Filed at content/popular-explainer/2026-04-15/article.md. Here's a summary of what's in the piece:
"Whose Body Did the Algorithm Learn From?" (~1,080 words)
Angle: Cultural bias in motion AI datasets — how the foundational training sets for modern motion models (Human3.6M, AMASS, HumanAct12) were built from a narrow slice of humanity and what that means for somatic practices outside that slice.
Opening: A concrete scene of how those datasets were actually assembled — specific, citable, accessible to general readers without jargon.
Core argument in three moves:
- Every dataset is an act of selection; names the real datasets and what they selected for
- Capoeira's ginga as the concrete example — technically precise, historically robust, globally legible, not exoticizing
- Somatic practitioners as expert witnesses to these gaps, framing perceptual training as a form of institutional critique
Closes with: the pragmatic and ethical stakes — for clinicians, for dataset designers, and for anyone using motion AI to evaluate movement
Further Reading (APA 7th): Buolamwini & Gebru (Gender Shades), Guo et al. (HumanAct12), Ionescu et al. (Human3.6M), Mahmood et al. (AMASS), Noble (Algorithms of Oppression), Röhrig Assunção (Capoeira history), Sheets-Johnstone (Primacy of Movement)
The internal annotation explains the series-escalation logic and flags the participatory research design relevance of the "expert witness" framing.