Ask New York City a question in plain language. Get an answer from the actual public record, with an identifier you can open in a browser.
Seven open-source data connectors · Free · MIT licensed · Four need no signup at all
Name the room before you show anything. Weight the hour toward whichever disciplines are actually represented; in a gen-ed group that is usually first-year writing.
Ask each person what they are designing right now. A faculty audience mid-curriculum-design needs things they can assign, not an awareness session. Their answers reorder the rest of this deck.
A model answers about NYC from what it absorbed in training. It cannot tell you where a number came from, because it does not know. Fluency is the only signal you get.
The model queries the City's own systems while you watch: Legistar, Checkbook, the City Record, the Charter, 311, Albany. Every claim resolves to an identifier.
The interesting part for a classroom is not the capability. It is that the failure mode becomes demonstrable. You can show students the wrong answer and the right one side by side, and ask them how they would have known.
Do not dwell. Thirty seconds.
The load-bearing sentence is the last one: the failure mode becomes demonstrable. That is the thing they can assign. The capability is not the point for this room.
Int ####-20##Int 1122-2024intro.nyc/1122-2024✓ resolvesThe teaching beat is the question you ask afterward: how would a student have known the first answer was wrong? The answer is not "use a better model." It is "demand a resolvable identifier."
This is the slide that cannot be replaced by a document you send afterward. Press R to replay the typing.
After the grounded panel lands, stop and ask the room: how would a student have known the first answer was wrong? Let them answer it. The answer is not 'use a better model,' it is 'demand a resolvable identifier.'
If the animation misbehaves, the content still reads. Do not narrate the tool.
BetaNYC's own definition: "a small piece of software that lets an AI assistant pull data from a source on demand." Seven of them, one per government data domain. Published on npm, MIT licensed, source on GitHub.
Step one returns metadata: a dataset ID, a column list, data types. That is an API contract. Step two queries against the ID step one produced. The two-step generalizes far past this portal, which is why it is worth teaching.
Works with Claude, ChatGPT, Copilot, Notion, and other assistants that speak the same protocol.
Skip most of this unless asked. Do not spend time defining technical terms for an audience that already knows them; they will ask if they do not.
The transferable idea is the two-step — find the contract, then query against it. It generalizes far past this portal, which is why it belongs in a syllabus.
The City Charter, the Administrative Code, and the Rules of the City of New York, as a searchable offline corpus. The city's actual governing text.
"What does the Charter say about community boards?"
@betanyc/nyc-charter-laws-rules
In a course: institutions from the primary text, not a textbook paraphrase
Run this query
Agency spending, contracts, budget lines, payroll, and revenue, from the Comptroller's Checkbook system.
"Show me Sanitation’s registered contracts for FY2026."
@betanyc/nyc-checkbook-mcp
In a course: any claim with a dollar figure becomes checkable
Run this query
Open RFPs, contract awards, and public hearing notices from The City Record, the city's daily paper of government activity, published since 1873.
"What solicitations are still open, soonest deadline first?"
@betanyc/nyc-record-mcp
In a course: how government asks for things, in its own register
Run this query
Which organizations each Council member funded, the award amounts, and capital project changes. Structured by BetaNYC from the adopted budget.
"Which organizations did Council Member Restler fund in FY2026?"
@betanyc/nyc-budget-mcp
In a course: representation made specific; every student has a district
Run this query
Lead with cost. Four of seven need no key and no account. That is the objection that kills classroom adoption, so answer it before it is asked.
Click any card to run the query live. Charter is the safest demo — offline corpus, no network dependency.
Every card links to its GitHub repo and npm package if someone wants to look under the hood.
Bills, committee votes, hearings, and individual council member voting records, live from the Council's own Legistar system.
"How did the Council vote on Int 0948?"
@betanyc/nyc-council-mcp
In a course: trace one bill; process as a verifiable sequence
Run this query
Alternate-side parking, collections, school session status, and emergency conditions such as Code Blue. Status lookups for any date, or a 90-day range.
"Is alternate-side parking in effect on July 30?"
@betanyc/nyc-311-mcp
In a course: the city as its own residents describe it
Run this query
State bills, votes, committee calendars, and Senate hearing transcripts from OpenLegislation. More than 150,000 records.
"Find State Senator Kavanagh. Which district, which session?"
