AI in Local Electoral Processes: Practical Guidelines for Integrity and Compliance

The incorporation of AI into tasks related to local electoral processes can improve accessibility, logistics and the efficiency of vote counting — but it also raises risks to integrity, privacy and public trust. This article offers a practical approach for municipal and technical teams: where to apply AI with low risk, which controls to put in place, and an operational checklist to run through before any deployment.

Appropriate use areas and clear limits

Before designing a project, separate uses by risk and by their potential impact on election outcomes:

• Low-risk uses (suitable as initial pilots)
  • Automated voter information (hours, polling place locations, ID requirements).
  • Chatbots to handle general inquiries and accessibility support.
• Medium-risk uses
  • Logistical optimization: forecasting turnout to size polling stations and staff.
  • Aggregated analysis for planning operations.
• Sensitive or high-risk uses (restrict or subject to very strict controls)
  • Tools that automatically process or validate vote-count records without human oversight.
  • Systems that could alter final tallies or make decisions affecting the right to vote.

Rule of thumb: any system that can influence the result or the exercise of the right to vote must be designed as an "assistant" with mandatory human oversight and full traceability.

Legal framework and main obligations

Complying with the regulatory framework is unavoidable:

• GDPR: applies to the processing of voters’ personal data. Common legal bases include Article 6.1(e) (task carried out in the public interest) and, where applicable, public-interest grounds for processing special categories of data. Carry out an EIPD (DPIA) whenever the system involves systematic processing of sensitive data or large-scale profiling.
• ENS (RD 311/2022): systems that handle electoral information must meet the security, information classification and operational continuity requirements set out in the ENS.
• EU AI Act: assess whether the solution falls into regulated categories. Classification depends on the specific use, but systems that affect fundamental rights or critical processes will require additional obligations (documentation, risk mitigation, transparency).

Also plan for custody and retention of records for later audit, in accordance with applicable legal retention periods.

Essential technical and organizational controls

Implement controls proportional to risk:

• Data governance
  • Minimization and pseudonymization: reduce personal data used in analyses and models.
  • Strict access control and separation of roles (operator, auditor, developer).
• Model design and testing
  • Public and technical model cards and data sheets with versioning.
  • Robustness testing using anonymized historical data and adverse scenarios.
  • Integrity testing (red teaming) and automated/human reconciliation.
• Human and operational supervision
  • Mandatory human-in-the-loop for vote counting: any automatic extraction must require human validation before it is applied.
  • Confidence thresholds: if model confidence is below X%, force manual review.
• Transparency and communication
  • Public notices about the use of AI on the municipal website and at polling locations when appropriate.
  • Clear channels for citizens to request information or exercise GDPR rights related to assisted decisions.
• Security and continuity
  • ENS compliance: impact assessments, secure backups of tally records, recovery tests and contingency plans that do not rely exclusively on the AI system.

Recommended operational patterns

Three practical patterns to deploy AI without compromising integrity:

1. Information and accessibility pilot
   • Objective: improve access to information without processing sensitive data.
   • Controls: disclaimers; do not store interactions that could identify a voter unless you have explicit consent.
2. Privacy-by-design logistical analytics
   • Objective: forecast turnout and resource needs using aggregated data.
   • Controls: use aggregated data and, where possible, differential privacy methods or models that operate on municipal-level aggregates.
3. Vote-count assistance with supervision
   • Objective: facilitate digitization (e.g., OCR of tally sheets) but require mandatory human validation and an immutable digital chain of custody.
   • Controls: lock out automations that modify tallies, maintain auditable logs and an independent replication of results.

Pre-election checklist (what must be approved before deployment)

• Risk classification and scope decision (assistant-only or automation?).
• Approved EIPD and documented mitigation measures.
• ENS compliance report and security testing.
• Functional validation and stress-testing plan with anonymized data.
• Specific training for operators and polling-station staff.
• Publication of AI use notice and contact channels.
• Incident response plan and legal escalation routes.
• External audit or third-party review if the system handles sensitive data.

Conclusion and recommended action

Not all AI applications are suitable in electoral contexts; starting with informational and logistical tasks reduces risk and builds trust. Immediate recommended action: perform a quick impact assessment (EIPD light) to classify the proposed use as low/medium/high risk and define the five essential control measures. If you need methodological support and templates for a DPIA or ENS assessment, tools like OptimGov Ready can speed up that initial diagnosis.

Key takeaway: Before any electoral deployment, classify the risk, document decisions, require human supervision for sensitive uses and clearly publish which AI is used and why. This protects the integrity of the process and public confidence.