Co-constructing sociotechnical AI governance: participatory system mapping using algorithm registers

1. Introduction

Municipal policymakers are increasingly required to publish algorithm registers that disclose which automated decision‑making systems are used, what data they consume, and where they are deployed. While registers improve visibility, they often remain static snapshots that fail to capture the dynamic interplay between algorithms, human operators, and institutional processes. This article explains how a participatory system‑mapping methodology—augmented by STPA—can convert a register into an actionable safety platform, using a Dutch municipal welfare‑benefits tool as a concrete case study. The insights are directly applicable to AI governance researchers, public‑sector technology officers, and any municipality seeking transparent, trustworthy AI.

The approach aligns with the AI governance principles championed by UBOS, leveraging its Enterprise AI platform to host collaborative workshops, automate documentation, and maintain living registers.

2. Background on Algorithm Registers

Since the EU’s AI Act and many national AI strategies, algorithm registers have become a policy staple. They typically contain:

• Algorithm name and version
• Purpose and scope
• Data sources and categories
• Vendor or developer information
• Deployment environment (cloud, on‑premise, edge)

However, three intertwined challenges limit their effectiveness:

1. Diverse stakeholder expectations: Citizens demand understandable impact narratives, auditors need technical metadata, and frontline workers require operational guidance.
2. Information overload: Registers often list technical fields without contextualizing how the algorithm interacts with legacy systems, legal frameworks, or human decision points.
3. Opaque sociotechnical coupling: The register captures a snapshot of the algorithm but not the surrounding processes, incentives, and feedback loops that can generate safety hazards.

To address these gaps, municipalities must adopt a participatory lens that brings together technical, legal, and lived‑experience perspectives. UBOS’s Workflow automation studio provides a low‑code environment for such collaborative mapping.

3. Participatory System Mapping Methodology

The methodology consists of two tightly coupled layers:

3.1. Co‑Creation Workshops

Stakeholders—including municipal staff, civil‑society NGOs, data‑protection officers, and end‑users—collaborate in facilitated sessions to construct a visual map of the algorithm’s sociotechnical ecosystem. The map captures:

• Data inflows (e.g., citizen‑submitted forms, tax records)
• Decision nodes (automated scoring, rule‑engine evaluation)
• Human control actions (caseworker overrides, manual audits)
• Feedback loops (appeal outcomes, performance dashboards)
• External dependencies (third‑party APIs, cloud services)

UBOS’s Web app editor can be used to embed these maps directly into the register, turning a static table into an interactive diagram.

3.2. STPA‑Guided Safety Analysis

Once the map is complete, analysts translate it into STPA artifacts:

1. Identify unsafe control actions (UCAs): Actions that could lead to loss of safety, such as “auto‑deny benefit without human review.”
2. Define loss scenarios: Situations where a UCA could cause harm (e.g., wrongful denial of benefits).
3. Trace causal factors: Technical bugs, ambiguous policy rules, or organizational incentives that could trigger the UCA.

The outcome is a set of concrete, actionable recommendations that can be fed back into the register, creating a living safety loop.

The entire workflow can be orchestrated within UBOS’s templates for quick start, allowing municipalities to launch new safety analyses in days rather than weeks.

4. Case Study: Dutch Municipal Welfare Benefits Tool

The research team selected a welfare‑benefits eligibility engine used by a mid‑size Dutch municipality. The tool ingests citizen‑submitted applications, cross‑references tax records, and outputs a binary eligibility decision.

4.1. Register Extraction

The starting point was the municipality’s public algorithm register entry, which listed:

• Name: BenefitEligibilityAI
• Purpose: Determine eligibility for social assistance
• Data sources: Citizen form data, national tax database, employment registry
• Vendor: Local AI consultancy
• Deployment: Cloud‑based microservice (Docker, Kubernetes)

4.2. Stakeholder Recruitment & Briefing

Eight participants were invited:

• Three caseworkers from the welfare department
• One representative from a citizen‑rights NGO
• One municipal ombudsman
• Two IT managers (infrastructure & data engineering)
• One legal advisor specializing in social‑security law

Each participant received a briefing that linked register fields to real‑world tasks (e.g., “rule engine” → “caseworker verification step”). The briefing material leveraged UBOS’s AI marketing agents to generate concise, role‑specific summaries.