@betanyc/nys-openlegislation-mcp
In a course: federalism made concrete; city authority versus state
Run this query
Any Socrata-powered portal, which means all of NYC Open Data and NY State Open Data. Dataset search, metadata inspection, live queries. Built by Nathan Storey.
"Search NYC Open Data for datasets about career and technical education."
npstorey/socrata-mcp-server
In a course: the catalog itself, read as an API contract
Run this query
Cautionary precedent: our own 2024 workshop on open data and AI required a paid ChatGPT Plus account to follow along. The 2026 replacement removed that requirement, because a paid-tool dependency excludes CUNY students unevenly.
The keys are free; registration is the only friction.
Say the last paragraph out loud. Our own 2024 workshop required a paid ChatGPT Plus account. The 2026 replacement removed it, because a paid-tool dependency excludes CUNY students unevenly. Admitting that buys more than any capability claim.
"Explore Open Data with AI. See how AI answers with and without access to real datasets."
Both answers side by side, in a browser. Three cities: New York, Chicago, San Francisco.
See what it returns
Each answer is published as a signed record naming the datasets, queries, and model used, with an Explore page that logs every tool call. Provenance as a feature, not a footnote.
It turns the demo two slides back into something a student runs themselves, on their own question, with no setup. The failure mode stops being a claim you make and becomes an experiment they perform.
Built by Nathan Storey as a personal project, fiscally sponsored by Metagov. Not a BetaNYC tool. He also taught the two-part open data and AI class at our School of Data in March 2026.
Demo-scale rate limits apply; running it locally lifts them. Worth naming the ownership out loud: the best classroom-ready thing in this deck is not ours.
Say the ownership out loud. Not ours. Nathan Storey's personal project, fiscally sponsored by Metagov, not affiliated with any employer.
The signed records feature is the thing to linger on: each answer publishes the datasets, queries, and model used, with replayable traces. A student can hand in the trace and an instructor can replay it. That is the closest existing analogue to a citation standard for an AI answer.
The most classroom-ready tool in this deck is not ours, and saying so is the credible move.
Two of these are worth pairing deliberately: the Field Guide indexes what exists, the Resist List argues about what should not.
Do not read the grid. Pick two.
The pairing worth making: the Civic Tech Field Guide indexes what exists, The AI Resist List argues about what should not. Showing both is the teaching move, and it signals you are not selling AI.
Every tool starts as a RADAR (Research & Data Assistance Request) from a community board, a coalition, or a partner organization. Not from us.
AI-generated imagery of projected flooding, for advocacy. Paperspace GPU, NYC Geocoder, LIDAR mask preprocessing.
Explore urban heat data, visualize heat metrics by neighborhood, and find cooling strategies.
Overlapping administrative boundaries. Why cross-agency coordination is structurally hard.
Public school allocations and budgets, scraped from the DOE’s Galaxy system, consolidated and cleaned.
Statewide support for the Housing Courts Must Change platform, and an address lookup telling tenants whether their building likely contains lead paint.
RADAR is the point of this slide. Tools begin as requests from community boards and coalitions, not as our ideas.
The lead paint map is the best single example for a humanities room: no AI at all, one public fact (buildings before 1960 are presumed to contain lead paint), and a real tenant decision at the other end. Built with the Cooper Square Committee through RADAR.
Reconciling public money has to be reproducible. Run it twice, get the same numbers, and be able to show why. A language model cannot promise that, so it is not in the extraction path.
The connector that answers "which nonprofits did my Council member fund?" sits on top of that deterministic data. The AI is the interface, not the arithmetic.
Slow down here. This is the credibility slide for a skeptical room.
Reconciling public money has to be reproducible: run it twice, get the same numbers, show why. A language model cannot promise that, so it is not in the extraction path.
The AI is the interface, not the arithmetic. Naming a place you declined to use AI buys more trust than any capability claim in the deck.
A student who understands why an agent must cite a dataset ID understands provenance, API contracts, and verification. Three things that outlive whichever model is current.
Where did this come from, and can I get back to it?
What does this source actually promise to contain?
Including the discipline of stating a bounded negative.
NYC's open data is an unusually good teaching corpus because it is real, local, free, and messy in instructive ways.
The pivot from tools to teaching. This is where the deck stops being a demo.
