Council Members
Including Speaker Menin and Chair De La Rosa, who asked most of the questions
Most distinctive phrases:
- sexualized
- image
- last week
- spacexai
- idea
- asked
- Grok
- pass
NYC Council Committee on Technology, October 16, 2026 › What each group talked about
What each group talked about
Oversight - Continuing Discussion about the Risks Posed by Artificial Intelligence. For each group of speakers, the phrases they used far more often than everyone else at the hearing. Each page has a word cloud, a bar chart of counts, and the data as a table.
Speaker groups
SpaceXAI
One witness, Gabriela Castillo Madrid
Most distinctive phrases:
- release
- model
- law
- bad actors
- policies
- control
- reports
- obviously
The other witnesses
Devin Kim, formerly of xAI, his counsel, and the public witnesses
Most distinctive phrases:
- xai
- data
- employees
- tech
- difficult
- critical
- OpenAI
- whistleblower
Individual speakers
Individual pages for the SpaceXAI witness, the former xAI employee, and the two Council Members who ran the hearing.
Each person is compared with everyone else at the hearing, including the rest of their own group. Phrases said at least 2 times count, because some spoke only briefly.
Gabriela Castillo Madrid (SpaceXAI)
Most distinctive phrases:
- release
- model
- law
- bad actors
- policies
- control
- reports
- obviously
See the word cloud and counts for Gabriela Castillo Madrid (SpaceXAI)
Devin Kim (former xAI employee)
Most distinctive phrases:
- xai
- employees
- difficult
- critical
- speak
- training
- safety
- researchers
See the word cloud and counts for Devin Kim (former xAI employee)
Julie Menin (Speaker)
Most distinctive phrases:
- Grok
- image
- sexualized
- spacexai
- idea
- images
- last week
- seconds
Carmen De La Rosa (Council Member, chair of the Committee on Technology)
Most distinctive phrases:
- reported
- type
- challenged
- joined
- sexual
- imagery
- giving
- developers
How this was made
“Distinctive” means a phrase a group said much more often than all other speakers, after allowing for how much each side talked, scored with the weighted log-odds test in Monroe, Colaresi and Quinn (2008), “Fightin' Words: Lexical Feature Selection and Evaluation for Identifying the Content of Political Conflict,” Political Analysis 16(4). Counts come from the clean verbatim transcript. Groups come from AI-assigned speaker labels. 56 turns at this hearing hold two voices; their words count toward the person named on the turn, so a cloud can include some words said by another speaker. This is a machine transcript; check the video before quoting anyone. Each page explains the details. Code: scripts/build_wordclouds.py.
Common words, speakers' names, “New York”, “city”, “AI” and hearing boilerplate are left out, using the word lists made for the 2026-10-05 hearing, which include filler an AI assistant proposed cutting from that hearing's lists. This hearing's lists were not reviewed by hand. Word lists: scripts/meta/wordclouds.json; this hearing's groups: scripts/meta/2026-10-16-technology-ai-wordclouds.json.