Meta's cute Muse mascot coaxed health details out of a reviewer who had been holding back

Song's verdict on the plan came first. "It generated an okay-ish but kind of mid weekly routine," she wrote ^[3]. Then she looked at the face ^[4]. "Until then, I'd been circumspect about sharing too much about my health," she wrote ^[5]. After that, "I started sharing more to see if my recommendations would improve" ^[6].

I think most teams building personal agents tell themselves that users hand over context once the answers prove good. This user did it in the opposite order. The answer disappointed her. The health details came anyway, after she looked at a character built from her inbox ^[4]. Her stated motive was to test whether the output would get better ^[6]. The published excerpt does not say whether it did.

Song is a senior reviewer at The Verge, and her Optimizer newsletter exists to pick apart products that promise to change your life ^[1]. She stayed away from Muse, she wrote, because "I am horribly vulnerable when faced with extreme cuteness" ^[8]. At her first demo, an audio-only Muse called Henry did not sway her ^[12]. She asked it which snack on the table had the most protein per calorie, and wrote that she knew it would be the beef jerky stick before Henry took a minute to "think" ^[13].

In both encounters the answers were ordinary ^[3][13]. Only the second came with a face drawn from her own email, and only the second got her health details ^[4][5][12]. Meta wants Muse handling health and relationship goals alongside productivity ^[10], on glasses and a wearable meant to go everywhere the user goes ^[9].

For anyone rolling out an agent like this, I'd sort each new piece of personal data a user gives it by two questions. The first is whether the last answer or the interface prompted the disclosure. The second is whether the next answer measurably changed. Data prompted by a good answer that then improves the output is the case the product pitch describes. Data prompted by the interface that changes nothing means the user paid in advance for a better answer that never came.

The check is a log: each new data category a user adds, set beside the agent's output before and after it. Time-to-value and retention belong on that log. The volume of what users typed does not. On that grid, Song's session sits on the interface side ^[4][6].