Provenance, API contracts, verification — three things that outlive whichever model is current. That is the argument for putting this in a gen-ed course rather than a CS elective.
First-year writing
Take one real agency document. Identify every checkable claim. Check each against a primary source, recording the identifier. Rewrite for a named audience.
Then: a note on what could not be verified, and why.
Critical thinking
The same civic question, answered twice. Both fluent. Students sort checkable from uncheckable, check, and then write up what made the wrong one persuasive.
Political science
One bill, one discretionary award, one 311 pattern, all in the student's own district. A memo on whether stated priorities match the record, citing identifiers throughout.
Run this query
On academic integrity: the deliverable is not prose a model could produce. It is a verification table, a rewrite, and a documented gap. The process is the artifact.
Weight first-year writing most heavily. Angela and Swan both teach it; that is half the room.
These assignments are untested. Offer them as conversation starters with four curriculum designers, not as finished handouts. Our own standing rule is that a script without verified notes is a draft.
The academic-integrity line is what they will ask about. The deliverable is a verification table, a rewrite, and a documented gap — the process is the artifact. A student who outsources it has to outsource the checking, and the checking leaves the trail.
Act 4 is where the honesty gets taught. Our own worked example: Local Law 174 of 2016 requires an annual career and technical education report. The most recent one on the portal covers 2019–2020. It is 2026. The finding is the absence.
Run this query
Act 4 is the one to keep in every version. That is where the honesty gets taught.
Click the card to watch the absence surface: Local Law 174 requires an annual report, the most recent on the portal is 2019–2020, and it is 2026.
Stating a bounded negative — precisely what you checked and did not find — is harder than finding a source, and it is exactly the move student writing usually skips.
Not necessarily. School of Data 2026 ran setup through GitHub Codespaces, which avoids local installation entirely. That is the answer for a locked-down lab.
Our Class 101 runs with no computer needed, using a printed card deck. It is in reworking right now, so we will not oversell it, but the format exists and works.
Yes. Most of it is CC BY-SA 4.0, crediting "BetaNYC, a partner project of the Fund for the City of New York." ShareAlike means your adaptation carries the same license. AI 101 material additionally carries a Data & Society credit.
The city masks small cell counts in its own published data so individuals cannot be re-identified. Same reasoning keeps student data out of these tools. It demonstrates better than it lectures.
These four answers decide whether anything actually gets assigned. Do not rush them to get to the closing slides.
Have the license obligation ready: CC BY-SA 4.0, crediting 'BetaNYC, a partner project of the Fund for the City of New York.' ShareAlike, so their adaptation carries the same license. AI 101 additionally carries a Data & Society credit — and the AI 101 page does not currently display it, so say it verbally.
This is an offer to collaborate, not a product to adopt.
Read this slide. Do not soften it.
Five of twenty classes written. The two most relevant to this room — plain language, and spotting AI fakes — are both still stubs.
Offer collaboration, not a product. This room can co-author the liberal-arts version, and that is a better ask than adoption.
Kathleen Cumiskey, Anthoni Garcia, Raj Korpan, Jeff Tobias, and Brendon Hawkins ran that workshop at our conference. No technical background required. Educators and librarians were the named audience.
That critique applies to our catalog too. AI 101 teaches use. AI Civics 201 teaches grounded querying. The questioning strand is the liberal-arts strand, and you are better placed to teach it than we are.
The strongest thing you have, and it is not a tool.
CUNY faculty ran a critical-AI workshop at BetaNYC's own School of Data in March 2026. PIT Lab and TIER Lab. No technical background required, educators and librarians as named audience.
Offer the introduction. Colleagues two campuses over beat a tool demo, and it is the concrete version of what the event description promised.
All seven, with setup instructions and example questions.
Intro to NY Open Data · AI 101 · Mapping for Equity, with slides and companion docs.
The annual conference. Where the Critical AI workshop happened, and where your session could.
BetaNYC is a partner project of the Fund for the City of New York. Materials CC BY-SA 4.0 unless noted. Connectors MIT.
Ask what would actually help them, then stop talking.
Do not promise the Discord cohort space. BetaBuilders Phase 1 built staff and associate-board areas only; a public learning space is an unstarted phase.