4.3. Co‑Construction of System Maps

Over three half‑day workshops, participants used a shared digital canvas (hosted in the Workflow automation studio) to sketch causal loops. Key discoveries included:

• Multiple input streams (online forms, paper submissions, third‑party data feeds).
• Two distinct human control points: a manual “data‑validation” checkpoint and a “final approval” step that can override the algorithm.
• A feedback mechanism where denied applicants can appeal, triggering a re‑run of the algorithm with adjusted parameters.
• An undocumented “batch‑processing” job that runs nightly, creating a hidden performance bottleneck.

4.4. STPA Integration

The map was translated into STPA tables using UBOS’s AI Survey Generator to capture stakeholder perceptions of risk. The analysis identified three high‑impact unsafe control actions:

1. Automatic denial without contestability – the system can close a case without exposing the rule that triggered the denial.
2. Performance degradation under load – the nightly batch job slows down during peak application periods, causing delayed decisions.
3. Insufficient audit trail – the register lacks fields for logging which rule fired and why, hindering post‑hoc investigations.

5. System‑Theoretic Process Analysis (STPA) Findings

STPA’s strength lies in its ability to surface control‑theoretic hazards that traditional testing misses. The following table summarizes the key findings from the Dutch case study:

Each mitigation was fed back into the register as a new metadata entry, ensuring that future auditors see both the original algorithm description and the safety enhancements.

6. Safety Hazards and Recommendations

Beyond the three UCAs identified, the participatory mapping uncovered additional latent hazards:

• Data‑quality drift: Tax‑record updates lag by two weeks, causing outdated eligibility calculations.
• Human‑in‑the‑loop fatigue: Caseworkers receive >30 automated alerts per day, leading to alert desensitization.
• Vendor lock‑in: The cloud provider’s proprietary model format prevents easy migration.

Recommended actions:

1. Implement a data‑freshness flag in the register, automatically turning off the algorithm when source data exceeds a freshness threshold.
2. Introduce priority‑based alert routing using UBOS’s AI Email Marketing module to summarize high‑risk cases for senior supervisors.
3. Adopt an open‑format model container (e.g., ONNX) and record the format in the register to reduce vendor lock‑in risk.

To operationalize these recommendations, municipalities can leverage the Enterprise AI platform by UBOS, which provides built‑in versioning, monitoring, and compliance dashboards.

7. Implications for AI Governance

The participatory system‑mapping + STPA framework delivers several governance benefits:

• Transparency that scales: By embedding maps directly into the register, citizens can explore the sociotechnical context without needing specialist knowledge.
• Accountability loops: Each unsafe control action is linked to a concrete mitigation that is tracked as a register update, creating an audit trail of governance actions.
• Stakeholder empowerment: NGOs and ombudsmen become co‑designers rather than external critics, fostering trust and reducing adversarial dynamics.
• Regulatory alignment: The approach satisfies emerging EU requirements for “risk‑based” AI oversight, as outlined in the AI Act’s Article 9.

Moreover, the methodology can be generalized beyond welfare benefits to any public‑sector AI system—such as AI Video Generator for public communications or AI Chatbot template for citizen services.

For municipalities seeking to institutionalize this practice, the UBOS partner program offers training, certification, and co‑development opportunities.

8. Conclusion

Turning algorithm registers into living governance artifacts requires more than data entry; it demands a structured, participatory process that surfaces hidden sociotechnical risks. By coupling system‑mapping workshops with STPA, municipalities can:

• Identify unsafe control actions before they cause harm.
• Translate stakeholder insights into concrete register updates.
• Maintain a transparent, auditable trail that satisfies regulators and builds public trust.

The Dutch case study demonstrates that even a modest eight‑person workshop can uncover dozens of previously undocumented control points and three high‑impact safety hazards. Scaling the approach is feasible using UBOS’s low‑code Web app editor, Workflow automation studio, and pre‑built templates for quick start.

As AI continues to permeate public services, municipalities that embed participatory safety analysis into their algorithm registers will be better positioned to meet regulatory demands, avoid costly litigation, and, most importantly, protect the citizens they serve.

For a deeper dive into the original research, consult the arXiv pre‑print. To explore how UBOS can help you implement these practices, visit the UBOS homepage or read more about the About UBOS team.