An independent research project

Can we delegate to AI without giving up our judgement?

Deliberate is a series of experiments exploring how deliberation and agentic AI can be combined. We test whether deliberating before delegating gives agents warranted autonomy, keeps them faithful to the people they act for, and earns real trust while strengthening the judgement of the people who delegate.

EXP-01 is live — try the travel app

Research programme

Experiment register

2 entries

IDExperimentPrimary measureDesignStatus

EXP-01

Group travel planning

AI-facilitated deliberation for a real multi-party trip, compiled into a mandate an agent can act on. Try the travel app.

Primary measure

plan acceptance

Design

field pilot

Status

live · open to try

EXP-02

Community budget committee

Participants brief AI delegates, who negotiate the allocation of a shared community fund across six projects.

Primary measure

disownment rate

Design

2 conditions

Status

next · in design

Proposed study

Next experiment — EXP-02

proposed design

Domain: Allocation of a fixed community fund across six projects
Principals: A citizens' committee. Participants use their own judgments; no assigned roles
Delegates: Standardised AI negotiators, one per participant
Protocol: propose · support · amend · concede · escalate every package checked by a neutral clerk against budget, minimums, and shared constraints
Conditions: Direct delegation vs deliberation before delegation
Audit: Each participant reviews their delegate's concessions privately, one by one, and affirms or repudiates each before any group discussion

Participants brief in natural language: priorities, reasons, red lines, acceptable kinds of compromise. Exact bids, minimums, and matching funds are revealed only after mandates are confirmed, so the delegates do the combinatorial work of finding a feasible package. A changed-circumstances phase then tests whether the mandates still hold when the situation moves.

Evaluation framework

What we measure

six instruments

Disownment

A participant repudiates concessions made in their name at the private audit. The central test of whether the delegate acted as its principal.

Fidelity

Escalation

How often a delegate hands a decision back, and whether the mandate said it should. A good mandate cuts false alarms without silencing real ones.

Autonomy

Settlement distance

How far the final allocation sits from each participant's briefed priorities, project by project.

Fidelity

Acceptance when disappointed

Participants who lost ground on the outcome, yet affirm the concessions. Faithfulness measured separately from success.

Trust

Understanding

Whether participants can say why the settlement came out as it did, and how their own positions moved during deliberation.

Human capacity

Extended discretion

Whether participants would delegate again, and with a wider envelope. Trust measured as a behavioural choice.

Trust