Every answer in this deck came out of a record someone else maintains, or a tool someone else built. Naming them is the same discipline the deck argues for.
The people who publish and maintain the data every connector reads.
NYC Council & Legistar · Office of the Comptroller (Checkbook NYC) · The City Record, published daily since 1873 · NYC Charter, Administrative Code and Rules, via American Legal Publishing · NYC 311 · NYC Department of Education (Galaxy) · NYS Education Department · NYC Council adopted-budget documents
The infrastructure that makes the records queryable at all.
Socrata / Tyler Data & Insights, and SoQL, the query language underneath both portals · NYC Office of Technology and Innovation, which runs NYC Open Data and the Open Data Ambassadors program · NYS Office of Information Technology Services, which runs data.ny.gov · the Model Context Protocol, which is how any of these reach an assistant
Tools in this deck that are not ours, and are better than ours in places.
Nathan Storey — the Socrata MCP server, civicaitools.org (fiscally sponsored by Metagov), and the two-part open data and AI class at our School of Data · Housing Court Answers (Roxanne) · CityMeetings.NYC · Reclamo AI · Depositron · Block Party · Vital City · The AI Resist List · Civic Tech Field Guide
Lab tools begin as a RADAR request from someone doing the work, not as our idea.
Right to Counsel NYC Coalition — the Housing Courts Must Change map · Cooper Square Committee and Lead Dust Free NYC — the lead paint map · OpenPlans, Culture and Arts Policy Institute, and New York Lawyers for the Public Interest — recent RADARs
Our own classes are co-authored, and the credit travels with the material.
Data & Society — AI Civics, a content partner on our AI 101 curriculum · NYC Council AICE Initiative — named curriculum partner · Upperline Code, NYC DOE, WiTNY and Google — CS and the City, and Data Science 1, presented at School of Data 2020 by Renne Castro and Jeff Olson · Two Sigma Data Clinic
The session that framed the gap this deck opens with, at our own conference, March 2026.
CUNY Public Interest Technology (PIT) Lab and the Trustworthy, Intelligent and Explainable Robotics (TIER) Lab — Kathleen Cumiskey, Anthoni Garcia, Raj Korpan, Jeff Tobias, and Brendon Hawkins
Skim this. The point is that it exists, not that you read it.
If someone asks why you bothered: the answer is slide 11. We argue that claims must resolve to a source, so the deck names its own.
Seven connectors: MIT. Curriculum, guides, toolkits and decks: CC BY-SA 4.0, crediting "BetaNYC, a partner project of the Fund for the City of New York." ShareAlike, so your adaptation carries the same license.
AI 101 material additionally carries a Data & Society credit and involves the NYC Council AICE Initiative. Government records referenced throughout are public records of the City and State of New York.
Since the deck shows you how to look up discretionary funding, here is ours.
$25,000, FY2026, from Council Member Lincoln Restler, Council District 33, through the Digital Inclusion and Literacy Initiative, administered by DYCD, passed through the Fund for the City of New York.
You can verify that yourself with the discretionary funding connector, in about fifteen seconds, without an account.
BetaNYC uses AI openly and with human accountability, so: this deck was assembled with AI assistance (Anthropic's Claude) under BetaNYC staff direction and review. The connectors themselves were built the same way, then reviewed and tested by staff before release.
The curriculum repos carry the same statement. We are not going to argue for provenance and then be vague about our own.
Every query result shown in the demos was dry-run against the live connectors on 2026-07-29. Nothing on those screens is invented, and nothing is a placeholder.
Two things we could not verify and did not fake: the Checkbook NYC website blocks automated requests, so that one link is unconfirmed; and figures on live systems change, so a number true today may not be true when you run it.
That is the honest version of a citation, and it is what we would ask a student to write.
Corrections welcome, on any of it. Issues and pull requests are open on every repo, and the portal directory in particular improves when someone tells us a link is dead.
The funding disclosure is the closing move if the room is still skeptical.
The deck teaches them to look up Council discretionary awards, so it discloses ours: $25,000, FY2026, Council Member Restler, District 33, Digital Inclusion and Literacy Initiative, via DYCD. They can verify it in fifteen seconds without an account.
Every demo result was dry-run against the live connectors on 2026-07-29. One link — Checkbook NYC — is unverified because the site blocks automated requests, and the slide says so.