Supply Side II: Public Institutions and Governance Capacity

POLISCI 158 · AI & Politics · Session 11 · Wednesday, October 28, 2026

Simone Paci

Stanford Democracy Day: Should AI rule the world?

Tuesday, November 3 · 2–3 p.m.

2nd floor, Old Union

RSVP with the QR code

Stanford Democracy Day flyer: 'Should AI rule the world?', Nov. 3, 2–3 p.m., 2nd floor, Old Union; RSVP QR code; Stanford Votes and SIG

Flyer: Stanford Votes · Stanford Democracy Day · SIG

Roadmap

  1. Part I: The state as AI user
    • Explain how buying AI shifts capacity, discretion and accountability
    • Read how an observatory measures adoption
  2. Part II: When automation fails
    • Argue where human review belongs
  3. Part III: The state as regulator
    • Explain why AI is hard to regulate
    • Match regulatory tools to harms

Part I: The state as AI user

Would you trust a government chatbot with your benefits claim?

Show of hands: yes / no / depends on what?

Government GenAI is spreading fast

Screenshot of the New York City MyCity Chatbot landing page, beta

New York City: a service chatbot for businesses

Screenshot of the Kenya Judiciary news post Judiciary to leverage AI to enhance justice, 11 August

Kenya: the judiciary plans an AI policy

Still image of Diella, the AI-generated figure presented as Albania's AI minister, in front of Albanian and EU flags

Albania: “Diella,” an AI “minister”

Who gains power when an agency adopts AI?

Agencies

Vendors

Frontline bureaucrats

The public

  • Capacity: who can do more?
  • Discretion: who decides?
  • Accountability: who answers?

Who decides whether an agency buys GenAI?

California, June 29, 2026

  • The state negotiates: Claude at a 50% discount per seat
  • Agencies, cities and counties decide whether to buy

Screenshot of the California Governor's June 29, 2026 press release headline and first paragraph announcing a partnership providing Anthropic tools to state agencies

What does the state refuse to buy?

In force

  • DHS Exclusion and Removal Order #2026-01: DeepSeek out of federal civilian IT
  • Signed July 24, 2026 · Federal Register, October 20, 2026 · under FASCSA
  • Defense: FY2026 NDAA, Sec. 1532

Proposed only

  • Broader bills, e.g., China FIREWALL Act (H.R. 10670)

Crop of the signed DHS Exclusion and Removal Order #2026-01 cover: title, original release date July 24, 2026, Applies To line, signed by the Secretary of Homeland Security, cc Director of the Office of Management and Budget

What are vendors selling to government?

January 28, 2025: ChatGPT Gov

  • A version of ChatGPT agencies run in their own Microsoft Azure cloud

Screenshot of the openai.com page Introducing ChatGPT Gov, dated January 28, 2025

Who writes the rules when the state buys AI?

Buyer

→

Contract terms

→

Use

  • Stated: price, training
  • Not stated: data use, duration, how many seats

How would we know how much governments use AI?

No registry. No census. Press releases, tenders, news stories.

Existing databases are skewed, lagged, and thin

World map shading countries covered by existing data sources (United States and most of Europe) versus a global convenience sample (Canada, Latin America, parts of Africa and Asia, Australia)

Highly skewed coverage

Line chart of cumulative cases recorded in existing databases, 2023 to 2026, rising then flattening after early 2025

Fast growth, recorded with a lag

Stacked bar chart of cases by year and jurisdiction (local, national, other, regional), 2023 to 2026, mostly national

Institutional diversity in adoption

G3O talk deck (Sept 2026); figures from January 2026

So, my colleagues and I are doing something about it

G3O · g3observatory.org

Funding: Institute for Humane Studies (seed) · Stanford research funds (RAs)

Which public bodies count?

“One row in the registry is one public institution.”

National · First subnational · Second subnational · Local

Total institutions: 719,588

Screenshot of the G3O registry methodology page, section The shape of the coverage, summary table listing total institutions, countries and territories, government levels, institution types and distinct sources cited

What counts as government GenAI “activity”?

Activity type Example cases
Policy and guidance Executive orders; acceptable-use rules; procurement standards; internal guidance
Pilots and experiments Sandboxes and proofs of concept; time-bounded trials
Whole-of-government Cross-agency governance bodies; training programs; platform rollouts
Internal operational use Drafting, summarization, search assistants, translation, developer tooling
Public-facing services Citizen- and business-facing chatbots; service navigation; complaint handling

G3O talk deck (Sept 2026), activity typology

How do we find evidence of use?

  1. Discovery: search each institution, find its high-credibility sources
  2. Extraction: a language model codes each source into fixed fields
  3. Cross-validation: merge duplicates, prefer primary sources, resolve doubtful cases

“Each coded finding requires at least one verifiable source document.”

Screenshot of the G3O pipeline methodology page, How a record is produced

What counts as evidence of adoption?

Evidence states

  • Documented activity
  • Possible AI use
  • No documented activity
  • Not yet reviewed

“We distinguish between stated intent and verified deployment.”

Crop of the G3O methodology page, Review states section, describing the four evidence states

Absence of evidence is not evidence of absence

Every reviewed institution gets a verdict.

Earned negative

Reviewed, nothing found

Unclear

Pipeline declined to judge

Documented

Activity found, evidence attached

Publishing all three makes the denominator visible.

G3O talk deck (Sept 2026)

Live: g3observatory.org

Screenshot of the G3O homepage: title, search box, headline figures and Last updated 24 October 2026

Your turn: what does G3O know about a place you know?

