Society-in-the-Loop: AI and Societal Values

• Society-in-the-Loop (SITL) is a framework that integrates technical oversight with a negotiated social contract to ensure AI systems reflect collective societal values.
• It employs methods like crowdsourcing, computational social choice, and participatory governance to translate public values into concrete technical requirements.
• Practical applications span content moderation, autonomous transportation, and robotics, using simulation platforms and real-time auditing to verify compliance.

The Society-in-the-Loop (SITL) paradigm designates a regulatory, technical, and participatory framework for aligning AI and algorithmic systems with broader, pluralistic societal values. Unlike the traditional Human-in-the-Loop (HITL), in which individual or expert oversight governs automated systems, SITL explicitly programs, negotiates, and maintains the algorithmic social contract—embedding ongoing collective negotiation, feedback, and compliance monitoring at societal scale (Rahwan, 2017). SITL has become foundational in theoretical analysis, governance models, simulation platforms, and practical tool development across domains including content moderation, embodied robotics, transportation, and fairness in algorithmic decision-making.

1. Conceptual Foundations and Motivation

SITL is formally defined as:

$\text{SITL} = \text{HITL} + \text{Social Contract}$

where the “social contract” represents negotiated tradeoffs and values among all stakeholders affected by an algorithmic system, directly inspired by political philosophy (Rahwan, 2017). The motivation stems from the observation that many modern AI deployments (self-driving vehicles, news algorithms, risk prediction models) have societal-wide impacts and generate contested distributive tradeoffs (e.g., privacy vs. security, fairness vs. utility), which exceed the remit of individual expert oversight. In these domains, legitimate governance demands explicit mechanisms for articulating, aggregating, operationalizing, and auditing societal values—not merely compliance with technical specifications.

SITL thus requires:

• Programming societal values into system design requirements.
• Negotiating tradeoffs using computational social choice, crowdsourcing, and contractarian reasoning.
• Continuous monitoring and compliance verification in live deployments.

2. SITL in Technical Systems and Simulation Platforms

SITL has been operationalized in platforms such as Virtual Community for embodied human-robot society simulation (Zhou et al., 20 Aug 2025) and Sky-Drive for collaborative autonomous transportation (Huang et al., 25 Apr 2025). These environments instantiate SITL via:

• Multi-agent architectures supporting both human and robotic avatars, governed by unified physics engines and grounded in large-scale, real-world 3D scenes.
• Scenario generation pipelines encoding social networks, daily routines, and value-aligned behaviors, validated by automated and LLM-based “social feedback engines.”
• Task and reward structures incorporating explicit societal utility, formalized as cost, efficiency, and influence metrics in social planning, resource usage, and agent interaction.

For example, Virtual Community leverages LLMs (e.g., GPT-4o) for agent generation with rich, socially-grounded profiles and schedules; reward functions penalize inefficiency and misalignment with socially preferred norms (e.g., extraneous travel, disruption of schedule, violation of traffic rules). State automata and RL planners encode social logic, with outcomes quantitatively benchmarked for societal effect (e.g., “friendship ranking wins,” “conversion rates”) (Zhou et al., 20 Aug 2025).

Sky-Drive integrates multimodal, distributed human-in-the-loop feedback (eye tracking, physiological signals, voice) into the RL policy optimization for autonomous vehicles, aligning behaviors both with human preferences and physics-based societal traffic norms. Explicit reward structures and hybrid policies (e.g., HAIM/AIHM) enforce minimum safety and efficiency floors, with ongoing curriculum and scenario adjustment reflecting evolving social context (Huang et al., 25 Apr 2025).

3. Mechanisms for Articulating and Negotiating Societal Values

Operationalization of SITL demands translation of high-level, often contested societal values into concrete technical requirements:

• Crowdsourcing and value-sensitive design: Public surveys and sentiment analysis (MIT Moral Machine) elicit societal preferences for value tradeoff resolution.
• Computational social choice: Preference aggregation algorithms, voting schemes, and contractarian methods (e.g., Rawlsian veil-of-ignorance reasoning) formalize negotiation and encoding of distributive ethics in algorithms (Rahwan, 2017).
• Participatory governance: Civil Society in the Loop embeds CSO stakeholders directly into feedback collection, model adaptation, and tool refinement for AI-assisted moderation, with workflows supporting explicit labeling feedback, prompt revision, and batch retraining (Pustet et al., 9 Jul 2025).

In this context, societal values must be repeatedly negotiated and updated (debugging and maintaining the social contract) as technical systems and social realities co-evolve.

