False positive
Unusual behaviour is not necessarily fraud.
Tax reform · Plan de Rupture
6.09 — Fight tax fraud with artificial intelligence
The French tax administration already uses data-based targeting. The Plan proposes expanding it, while an algorithmic flag must remain an alert rather than automatic proof of fraud.
What the Plan proposes
Plan source: Recette nette — chiffrage détaillé à expertiser. The title and stated effect are preserved from the public edition; this page adds an initial legal and operational architecture for expert review.
The French tax administration already uses data-based targeting. The Plan proposes expanding it, while an algorithmic flag must remain an alert rather than automatic proof of fraud.
Status: structured civic prototype requiring expert review before filing or implementation.
Current system
Tax control involves more than 10,000 staff; the ministry reported €17.1 billion in tax rights and penalties notified in 2025. A June 2026 Act strengthened anti-fraud powers.
Implementation path
- Define purposesList targeted fraud and prohibit unrelated uses.
- Control dataQuality, source, retention, access and correction.
- Test biasMeasure false positives across taxpayer groups.
- Keep human investigationNo reassessment should rest on a score alone.
- Publish net resultsSeparate detected, notified, collected and control cost.
Key risks
Model secrecy
Taxpayers need enough information to challenge evidence.
Cybersecurity
Tax databases are sensitive targets.
Misleading figures
Notified amounts are not always collected.
Indicators to publish
| Indicator | Before | After | Purpose |
|---|---|---|---|
| Scope and population | Baseline | Annual value | Prevent perimeter changes |
| Budget effect | Estimate | Cash collected or spent | Separate promise from execution |
| Transition cost | Disclosed | Multi-year tracking | Calculate net effect |
| Appeals and errors | Initial rate | Corrected rate | Protect rights |