Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·May 21·6 min read
AffectAI-Capture: A Reproducible Multimodal Protocol for Small-Group Meeting Research
TL;DR
Studying group emotions and interactions requires collecting many data types simultaneously, but existing methods lack standardization and reproducibility. AffectAI-Capture is a protocol that synchronizes eye tracking, wearable heart-rate sensors, audio recordings, video from multiple angles, and self-reported feelings during structured four-person meetings, organizing all data around a single timeline.
✦ Why It Matters
Researchers can now collect and share standardized group-interaction datasets, accelerating reproducible studies of emotion and collaboration in meetings.
Key Takeaways
- Use unified event timeline to synchronize eye tracking, audio, video, and physiological sensors across distributed data streams.
- Ground meeting tasks in established group-interaction paradigms to ensure research reproducibility and comparability.
- Collect close-talk microphone data per participant to isolate individual speech from room noise.
- Implement structured self-report alongside sensor data to capture subjective emotional and social states.
Full Summary
Understanding how people interact in groups requires measuring multiple signals—eye gaze, heart rate, facial expressions, speech—but collecting and aligning these streams reliably has been difficult. AffectAI-Capture addresses this by defining a complete protocol for four-person meeting-like interactions that integrates eye tracking (gaze direction), wearable physiology (heart rate, skin conductance), close-talk microphones (individual speech), room audio, multi-view video, event logs, and structured questionnaires.
Sessions follow fixed task blocks derived from established group-interaction research paradigms, ensuring consistency across studies. All data streams are synchronized to a single authoritative event timeline, eliminating timing misalignment.
This standardized approach enables reproducible collection and comparison of emotional and behavioral responses across different research teams and studies.