Demand Side II: Firms, Platforms, and Lobbying

POLISCI 158 · AI & Politics · Session 4 · Mon, Oct 5, 2026

Simone Paci

Roadmap

  1. Firms: political power and product safety
    • Ask when market discipline needs public regulation
  2. Six channels of political influence
    • Identify a mechanism and evaluate the evidence in an AI case
  3. Access, labs and competing principals
    • Compare model access and competing principal–agent relationships

Part I: Firms, power and product safety

Why look at AI firms?

Political economy

AI becomes more important. Which firms gain resources and political power?

Product safety

Who bears the risks? Are market discipline and liability enough?

Firms have strong policy interests

A firm’s business model gives it reasons to care about policy.

Labor

Who can do the work, and on what terms?

Finance

Who pays for expansion, and who bears the risk?

Market rules

Who can enter, compete and reach users?

Who solves the collective-action problem?

  • Money to sustain an effort
  • Technical knowledge and specialist staff
  • Fewer actors to coordinate
  • Dispersed benefits can encourage free riding

Having resources does not prove that a firm gets its way.

Mancur Olson, The Logic of Collective Action, 1965. Shared interests do not automatically produce collective action.

What happened with the last Industrial Revolution?

Historical cartoon of four wealthy industrialists treating American industry as a game board

Frederick Burr Opper, Let Them Have It All, And Be Done With It!, Puck, 1882. Cornell University Library, P.J. Mode Collection.

Antitrust: laws and a court remedy

1890

Sherman Act

  • Prohibited restraints of trade and monopolization.
  • Created a federal basis for action against trusts.

1911

Standard Oil dissolution upheld

  • The Supreme Court affirmed the dissolution order.
  • Standard Oil had to separate its constituent companies.

1914

Clayton Act

  • Targeted specific practices that could reduce competition.
  • Addressed mergers and interlocking directorates.

National Archives: Sherman Act; Standard Oil, 221 U.S. 1, pp. 1, 80–82; FTC: the antitrust laws.

Will the market keep AI safe?

The market-discipline argument

  • Unsafe products lose customers.
  • Reputation and liability deter risk-taking.

The questions to test

  • Can customers recognize the risk?
  • Who bears harms outside the transaction?

Did social media protect young users?

A familiar product-safety problem

Facebook and Instagram: children’s safety, product design and what users were told.

An external response

States challenged platform designs; litigation led to judgments and a multistate settlement framework.

New Mexico v. Meta, trial-court judgment, filed August 6, 2026.

What changed when states intervened?

New Mexico judgment

$942m

Civil penalties + abatement funding

Multistate settlement framework

Up to $17bn

Payments over ten years

Product rules: teen time limits, overnight blocks and independent monitoring.

New Mexico judgment, August 6, 2026; California AG’s framework announcement, August 26; North Carolina reports its first payment, October 1.

Part II: Six channels of influence

1. Campaign finance and candidate selection

Resources → Campaign messages → Candidate incentives

  • Who receives support? Who faces opposition?
  • Which policy dispute becomes an election issue?
  • What would show that spending changed the result?

NY-12: a policy dispute enters a primary

  • Alex Bores: Assembly sponsor of the RAISE bill
  • Think Big PAC: committee behind the attack ad
  • The dispute: rules for frontier AI developers

RAISE was enacted in 2025 and amended in March 2026. The amended act sets January 1, 2027 as its effective date.

New York Assembly, June 13, 2025; New York Senate, S8828, actions and §3; Think Big PAC, FEC C00923417.

Two sides, two messages

Still from Think Big PAC's campaign ad about Alex Bores

Think Big PAC

Alex Bores is Wrong for AI & Wrong for Congress

Still from Alex Bores for Congress's response ad

Alex Bores for Congress

Stop Us

Who is the audience? What claim should we check?

Campaign videos, 2026; stills illustrate the speakers’ messages. Linked playback requires internet.

Do influence campaigns always win?

Micah C. Lasher

41,940

votes

Alex Bores

37,637

votes

Observed: Lasher won the primary.

Unanswered: what would have happened without the spending?

NYC Board of Elections, certified NY-12 Democratic primary return, pp. 10–12; certified July 14, 2026. Named-candidate lines shown.

2. Lobbying: information and persuasion

Information

Expertise, policy drafts and constituency knowledge

versus

Persuasion

Arguments about costs, benefits and priorities

AI lobbying grew — what are we counting?

Public Citizen Figure 1: annual unique AI-issue lobbyists rose from 1,332 in 2022 to 3,570 in 2025

Unique people

1,332 → 3,570

2022 → 2025

About +168%

Public Citizen, Generative Influence, Feb 24, 2026, Figure 1 and Table 1. Search includes AI, autonomous vehicles/systems and data centers.

