In 2020, a team of researchers ran a simple experiment at unusual scale. They matched strangers across the United States, put each pair on a video call, and asked them to talk for at least 25 minutes about anything they liked. No script, no assigned topics.
The result, published in Science Advances in 2023, is the CANDOR corpus: one of the largest open collections of naturalistic conversation ever assembled.
By the numbers
- 1,656 conversations, recorded in spoken English
- About 850 hours of audio, video and transcripts, more than 1 terabyte of data
- 1,456 unique participants, aged 19 to 66, from across the U.S.
- Each participant paid up to $15 for the conversation and surveys
Every participant agreed to have their conversation's audio and video recorded and released for research. That consent is what made the corpus possible to share at all.
Why researchers value it
Because the conversations are unscripted and video-based, CANDOR captures what most text datasets lose: pauses, overlaps, laughter, facial expressions and the way two strangers gradually find a rhythm. Researchers have used it to study conversational dynamics, including how personality shapes the way people talk with each other.
The scale problem
For all its value, CANDOR is small by AI standards. 850 hours is a rounding error next to the text and video that large models are trained on. It is research-licensed, two-person and English-only. It does not include group meetings, presentations, physical tasks or the long tail of settings, such as negotiations and cross-cultural calls, where AI most often fails.
CANDOR proves that consented, high-quality interaction data can be collected. The challenge now is doing it at the scale, diversity and licensing terms that Social AI needs.
Why it matters for Social AI
The CANDOR design is a template: recruit real people, pay them fairly, get explicit consent for how the data will be used, and record every participant clearly. Crosstalk applies the same principles across many more settings and group sizes, with licenses for both research and commercial use.
Crosstalk sources consented interaction data across presentations, conversations, physical AI and social settings. Explore the catalog, request a custom collection, or read the paper.
Sources: CANDOR paper in Science Advances (PMC), Cornell ConvoKit, BetterUp CANDOR page.