4. Monitoring, Compliance, and Auditing

SITL frameworks require systematic monitoring and auditing to ensure that algorithmic outcomes remain aligned with negotiated societal values in practice—not only as design intention:

• Algorithm auditors and black-box tests: Professionals or automated probes (e.g., input fuzzing, distribution shift tests) assess for emergent bias, unfairness, or policy non-compliance (Rahwan, 2017).
• Algorithmic oversight programs: Real-time algorithmic monitoring detects unwanted behaviors and adapts policies as needed (e.g., domain adaptation in fairness metrics).
• Continuous, live auditing: Especially for critical and infrastructurally embedded systems, ongoing outcome-based evaluation is necessary to catch “defeat device” strategies and emergent risks post-deployment.

In Civil Society in the Loop, model retraining is triggered by detected concept drift, using an accumulating gold-standard dataset built from stakeholder feedback; prompt-based LLM adaptation offers immediate, but monitored, deployment changes (Pustet et al., 9 Jul 2025). Virtual Community and Sky-Drive platforms incorporate replay, scenario editing, and auditor interfaces supporting direct societal input and scrutiny (Zhou et al., 20 Aug 2025, Huang et al., 25 Apr 2025).

5. SITL in Fairness, Feedback Loops, and Societal Risk

A critical application domain for SITL is algorithmic fairness, as the absence of SITL oversight enables self-reinforcing feedback loops that amplify societal inequality and risk (Bohdal et al., 2023):

$\text{risk} \sim \beta^t$

• $\beta$ = bias amplification rate ($\beta>1$), $t$ = iteration.
• Biased AI systems, once deployed, produce biased outcomes; these outcomes contaminate future training data, compounding disparities across generations.
• Feedback loops “lock in” marginalization and erode institutional trust; interactive effects with external stressors (e.g., climate change, economic shock) exacerbate risk up to existential levels (historical analogues: French Revolution, Arab Spring).

The SITL paradigm is identified as an essential safeguard, requiring:

• Regulatory mandates for ongoing human-in-the-loop review.
• Policy adaptation with continuous algorithmic audit and transparency mandates.
• Post-deployment evaluation to identify and remediate emergent or cumulative bias (Bohdal et al., 2023).

6. Limitations and Challenges in Practical SITL Deployment

Current technical and institutional approaches to fairness, social alignment, and participatory governance face notable limitations:

• Domain generalization: Most fairness methods only guarantee in-domain performance; real-world distribution shifts degrade fairness and utility functions (Bohdal et al., 2023).
• Data bias and infrastructure: Deeply embedded systems are difficult to update or remove, and unbiased data curation is costly and challenging.
• Feedback bias and resource constraints: In participatory workflows (e.g., Civil Society in the Loop), explicit feedback is skewed toward negatives, local retraining requires resources often unavailable in stakeholder organizations, and privacy requirements limit cloud-based operation (Pustet et al., 9 Jul 2025).

Paths forward, as proposed, include:

• Development of unbiased synthetic datasets with broad stakeholder input.
• Expansion of benchmarks to stress-test societal value alignment in rare or underserved populations.
• Investment in mechanistic interpretability (fine-grained analysis and correction at neuron/module levels).
• Continuous regulation and society-in-the-loop oversight at all lifecycle stages (Bohdal et al., 2023).

7. Summary Table: SITL Features Across Domains

Domain / Platform	SITL Mechanisms	Outcome / Principle
Virtual Community (Zhou et al., 20 Aug 2025)	LLM-based agent planning, scene-grounded profiles, social feedback loops	Embedding social context and societal norms in multi-agent interaction
Sky-Drive (Huang et al., 25 Apr 2025)	Distributed HITL, multimodal feedback, societal scenario editing	Societal value alignment in autonomous transportation
Civil Society in the Loop (Pustet et al., 9 Jul 2025)	CSO feedback, batch retraining, prompt-based adaptation	Participatory content moderation, transparency, contextual responsiveness
AI Fairness (Bohdal et al., 2023)	Societal feedback, algorithmic audit, regulatory mandates	Mitigation of feedback-induced social risk and bias amplification

8. Implications and Directions

The SITL paradigm integrates the democratic social contract tradition with algorithmic governance, calibrating technical design and oversight to evolving pluralist societal values. It provokes methodological innovation in participatory design, technical adaptation, continuous audit, and ethical operationalization. The approach is not universally solved; technical and institutional challenges to value translation, stakeholder negotiation, outcome monitoring, and robust adaptation persist across domains. Nevertheless, the embedding of society in the loop—rather than the individual human—marks a conceptual and practical shift in the governance of algorithmic systems affecting collective social, economic, and ethical life.