3. The revolving door

Government staff ⇄ Corporate staff

  • Knowledge of how institutions work
  • Professional networks and access
  • Preferences carried between organizations

What would separate expertise from preferential access?

Who has access to the administration?

President Donald Trump at a White House dinner with technology executives on September 4, 2025

White House technology-leaders dinner, September 4, 2025. Official photograph by Andrea Hanks; displayed crop.

4. Private governance: can firms preempt regulation?

Voluntary rules → Less demand for public intervention?

  • Firms can respond before governments impose rules.
  • Visible self-regulation may reassure voters and officials.
  • Industry-wide participation makes the signal harder to dismiss.

Malhotra, Monin & Tomz, Does Private Regulation Preempt Public Regulation?, APSR 113(1), 2019.

Private governance in practice: Project Glasswing

Anthropic sets conditions for access

  • Selected partners received Claude Mythos Preview for defensive cybersecurity
  • Anthropic committed up to $100m in usage credits
  • Expanded access still required Anthropic’s security conditions

Who sets the conditions, and who checks compliance?

Anthropic, Project Glasswing, Apr 7, 2026 launch description. Company announcement; credits are not cash disbursed.

A White House accord: voluntary oversight

Six firms signed the September 2026 accord.

Inside the firm

Internal controls + an empowered oversight team

Outside management

Independent evaluator + independent board committee

Who monitors the commitments? What follows a breach?

White House Accord on Super Intelligence, September 29, 2026. Google, Anthropic, Meta, OpenAI, xAI and NVIDIA; voluntary commitments.

5. Narrative power

A company frames the policy problem

  • OpenAI calls for mandatory national safety requirements
  • Its proposal ties rules to model capabilities
  • The framing invites government action

Which interests become visible? Which fade into the background?

Chris Lehane, OpenAI, The AI policy window is open. We need to act., Sep 9, 2026. An attributable company position.

Two companies, two narratives

Poster from Claude’s There’s hope in hard questions film

Anthropic: “There’s hope in hard questions”

Watch the film

Poster from Jensen Huang’s interview on The Ezra Klein Show

Jensen Huang × Ezra Klein

Watch the interview

Anthropic campaign, July 2026; The Ezra Klein Show, September 23, 2026. Click a poster for online playback.

6. Workplace politics

Workers can contest their employer’s choices

  • DeepMind staff sought union recognition
  • CWU reported demands about moral refusal and ethics oversight
  • Management rejected voluntary recognition; talks were planned

Communication Workers Union, May 20, 2026. A dated union account; no completed recognition is claimed.

Who speaks out from inside the labs?

Public departures and warnings

  • William Saunders · OpenAI safety/testing concerns
  • Daniel Kokotajlo · OpenAI governance concerns
  • Jan Leike · OpenAI Superalignment resignation
  • Geoffrey Hinton · left Google to discuss AI risks

Selected collective statements

Saunders, Senate testimony; Leike’s public statement; University of Toronto on Hinton. Letters link to their original statements.

Part III: Access, labs and principals

Open weights, open source, hosted access

Open weights

Parameters available to download; inspect the license.

Open source AI

OSI: permissions plus data information, code and parameters.

Hosted access

Use through an interface; access governed by provider terms.

Open Source Initiative, Open Source AI Definition 1.0. OSI’s definition is not a statutory classification.

Downloadable does not mean unconditional

Llama 4: one license-specific example

  • Look at the licensee’s and affiliates’ products or services
  • More than 700m monthly active users in the month before release?
  • Section 2 requires an additional Meta license, granted at its discretion

Access to weights does not remove hardware, hosting or license dependencies.

Meta, Llama 4 Community License, §2. Condition assessed on that version’s release date; subject to the agreement’s other provisions.

Labs choose how models are governed

Training choices + Instructions + Deployment rules

  • Which requests should the model accept?
  • Which conflicts should trigger a refusal?
  • Who can inspect or change the rules?

Anthropic, election safeguards update, Apr 24, 2026: training, system prompts and enforcement. Company description of its safeguards.

A lab and its model: an agency problem?

Conceptual principal-agent diagram: the principal hires the agent, which performs; each has its own interests

Can the principal observe whether the agent follows its goals?

Several principals, conflicting demands

Teaching network: one overarching AI firm sets rules for three AI agents; each agent also receives instructions from a different user, creating multiple overlapping principal-agent relationships

Whose instructions govern when the firm and user disagree?

You write the lab’s instruction

Hypothetical: a lab asks you to write rules for a new model.

  1. Write two sentences about what it should do
  2. Compare with a partner
  3. Find a conflict between the lab, a user and a government

5 minutes · paper or quiet reflection · no account or paid tool needed

Put the mechanisms together

Return to one AI case.

  • Who controls a scarce resource?
  • Through which channel could it affect a decision?
  • Whose rules govern the model?
  • What evidence would change your interpretation?

See you next class