evidoria

← Back to browse

Good practice

CNIL's AI Sandbox for Public Services — Piloting Guardrails for Government AI Before Deployment

France · Paris · See the France profile

France's data-protection authority ran a 2023-24 regulatory sandbox pairing four public bodies — including France Travail and RATP — with AI projects, then published recommendations in April 2025 on bias, oversight and human review before nationwide rollout.

CNIL's AI Sandbox for Public Services — Piloting Guardrails for Government AI Before Deployment

Details

Maturity
Established
Promoter
CNIL (Commission Nationale de l'Informatique et des Libertés)
Period
2023–2025
Keywords
data protection, AI oversight, public administration

Context

CNIL, France's data-protection authority, opened the third edition of its regulatory sandbox in 2023, this time dedicated specifically to AI projects in public services, aiming to give hands-on regulatory guidance before deployment rather than sanctioning after the fact.

Objectives

Give participating public bodies direct regulatory guidance on bias, oversight and human review for their AI projects before any nationwide rollout.

Activities

Selected cases included France Travail's 'Conseils Personnalisés' chatbot (Jan-Jun 2024), recommending training courses to jobseekers; RATP's 'PRIV-IA' project (Jan-Apr 2024), testing privacy-preserving lidar-type sensors for passenger counting; and a Nantes Métropole project on water-consumption awareness.

Results

On 11 April 2025, CNIL published a recommendations document flagging risks of automation bias and excessive or biased data use, and requiring documented human oversight before any scale-up.

Conclusions

Sources do not publish quantitative outcome metrics for the AI tools themselves - the sandbox format itself, already run three times since 2021, is presented as the main transferable asset, more so than deployment metrics.

Implementation
Indicative cost
Low (< €50k)
Time to results
Medium (1–3 years)
Staffing & skills
CNIL sandbox coaching team, Participating public-body project teams (France Travail, RATP, Nantes Métropole)

Conditions for success

  • Voluntary selection of a small cohort of named public bodies for close regulatory coaching
  • An independent regulator (CNIL) with a pre-existing legal mandate and three prior sandbox editions of institutional experience since 2021

Common failure modes

  • No quantitative outcome metrics are published for the AI tools themselves, so success against the recommendations cannot be independently measured

Data sources

Where this practice's information was retrieved from, and when.

Attachments

Similar practices you may find useful