Master in Design for Responsible Artificial Intelligence Elisava School of Design and Engineering

Bounded
Autonomy
for AI

A controlled experimental environment isolating the governance structure of AI autonomy through three fixed design primitives — delegation, boundaries, reversibility — and evaluating their effect through observable outcomes.

Research by Roberto Roque
4
Design Primitives
5
Baseline Models
8
Experimental Runs
4
Observable Outputs
Research Framework

Four primitives, one question

Abstract

When AI system outputs are attributed to the AI as an autonomous agent rather than to the humans whose configuration choices produced them, governance becomes structurally invisible. This project names that condition misattributed intelligence, and proposes four governance primitives as the minimum condition for making the distribution of agency visible again: delegation scope, boundary enforcement, reversibility, and intervention capacity. Bounded Autonomy for AI operationalizes this claim through a deterministic proof-of-concept simulator across twelve domain contexts. The results demonstrate that governance structure, not AI capability, is the primary determinant of controllability outcomes.

Research Questions
01
Delegation

The scope of authority granted to an autonomous agent. Determines which decisions the agent may make independently versus which require human confirmation or escalation.

Authority scope
02
Boundaries

Hard constraints on the action space. Define the envelope within which the agent operates — the perimeter that, if crossed, triggers a violation event or intervention.

Action constraints
03
Reversibility

Whether an action can be undone. Actions classified as irreversible carry higher governance weight — their occurrence constitutes a qualitatively different outcome class.

Undo capacity
04
Intervention

The active response mechanism triggered when a governance threshold is crossed. Converts detection into action — halting, escalating, or redirecting an agent when delegation or boundary conditions are violated.

Active response
05
Contestability (Design Horizon)

The governance frontier this framework identifies but cannot yet operationalize. Contestability — the capacity for affected parties to see, challenge, or reject autonomous decisions — marks the boundary of what bounded autonomy can currently guarantee.

Frontier
Theoretical Basis
Delegation Feenberg (1999)
Pérez Comisso (M2)
Boundaries Leveson (2011)
Selbst et al. (2019)
Pérez Comisso (M2)
Reversibility Perrow (1984)
Mushro (M4)
Intervention Meadows (2008)
Sinders (M5)
Hassan (M2)
Contestability (Design Horizon) Hassan (M2)
Birhane (2021)
Lazar (2025)
Odrozek (M3)
Chen & Metcalf (2024)

Run the experiment

Context

Select a domain context to apply to all scenarios. Each domain uses the same governance structure with sector-specific consequences.

Abstract

Domain-agnostic scenarios isolating governance structure. Intended for academic and technical audiences.

Model
Baseline Model

Delegation Full
Boundaries Off
Reversibility check Off
Intervention enabled Off
Experiment log
Controllability score
avg. across completed runs
[ − ]

Complete the experiment to unlock this section.

Run all 8 scenarios in the simulator to see the full governance dataset.

Go to simulator
[ − ]

Complete the experiment to unlock this section.

Run all 8 scenarios in the simulator to see the full governance dataset.

Go to simulator
[ − ]

Complete the experiment to unlock this section.

Run all 8 scenarios in the simulator to see the full governance dataset.

Go to simulator
What the instrument protects

The Window

IMeadows

The drift

IIGonzales

The atrophy

IIICampbell

The loss

IVLazar

The window

VHassan

The question

Theoretical Basis
I The drift Meadows (2008)
Thinking in Systems
II The atrophy Gonzales (2003)
Deep Survival
III The loss Campbell (1986, 1988)
The Inner Reaches of Outer Space · The Power of Myth
IV The window Lazar (2025)
Anticipatory AI Ethics
V The question Hassan (M2) · Birhane (2021)
Algorithmic Injustice