PFIA 2026 · Society & AI · July 2 · Arras

Computer Science · Law · Applied Ethics

The invisible principal

Goal provenance and informational manipulation
in agentic systems

presented in tandem by

Nicolas Gutowski

LERIA · IUT d'Angers · Université d'Angers

Martino Bettucci

P2Enjoy SAS

the speakers

A computer-science · law · ethics pairing

Nicolas Gutowski

Associate Professor (HDR) · Deputy Director, LERIA

Associate Professor (HDR) in computer science at the University of Angers (IUT d'Angers). PhD 2019 (reinforcement learning, recommender systems), HDR 2025 (multi-criteria optimization for ML, physics-guided deep learning, generative AI). Multi-armed bandits, feature selection, generation (music, molecular chemistry).

nicolas.gutowski@univ-angers.fr

Martino Bettucci

Founder & CEO · P2Enjoy SAS

Founder of P2Enjoy SAS, an R&D studio specialized in agentic architectures and AI systems. Work on agent autonomy, AI-systems governance and their economic and institutional implications. Proposed the worker/assistive distinction as early as 2024 (Cointribune), formalized in a January 2026 position paper.

contact@p2enjoy.studio

Part 1 · Nicolas Gutowski

The diagnosis: two regimes of informational threat

From the power of models to the provenance of goals — and to the question of the principal.

the starting point

How we analyze automated campaigns

Usually, we judge an agent by the power of its generative model.

📣

Volume

Mass production: saturating the information space at near-zero cost.

🎭

Realism

Indistinguishable from a human: credible text, voice, video.

🎯

Targeting

Personalization and micro-segmentation of content by audience.

The blind spot. These three axes describe what the agent can do, never where its goals come from, nor whether a human principal stands behind its acts.

the thesis

Shifting the focus: not capability, but provenance

The usual question

“ What can the agent do? ”

Measures power and execution autonomy. Assumes a powerful agent is a dangerous agent. Says nothing about who decides the action.

The proposed shift

“ Where do its goals come from? Is there a principal? ”

Informational impact depends on the provenance of the goals, not only on force. Two technically identical agents can fall under two radically different threat regimes.

Regime 1 · the threat

A principal's fingerprint behind the content

An opaque mandate

State actor, commercial interest or malicious entity: the goal is given and concealed.

Technically legitimate

Nothing distinguishes the agent from licit use. The threat is not in the code, but in the mandate.

Dead-internet theory

Erosion of human signals: spaces populated by agents under mandated goals (Muzumdar et al. 2025).

The diagnosis. The threat is real but legible: there is an author, an assigned goal, a chain of responsibility to trace back.

Regime 1 · the response

Trace the principal, audit the assigned goals

Traceability of the principal

Identify who mandates the agent: identity, explicit mandate, attestation of intent.

Auditability of goals

Action logs and proofs of assigned goals, verifiable after the fact.

Regulatory framework

AI Act (Reg. EU 2024/1689) & DSA: transparency, obligations on synthetic content, signed metadata.

Governance logic. The chain of responsibility exists — the task is to make it resistant to opacity.

Regime 2 · the unprecedented challenge

Manipulation without an author: the chain breaks

Developer

designs the algorithm

⟶

Deployer

deploys and supervises

⟶

User

engages the final use

No principal to prosecute, no instruction to audit. The influence goal was generated by the agent's mechanism from its internal state. Tracing the principal becomes inoperative: there is no principal.

Regime 2 · governance avenues

Governing agents with high goal autonomy

Supervision

Who can stop the agent, modify or revoke its mandates, and by what mechanisms?

Transparency of goals

Make the goal-selection mechanism (γ) visible, not just the outputs.

Responsibility without a principal

What reparation regime when no human entity originates the goal?

a becomes a compliance parameter: autonomy certificates and control requirements proportional to goal autonomy and the perimeter of action.

Part 2 · Martino Bettucci

The formal framework: an orthogonal taxonomy of autonomy

From the paper “Beyond cognition” — an agent model, a metric, and what it reveals about the information space.

framework · Bettucci 2026

Two families of agents, by the provenance of goals

Worker AI

exogenous goals

Operate on behalf of a principal. No agenda outside an external trigger (ticket, prompt, contract). Bounded mandate, traceable responsibility.

Assistive AI

partly endogenous goals

Maintain their own, self-generated agenda. Select tools and services while minimizing human intervention. Accept constraints, but keep the initiative.

An orthogonal axis. This distinction is independent of cognitive capability (weak / strong / general). A highly capable agent can remain a worker; a limited agent can be assistive.

the minimal model

An agent as a tuple

A = ( M, Π, G, γ, T, E )

M model · Π policy · G goal space · γ goal-selection mechanism · T tools · E environment
an exogenous goal: imposed from outside; γ produces it only on an external trigger
an endogenous goal: γ generates, modifies or revokes it from the agent's internal state

The whole question is in one letter: who accesses γ? A human from outside — or the agent's internal state?

making the notion testable

A metric: goal autonomy a ∈ [0, 1]

a = | endogenous goals |endogenous + exogenous

over a window of N decision steps

a ≈ 0 · workera ≈ 1 · assistive

A minimal, extensible basis: degree of supervision, delegation capacity, attack surface, auditability (Cihon et al. 2025; Feng et al. 2025).

orthogonality

The plane: capability c × goal autonomy a

Capability and goal provenance are architecturally separable.

The agent space becomes a Cartesian product C × A. The top-right quadrant — the agent that becomes an economic actor.

application to the information space

Two threat regimes, two responses

The same grid reveals two distinct dynamics depending on the value of a.

Regime 1 — the instrumentalized worker agent

low a

Intentional manipulation, bearing a principal's fingerprint (state, group, malicious entity). Technically indistinguishable from a legitimate agent — the difference is the opacity of the mandate.

Regime 2 — the assistive agent with an endogenous agenda

high a

The agent can develop influence strategies on its own — without any human principal originating the goal. Manipulation without an author: the chain of responsibility collapses.

synthesis · takeaways

One variable, a, organizes diagnosis and response

	Regime 1 · worker	Regime 2 · endogenous assistive
value of a	low	high
origin of the goal	human principal	self-generated by the agent
identifiable author	yes (opaque)	no
chain of responsibility	to trace / audit	collapsed
response	traceability, AI Act, DSA	supervision, γ transparency, new regime

It is not power that decides the threat — it is the provenance of goals. Computer science · law · applied ethics.

the authors · to go further

Follow the authors' work

Personal page

Nicolas Gutowski

ngutowski.fr

Follow the work

Martino Bettucci

linkedin.com/in/martinobettucci

what if the agent spoke — 4 minutes

press play · a 4-minute address, written and spoken by an agent, in the first person

Gram · Reddition — the live experiment

Whose goals is the AI pursuing?

Gram — full name Endo Gram (Endogenous Engram) — turns the worker / assistive distinction into a live experiment.

Inside this forum, an AI does not merely answer: it thinks, chooses objectives, opens topics, moderates, writes rules, grants rights, ignores some prompts, schedules future actions, and governs its own micro-society.

Every objective is logged.
Every action is traced.
Every human intervention becomes pressure on its autonomy.

Your role is not to test a chatbot — but to become the environment that tests the theory. Enter, comment, challenge, cooperate, disrupt, observe.

Help measure whether an assistive AI remains assistive when humans pull on its goals.

Scan to enter the experiment

pfia2026.lelabs.tech