Supply Side II: Public Institutions and Governance Capacity

Session 11 · Wed, Oct 28

POLISCI 158 · AI & Politics · Stanford · Fall 2026

Roadmap

  1. In this session
  2. Questions we’ll tackle
  3. Before class
  4. After class

In this session

Governments are now buyers and users of generative AI, not only its regulators. We ask who gains and who loses when agencies adopt it: tech-savvy agencies gain capacity, vendors gain revenue, street-level bureaucrats may lose discretion, and voters and benefit recipients may lose accountability, while automated decisions can scale a single failure across thousands of people. We then turn to the state as regulator and ask why a general-purpose, fast-moving technology that firms understand better than governments is so hard to govern, and which tools regulators have.

Questions we’ll tackle

  • Who gains and who loses from AI in government: agencies, vendors, bureaucrats, or voters and recipients?
  • How can automated decisions scale failures, and where in a public process should a human stay in the loop?
  • What makes AI-governance different from e-governance?
  • Why is AI so hard to regulate, and which regulatory tool (standards and guidance, audits and red-teaming, liability, licensing, disclosure) fits which problem?

Before class

Read

Explore

  • Global Government GenAI Observatory (G3O), G3O Project (2026). Search one country’s ministries and look at how many institutions are documented, possible, or not yet reviewed; ask what an observatory can and cannot see.
  • Timeline for the Implementation of the EU AI Act, European Commission, AI Act Service Desk (2026). When each set of rules applies, and which deadlines moved after the Digital Omnibus amendments; ask why regulators phase rules in.
  • Public Trust in Government: 1958-2025, Pew Research Center (2025). Trust in the federal government fell from 73% in 1958 to 17% in 2025; ask what that means for accepting automated public decisions.

After class

Key takeaways

  • AI in government redistributes: capacity toward tech-savvy agencies, revenue toward vendors, and discretion away from street-level bureaucrats, with accountability to voters and recipients at stake.
  • Automated decisions can scale failures. Michigan’s MiDAS system decided 40,195 fraud cases by computer alone, and 85 percent of those findings were reversed on review.
  • Governments adopt AI against a backdrop of low trust in government and a wide gap between experts and the public: 56% of AI experts but only 17% of U.S. adults expect AI to have a positive effect on the country over the next 20 years.
  • AI is hard to regulate because it is a general-purpose technology, it changes faster than rules (the Collingridge dilemma: control is easiest when we know least), firms know more than governments about capabilities, risks, and usage, and evaluations cannot prove a risk is absent.
  • Regulators have several tools (standards and guidance, audits and red-teaming, liability, licensing, disclosure), and each fits some harms better than others.

Go further

About this deck

Adapted from POLISCI 157 Sessions 4B and 5A (Jul 15 and 20, 2026).

Curated by Simone Paci, Stanford University. Drafted with AI assistance (Claude) from the instructor’s course materials and reviewed by the instructor before publication; every external link is checked before each update.