Citizen participation with AI: preventing manipulation and ensuring inclusive deliberation

Artificial intelligence can enhance citizen participation by synthesizing opinions, facilitating debate and extending reach. But it also introduces new avenues for manipulation: bots, coordinated campaigns, amplification of disinformation and algorithmic biases that distort deliberation. This article offers practical measures —technical, organizational and contractual— to deploy AI-powered participation platforms that protect the integrity of the process and ensure inclusion and transparency.

Concrete risks you must mitigate

• Bots and fake accounts that inflate a viewpoint or flood forums.
• Algorithmic coordination that amplifies specific narratives (astroturfing).
• Filter bubbles: recommendations that only surface certain voices.
• Disinformation amplified by automated summaries or generated content.
• Profiling and improper handling of personal data (GDPR risks).
• Digital exclusion due to inaccessible interfaces or technical language.
• Lack of traceability and contestability when public decisions rely on automated outputs.

Relevant regulatory framework (brief)

• GDPR: requires a legal basis for processing, data minimization, rights of access and portability, and Data Protection Impact Assessments (DPIAs) when applicable.
• ENS (RD 311/2022): systems handling public information must meet security requirements affecting hosting, access control and incident management.
• Law 9/2017 on Public Sector Contracts: platforms and algorithms procured by public entities must include clauses for control, auditing and continuity.
• EU AI Act: introduces a risk-based approach; transparency obligations and technical documentation may apply depending on the system category.

This is not just about compliance: it's about designing for trust and legitimacy.

Operational design: practical measures (what to do now)

1. Identity verification and bot mitigation
   • Implement gradual onboarding (minimal verification), limits on account creation rates and stepped checks for sensitive actions.
   • Use anomaly detection and network analysis to spot coordination patterns.
2. Traceability and labeling
   • Clearly label content generated or assisted by AI (provenance). Every summary or recommendation should include links to the original sources.
   • Keep logs of algorithmic decisions: inputs, model versions, parameters and timestamps for auditing.
3. Algorithmic transparency and alternatives
   • Offer simple explanations of why a particular ordering or synthesis is shown (for example, ranking criteria).
   • Provide user controls: sort by chronology, show "alternative views" or compare multiple summaries.
4. Human+AI moderation
   • Combine automated filters with human review for potentially manipulative content.
   • Establish escalation thresholds (e.g., when unusual activity affects a key debate).
5. Prevention and detection of manipulation
   • Monitor signals of astroturfing: bursts of activity, new accounts with similar behavior, implausible geographic clusters.
   • Define response playbooks (temporary suspension, freezing votes, public communication).
6. Privacy and data minimization
   • Collect only what is necessary; apply pseudonymization and defined retention policies.
   • Conduct a DPIA before deployment and consult with the Data Protection Officer (DPO).
7. Accessibility and inclusion
   • Provide multilingual interfaces, plain-language summaries and alternative channels (in-person/phone) for those without digital access.
8. Contestability and governance
   • Publish minimum technical documentation (model cards) and a complaint/appeal procedure for AI-assisted decisions.
   • Audit results periodically and publish public reports on detected biases.

Procurement and essential contract clauses

• Require in tender documents (Law 9/2017) technical audit rights and access to relevant logs.
• Define SLAs for data integrity, availability and incident response times.
• Continuity clauses: data export plans and migration options if the provider stops servicing.
• Security requirements according to ENS (RD 311/2022) for hosting in certified environments.

Operation and practical metrics

• KPIs to measure ecosystem health: degree of representativeness (by demographics), ratio of content flagged as suspicious, average incident resolution time, percentage of users with access to plain-language summaries.
• Three-phase pilots: (1) closed beta with representative groups, (2) expanded pilot with independent audit, (3) deployment with continuous monitoring.
• Quarterly reviews of models and policies, with citizen participation and external evaluation when appropriate.

Integration with municipal practice

Administrations can start with controlled modules: for example, AI to synthesize input in public consultations, but with operational limits (do not use automatic summaries for decisions without human review). Procurement conditions and security and privacy requirements should be defined from the technical specification phase.

OptimGov can be integrated into these processes as a modular component that preserves traceability and facilitates technical and contractual compliance, but design and governance decisions should be led by public entities.

Takeaway / Recommended action

Before launching any AI participation tool, carry out a DPIA and a short pilot (30–60 days) with verification controls, labeling of AI-assisted content and a manipulation response protocol. Document these requirements in the procurement specifications (Law 9/2017) and include ENS (RD 311/2022) security obligations and audit rights in the contract.

Immediate action: convene a multidisciplinary workshop (IT, legal, citizen participation and DPO) to define the five minimum rules for your platform: verification, traceability, transparency, human moderation and alternative access. This ensures AI enhances deliberation without undermining the legitimacy of the process.