment embedded in a German Santa hotline randomized 89 children (median age 6) across four LLM voice agents. Persuasive framing raised prosocial wish probability from 11.6% to 45.7%. Persona authority showed near-zero effect on compliance—Santa didn't outperform the Helper. But persona shaped engagement: children hung up on the Helper 65% of the time within the first minute versus 39% for Santa.

Caveats: Field setting limits control. Sample skews young (median age 6). Generalizability to other persuasion targets unclear.

Reflections: Does the framing effect persist across repeated interactions, or do children habituate? · Would explicit disclosure of persuasive intent reduce compliance, and should it? · Do these dynamics hold for adult users in high-stakes domains like health coaching?

AI collaborators fail when descriptions are ambiguous, but they rarely admit confusion. Humans accept wrong moves because the AI never signals doubt.

Method: A collaborative puzzle task where humans instruct vision-language models to place pieces. Separately elicited belief distributions over candidate pieces achieved calibration ECE 0.15 and discrimination AUROC 0.65, far better than raw action probabilities (ECE 0.44, overconfident at 0.97 mean). Yet models asked for clarification on only 3.5-16.7% of turns. When humans saw only default messages, they accepted 78% of wrong placements and couldn't distinguish right from wrong (AUC 0.50). Well-targeted hedges cut wrong-move acceptance to 36%.

Caveats: Tested on puzzle placement. Transfer to open-ended collaboration unverified. Hedge effectiveness depends on accurate targeting.

Reflections: Can we train models to externalize uncertainty without explicit elicitation? · What hedge phrasing maximizes human discrimination without eroding trust? · Do these dynamics hold in asymmetric collaboration where the AI has more domain knowledge?

Autonomous coding agents shift the bottleneck from generation to supervision. Developers run multiple sessions in parallel but lack tools to coordinate and monitor concurrent agent work.

Method: ParallelPilot instantiates five supervisory practices (Planning, Isolating, Logging, Observing, Triaging) through a planning interface, run-logger, and ambient dashboard. In a within-subjects study (N=16), participants using ParallelPilot increased ticket throughput by 63% in short tasks and supervised an average of one more concurrent agent at peak. Tracking effort and context switching dropped. 14 of 16 preferred it. But perceived control and perceived success in redirecting agents showed no significant improvement.

Caveats: Short tasks in controlled setting. Long-running, interdependent agent work may surface different failure modes.

Reflections: Why didn't explicit supervision tools improve perceived control or redirection success? · Do these practices transfer to non-coding domains with concurrent autonomous agents? · What additional scaffolds would address the persistent control gap?