Multimodal Thematic Analysis - Live | AI in Academia - Dataspheres AI

What we did We took one hearing and six panel recordings and treated them the way a research team would: transcribe, attribute speakers, code the claims, a...

What we did We took one hearing and six panel recordings and treated them the way a research team would: transcribe, attribute speakers, code the claims, and count what comes up. The corpus is a 1 hour 30 minute (trimmed of pre-gavel air) US Senate HELP subcommittee hearing on AI in K-12 education — public domain, pulled from the Senate’s own stream and diarized into ten speakers — plus five Stanford HAI summit panels and a UNESCO Digital Learning Week talk, linked from their sources and summarized with Gemini’s native video analysis. Each coded segment is a verbatim quote with its source, speaker, timestamp where we have one, a theme code, and a stance. The theme codes are the same eight this datasphere already uses for its document corpus — integrity, assessment, pedagogy, acceleration, labor, equity, governance, publishing — so the audio, video, and web material lands in the same analytical frame as the text. What the coding shows Governance dominates, and it skews cautious. The hearing’s sharpest moments are all evidence-and-safeguards moments: no high-quality causal studies on long-term effects, a breach that exposed 62 million students’ records, consumer chatbots arriving in classrooms outside FERPA’s reach. The pro-side arguments cluster in pedagogy and labor — AI literacy curricula, administrative burden lifted off teachers — and even those come wrapped in conditions: human in the loop, teacher training first, vision before vendor. The panels rhyme with the hearing. Stanford’s researchers push augmentation over replacement; UNESCO widens the equity lens to geography and language. Where the senators worry about surveillance and cognitive surrender, the panelists worry about cheating and the loss of human connection. Different rooms, same fault lines. The live data Both tables below are live datasets, not snapshots. As more sources get coded, the counts change here and in every presentation chart bound to them. Theme sentiment summary Coded segments Method notes Speaker attribution comes from Deepgram diarization checked against the hearing’s official captions; names in parentheses are our identifications from self-introductions in the transcript. Video segments carry no timestamps because they are coded from Gemini’s summary of the full recording rather than from a diarized transcript. Stances are coded per segment, not per speaker — the same witness argues pro on access and con on chatbots, which is exactly the texture a single sentiment score would erase.