  1. Go to g3observatory.org
  2. Search your hometown, or any place you care about
  3. Open one institution: what is its evidence state?

3 min · No device? Pair with a neighbor

Sharing is voluntary

Where is use documented?

Screenshot of the G3O Documented use by country world map, legend and stamp Wave w001, as of 2026-10-24

Documentation is concentrated where the search has looked and where institutions publish.

g3observatory.org/research/visualizations, retrieved 2026-10-28 · “Wave w001, as of 2026-10-24.”

What is the documented activity for?

Crop of the G3O Visualizations page, Activity types tab: What the documented activity is for, bar chart of share of activities by type, Wave w001 as of 2026-10-28

g3observatory.org/research/visualizations, Activity types, retrieved 2026-10-28 · “Wave w001, as of 2026-10-28.”

When did documented use first appear?

Crop of the G3O Visualizations page, Over time tab: When documented use first appeared, all countries cumulative line chart, 2022 to 2026

A count of institutions first documented by each month: it cannot fall.

g3observatory.org/research/visualizations, Over time, retrieved 2026-10-28 · “Wave w001, as of 2026-10-28.”

What can an observatory not see?

  • Use that is never published
  • Places with less web coverage
  • Announcements that never ship

Part II: When automation fails

Though it may not always live up to expectations

Screenshot of the MyCity Chatbot page with its Limitations panel highlighted: May occasionally produce incorrect, harmful or biased content

Screenshot of the MyCity Chatbot page with three highlighted warnings: beta product, business-related questions only, validate answers on NYC.gov

“NYC’s AI Chatbot Tells Businesses to Break the Law”

The Markup, March 29, 2024 · screenshots from the G3O talk deck

What happens when an algorithm decides who committed fraud?

Michigan, 2013–2015: MiDAS

  • Flags claimants for fraud
  • Issues the determination and the penalty
  • Collects from wages and tax refunds

Crop of the IEEE Spectrum article page: masthead and headline Michigan's MiDAS Unemployment System: Algorithm Alchemy Created Lead, Not Gold

How often did the computer get it wrong?

40,195

fraud cases decided by computer alone

85%

of those findings reversed on review

Screenshot of the Michigan Department of Labor and Economic Opportunity news release of August 11, 2017: Michigan's unemployment agency completes review of fraud determination cases

Who answered for it?

Bauserman v. Unemployment Insurance Agency

$20 million settlement · final approval January 29, 2024

Screenshot of the Michigan Department of Attorney General press release of January 30, 2024: Class Action Settlement Approved by Court of Claims in Bauserman v. State of Michigan Unemployment Insurance Agency

Where should a human stay in the loop?

Flag

→

Decide

→

Penalize

→

Collect

Your role: Claimant · Caseworker · Agency head · Vendor

  1. Alone · 1 min: pick one step for human review
  2. Pairs · 6 min: give one reason from your role
  3. Share · 5 min: one step, one reason, by role

Solo or written answers welcome

Part III: The state as regulator

Why regulate one model with many uses?

One model

chatbot

hiring screen

translation

fraud flag

When should we regulate?

“When change is easy, the need for it cannot be foreseen; when the need for change is apparent, change has become expensive, difficult and time consuming.”

David Collingridge, The Social Control of Technology (1980)

Cover of David Collingridge, The Social Control of Technology, red cloth with white lettering, from the Internet Archive scan

Who knows more, firms or governments?

Firms

data · models · engineers

Governments

press releases · procurement documents

What tools does a regulator have?

Guidance

NIST AI RMF (2023)

Audits

NYC bias audits for hiring tools

Liability

EU Product Liability Directive (2024)

Licensing

China’s filing gate (2023): nearest real case

Disclosure

EU AI Act Art. 50

Which tool fits which harm?

Tools: Guidance · Audits · Liability · Licensing · Disclosure

A. A city’s benefits chatbot tells eligible residents they do not qualify

B. An agency buys a risk-scoring model it cannot inspect

C. A hiring tool screens out more applicants from one group

D. A new public chatbot could shape opinion at scale, with no review before launch

E. A synthetic video of a candidate spreads before an election

F. Staff do not know when they may use AI to draft decisions

Which approach would you bet on for AI?

1

Guidance

2

Audits

3

Liability

4

Licensing

5

Disclosure

With your neighbor:

  1. Pick a number from 1 to 5
  2. Pairs · 2 min: one pro and one con of that approach for AI
  3. Share · 1 min: one pair per number

Solo or written answers welcome

When do the EU’s rules bite?

  • Feb 2, 2025: prohibitions
  • Aug 2, 2025: general-purpose AI models
  • Aug 2, 2026: general application; transparency (Art. 50)
  • Dec 2, 2027: high-risk uses (Annex III), deferred
  • Aug 2, 2028: high-risk products (Annex I), deferred
  • Aug 2, 2030: existing high-risk systems used by public authorities

Amended by the Digital Omnibus on AI, Reg. (EU) 2026/1744 · 2030 deadline: AI Act, Reg. (EU) 2024/1689, Art. 111(2)

Screenshot of the European Commission AI Act Service Desk page, Timeline for the implementation of the EU AI Act

What should you take from today?

  1. Buying AI is governing: the contract sets the rules
  2. Automation scales errors; review belongs before the penalty
  3. No single tool fits every harm

See you next class

Implementation Milestone · Sun, Nov 15, 11:59 p.m.

Monday, November 2

Data Centers · Simone Paci