Skip to content
Collective Decision Laboratory
Contact

Ensuring human-AI decision systems are safe and effective.

Artificial intelligence is being integrated into decision systems, including in high-stakes settings. AI systems may act indirectly, by shaping the beliefs and preferences of human decision-makers, or directly, as decision-makers themselves. The Collective Decision Laboratory studies how hybrid decision systems succeed and fail, builds and tests safeguards, and works with practitioners to make these systems safe and effective—focusing initially on national security.

How hidden dependence enters a collective decision Decision system

Five decision-makers appear independent, but four are shaped by the same, hidden source.

  • Human
  • AI system
  • Hidden common source
  • Collective decision

The lab’s work

  1. 1 · Research

    Formal and empirical work on how competence, independence, aggregation rules, and more affect the safety and effectiveness of hybrid human-AI decision systems.

  2. 2 · Safeguard development

    Design and adversarial testing of protections for hybrid decision systems, targeting hidden dependencies and manipulation while anticipating advances in AI capabilities.

  3. 3 · Implementation

    Collaboration with practitioners to identify vulnerabilities in existing and planned systems, adapt safeguards to operational constraints, and harden institutional decision-making.

Research

Safeguarding Human-AI Decision Systems Against Hidden Dependence and Covert Capture

November 2026

Abstract

AI is being integrated into high-stakes decision systems like policy committees and military command & control. This integration is both direct (AI agents within these systems) and indirect (AI shaping the beliefs and preferences of human decision-makers), and it promises to improve the speed and accuracy of decision-making. But it also poses a serious risk: AI can introduce hidden dependencies which undermine collective accuracy or make decision systems susceptible to covert capture. A jury theorem model illustrates how (i) AI integration can improve individual competence while reducing collective accuracy, and (ii) manipulating a shared source of dependence can steer decisions towards a malign actor's preferred outcome even as measured collective accuracy rises. In one example, a committee is vetting AI systems for safety, and a common source shapes both the votes cast by AI decision-makers and the advice that human decision-makers receive. When an outside actor steers that source to favor approval, the committee's probability of approving an unsafe system rises from 17% to 65% even as its overall accuracy rises from 83% to 93%.

Building on a taxonomy of dependence mechanisms, this paper proposes a research and implementation agenda for identifying sources of dependence in hybrid decision systems under limited observability; assessing their effects on collective accuracy and vulnerability to capture; developing safeguards; and working with practitioners to integrate them into real-world decision systems.

Read the preprint →

About

The lab is led by Thomas Mulligan, PhD, a philosopher and decision theorist who served in the US Navy and the CIA. His work focuses on collective decision-making and the governance of advanced AI.