Autonomy, Social Norms, and Alignment: Towards a Developmental Framework for Autonomous Artificial Agents
Marica Notte, Ludovica Marinucci, Vieri Giuliano Santucci
cs.AI
Quick verdict
Position paper: agents should earn autonomy as children earn norms; conceptually plausible but unoperationalized — no experiments, metrics, or implementation.
Summary
This position paper argues that aligning highly autonomous embodied AI agents should be treated as an epistemic, developmental process rather than a design-time guarantee. The authors observe that intrinsically motivated open-ended learning expands agent autonomy while making pre-programmed normative rules unworkable. They then review how children acquire autonomy and social norms: imitation of caregivers, gradual internalization (understanding and applying norms around age three, generating norms by five, per Tomasello), self-regulation as mediator, and participation in collective practices. From this they distill a four-part model: norm alignment requires learning that is (i) embodied and situated, (ii) intrinsically motivated, (iii) socially interactive, and (iv) staged developmentally. Following Dennett, they propose that machine autonomy be granted progressively, contingent on demonstrated norm competence — agents 'earn' their degrees of freedom — and that EU AI Act regulatory sandboxes could host such staged training. The paper explicitly denies that such agents acquire moral responsibility. Main limitations: the framework is entirely conceptual — no experiments, no operational metrics for 'demonstrated alignment,' no engagement with the multi-agent norm-learning literature, and the child-agent analogy is asserted more than argued. It reads as a research agenda rather than a validated method, with its value lying in framing and design guidance.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 16 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Pretraining and human feedback are insufficient for dynamic, unstructured environments, so highly autonomous agents must acquire norms through embodied interaction rather than pre-programmed rules (Section 1).
- Norm internalization requires learning that is simultaneously embodied/situated, intrinsically motivated, socially interactive, and structured into developmental stages (Section 3).
- Machine autonomy should be granted progressively, contingent on demonstrated normative competence, mirroring Dennett's gradual attribution of moral status (Section 3).
- Even systems with sophisticated norm-management capacities are not morally responsible, because responsibility requires free will and intentionality beyond foreseeable systems (Section 3).
- EU AI Act regulatory sandboxes provide a suitable institutional framework for staged normative training and progressive autonomy expansion (Section 3).
Strengths
Empirically anchored interdisciplinary bridge
Section 2 grounds the framework in concrete developmental findings — Tomasello (2019) on imitation as conformity, norm understanding/application around age three and norm generation by five (Tomasello 2025), Barkley (2012) on self-regulation as mediator, Andrews et al. (2024) on primate norms — giving the AI argument a specific empirical anchor rather than vague analogy.
Crisp, rationale-backed four-condition model
Section 3 states each necessary condition with an explicit justification: norms are context-constituted so learning must be embodied/situated; internalization requires sustained exploration so agents must be intrinsically motivated; norms are intersubjective so interaction is required; norms are hierarchical so training must be staged.
Actionable governance hook
Philosophical honesty about moral status
Weaknesses
Core gating mechanism never operationalized
Missing engagement with normative MAS / machine normativity literature
Child-agent disanalogies unexamined
No falsifiable commitment or minimal experiment
Thin single-source empirical claims
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a short conceptual/position paper in three numbered sections by authors at ISTC-CNR and CID Ethics (Rome), produced under the EU Horizon 2020 PILLAR-Robots project (GA 101070381). It addresses one question: how can increasingly autonomous embodied AI agents remain aligned with human values and social norms when pre-programmed rules and dataset-scale pretraining fail in open, unstructured environments? The central move is reframing alignment as an *epistemic, developmental* process. Section 1 motivates this via intrinsically motivated open-ended learning (Santucci et al. 2020; Colas et al. 2022; Romero et al. 2025): techniques that expand robots' learning autonomy also expand decisional autonomy, making ex-ante rule specification unworkable. Section 2 reviews child development: autonomy as a gradually emerging capacity (Gillon 1985), imitation as a conformity mechanism (Tomasello 2019), norm understanding/application around age three and norm generation by age five (Tomasello 2025), self-regulation as mediator between experience and social competence (Barkley 2012), and norm transmission in primates (Andrews et al. 2024). Section 3 states the model: norm alignment requires learning that is (i) embodied and situated, (ii) intrinsically motivated, (iii) socially interactive, (iv) staged. Following Dennett (1996), autonomy should be granted progressively, contingent on demonstrated norm management — agents 'earn their autonomy through performance, not receive it upfront.'
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the main idea of this paper?
It argues that aligning highly autonomous AI agents with human norms should resemble a child's gradual education: agents should learn norms through embodied, intrinsically motivated, socially interactive, staged experience, and should 'earn' increasing autonomy only as they demonstrate norm competence — not receive it upfront.
Does this paper include experiments or data?
No. It is a conceptual position paper with no experiments, datasets, code, or quantitative results. It synthesizes developmental psychology (how children acquire autonomy and social norms), philosophy of autonomy and moral status, and AI research on intrinsically motivated learning into a framework with design implications.
Do the authors think AI agents can be morally responsible?
No. They explicitly argue that even agents exhibiting sophisticated norm-management capacities are not morally responsible, because moral responsibility requires free will and intentionality that current and foreseeable systems lack — building on the debate opened by Floridi and Sanders (2004).
What role does the EU AI Act play in the proposal?
The authors suggest the AI Act's regulatory sandboxes — controlled environments for supervised testing — could serve as institutional venues for the staged normative training and progressive, performance-gated expansion of agent autonomy that their model proposes.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Extraction contains no figures or tables; whether the published PDF includes any cannot be confirmed from the provided text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.