NYC Council Committee of the Whole, October 5, 2026 › What each group talked about
What the AI companies talked about
Morgan Dwyer (62 turns), Logan Graham (55 turns), Shane Cahill (54 turns), Alice Friend (50 turns) · OpenAI, Anthropic, Meta, Google · 221 turns, about 165.9 minutes, 23,891 words
Each speaker in this group on their own: Morgan Dwyer (OpenAI), Logan Graham (Anthropic), Shane Cahill (Meta), Alice Friend (Google).
Phrases the AI companies used far more than everyone else
Bigger means more distinctive: used often here, rarely by everyone else. The 12 most distinctive are blue and largest. Hover or tab to a phrase for its counts. The full list is in the table below.
How many times the AI companies said it
The 20 most distinctive phrases, ordered by how many times the AI companies said each one in about 165.9 minutes of talk. The gray number in parentheses is how many times everyone else said it across the rest of the 10-hour hearing.
Show the data as a table
| Phrase | This group | Everyone else | Distinctiveness score |
|---|---|---|---|
| safety | 140 | 104 | 9.79 |
| framework | 51 | 9 | 7.51 |
| models | 85 | 76 | 6.84 |
| incidents | 49 | 25 | 6.66 |
| federal | 47 | 23 | 6.59 |
| training | 45 | 26 | 6.14 |
| content | 30 | 7 | 5.78 |
| committed | 34 | 3 | 5.74 |
| users | 32 | 12 | 5.74 |
| release | 28 | 6 | 5.58 |
| team | 39 | 26 | 5.40 |
| Meta | 31 | 15 | 5.37 |
| approach | 26 | 9 | 5.23 |
| teams | 30 | 2 | 5.18 |
| evaluators | 24 | 6 | 5.16 |
| safeguards | 30 | 17 | 5.04 |
| policies | 25 | 10 | 5.02 |
| OpenAI | 48 | 45 | 4.96 |
| personal | 22 | 6 | 4.92 |
| risks | 52 | 53 | 4.85 |
| security | 40 | 34 | 4.83 |
| supported | 23 | 9 | 4.83 |
| safely | 19 | 4 | 4.60 |
| safe | 33 | 26 | 4.58 |
| model | 50 | 56 | 4.39 |
| independent | 33 | 29 | 4.30 |
| commitment | 17 | 6 | 4.22 |
| transparent | 16 | 3 | 4.21 |
| information | 47 | 54 | 4.15 |
| improve | 19 | 10 | 4.10 |
| evaluations | 21 | 1 | 4.09 |
| frameworks | 14 | 3 | 3.95 |
| robust | 14 | 4 | 3.91 |
| user | 14 | 4 | 3.91 |
| third parties | 14 | 2 | 3.88 |
| privacy | 27 | 24 | 3.86 |
| processes | 14 | 5 | 3.82 |
| practices | 13 | 4 | 3.75 |
| results | 13 | 4 | 3.75 |
| Muse | 16 | 1 | 3.74 |
| superintelligence | 14 | 6 | 3.71 |
| support | 32 | 34 | 3.67 |
| evaluate | 15 | 1 | 3.66 |
| teens | 12 | 2 | 3.63 |
| external | 12 | 4 | 3.57 |
How this was made
- “Distinctive” means a phrase this group said much more often than all other speakers at the hearing, after allowing for how much each side talked. A phrase someone says often but everyone else also says often is not distinctive.
- Each phrase gets a score from 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). We keep phrases said at least 3 times with a score above 1.96, a common cutoff; with thousands of phrases tested, a few may pass by chance, so read the list as a guide, not proof. When a short phrase almost always appears inside a longer one, only the longer one is shown.
- Counts are how many times each 1- to 3-word phrase occurs in the clean verbatim transcript. Hyphenated words are counted as separate words (“self-improvement” is “self improvement”), and a few terms the transcript spells more than one way are counted as one (“super intelligent” as “superintelligent”, “kill switches” as “kill switch”, “open-weight” and “affected parties” as one term).
- Common words, speakers' names, “New York”, “city”, “AI” and hearing boilerplate are left out. Beyond a standard list of common words, filler (polite forms, connecting words, generic verbs) was removed by hand: an AI assistant, working at BetaNYC's direction, reviewed each group's and each person's list and proposed the cuts. The full list of removed words is in scripts/meta/wordclouds.json.
- Speaker groups come from the speaker labels, which were assigned by AI agents working from self-introductions, the chair calling on people, and voice clusters, at BetaNYC's direction. Each label's role sets its group. A mislabeled turn counts toward the wrong group.
- This is a machine transcript, not an official record. Check the video before quoting anyone; search the clean transcript to find where a phrase was said.
- Data: ai-companies.csv (every phrase that passed). Code:
scripts/build_wordclouds.py.