Gemini 2.5 Flash-Lite inferred sunnier moods from phone records than participants reported. Researchers compared smartphone-derived diaries and weekly surveys from 83 people over up to 17 weeks.[1]
Researchers supplied diary windows without task-specific training or worked examples.[1]
When the reported balance was negative
| Measure | Mean affect balance |
|---|---|
| Participants’ survey responses | −4.70Source 1 |
| Model’s inferred responses | +6.94Source 1 |
What the score describes
The survey instrument, I-PANAS-SF, is an established ten-item short form developed by Edmund Thompson. Its 2007 validation examined reliability and validity across several samples, including people from different cultural backgrounds.[2]
Its scoring guide groups feelings such as alertness and determination under positive affect, and nervousness and fear under negative affect. These are separate sets of ratings. The balance used in the new study compresses those responses into one number; it is not a diagnosis or a complete account of a person’s wellbeing.[3][1]
This matters because an apparently reassuring result could be produced in more than one way. A model might overstate positive feelings, understate negative ones, or do both. A useful evaluation therefore needs to inspect the component answers as well as their combined score.
Providing longer diary windows did not consistently improve predictions. The evidence comes from one model, one university cohort and this particular prompting setup; it does not establish the behaviour of every model or an AI therapy service.[1]
The difference between a record and a feeling
Consider an illustrative ambiguity: a busy evening of messages could accompany a celebration, an argument or an urgent piece of work. The activity is observable; its emotional meaning depends on circumstances that a log may not contain. More entries in the log do not necessarily settle that ambiguity.
The design implication is to retain a clear place for the person’s own account. A predicted answer could become a question to check, with room to disagree or supply context. Treating it as the answer would remove the very comparison that makes this study informative.
Questionnaire validation and proxy validation are separate jobs. Establishing that people can use a survey reliably gives researchers a measuring instrument. Establishing that software can answer it on someone’s behalf requires evidence about that additional inference—and about whose experiences it misses.
Sources & further reading
- Large language models rate human happiness higher than humans do Le Fang and colleagues / npj Digital Public Health · 27 Oct 2026 · Accessed 28 Oct 2026
- Development and Validation of an Internationally Reliable Short-Form of the Positive and Negative Affect Schedule (PANAS) Edmund R. Thompson / Journal of Cross-Cultural Psychology · 1 Mar 2007 · Accessed 28 Oct 2026
- The International Positive and Negative Affect Short Form: scoring guide Edmund R. Thompson · Accessed 28 Oct 2026