On 12 May 2026, Senator Pierre-Jean Verzelen (Les Indépendants) filed proposition de loi n° 609, "visant à généraliser la vidéoprotection algorithmique." It would move France's AI-assisted street cameras from one-off experiment to permanent law — letting software flag predefined events (a weapon drawn, a fire, a collapsed pedestrian, an abandoned bag) so human operators can respond faster. The bill is significant less for what it adds than for what it refuses to add. Read carefully, it is a more disciplined proposal than the surveillance debate usually produces — though it leans on an experiment that France's own evaluators found wanting.
Steelman the objection first
Civil-liberties groups have a serious case, and it deserves to be stated plainly. Permanent algorithmic monitoring of public space changes the texture of public life: people behave differently when they know an algorithm is watching, and the chilling effect on protest and ordinary movement is real, not hypothetical. France's own Conseil constitutionnel, striking down an earlier extension on 24 April 2025, grounded its reasoning partly in the freedom of movement enshrined in the 1789 Declaration. "Event detection" is also a slippery category: today's anomaly list is short, but the same camera and pipeline can host a longer one tomorrow. Function creep is the default trajectory of surveillance infrastructure, not the exception. Anyone who waves this away is not arguing in good faith.
What the bill actually constrains
That said, the Verzelen text is built around the constraints privacy advocates have long demanded. Its exposé des motifs and articles are explicit on four points:
- No faces. The bill prohibits facial recognition and any creation of biometric files. The CNIL, France's data regulator, has drawn exactly this line: augmented cameras may categorise events but may not identify individuals, and "devices cannot involve processing biometric data, notably facial recognition."
- No data fusion. The text bars interconnection or cross-matching with other databases — closing off the path by which anonymous event-detection quietly becomes individual tracking.
- Humans decide. The algorithm, in the bill's own words, "ne sera à l'origine d'aucune action" — it raises an alert; a person decides what, if anything, happens next. No automated enforcement.
- Prefect authorisation, necessity and proportionality. Each deployment must be authorised by the prefect and justified as necessary and proportionate, the GDPR Article 23 standard for limiting data rights.
This is the right architecture. The meaningful danger in public-space AI is biometric mass identification — the persistent, suspicionless tracking of named individuals. A system that detects a fire or a fall but is legally and technically barred from recognising who you are sits on the defensible side of that line. Banning a smoke-detection algorithm because a different, prohibited system could recognise faces is a category error that would forfeit genuine public-safety gains for no privacy benefit.
The evidence problem
The weakness is not the design; it is the record the bill cites to justify itself. France ran exactly this experiment under the 19 May 2023 Olympics law, which authorised eight use cases through 31 March 2025 with no facial recognition. The verdict was unflattering. A parliamentary flash mission reported on 19 March 2025 that, of the eight authorised use cases, four did not work at all, and the remaining four were "not particularly useful." Rapporteur Stéphane Peu's standout success story was detecting a lost mushroom-picker. The government's official evaluation committee, chaired by Christian Vigouroux and delivered in January 2025, was only marginally kinder, calling the operational interest "real" but "limited," with performance varying sharply by operator and scenario.
The bill's claim that "aucune dérive n'a été démontrée" — no abuse was shown — is true but incomplete. Absence of abuse is not the same as demonstrated usefulness, and a permanent national rollout demands the second, not just the first. A proportionality test cuts both ways: a measure that intrudes on public space while delivering thin, unproven benefit is, by definition, not proportionate.
How to legislate this well
The pro-innovation answer is not to kill the bill but to make it earn its permanence. The earlier extension fell on a procedural "cavalier législatif" technicality — smuggled into a transport-safety law with no link to it — which is precisely the wrong way to enact lasting surveillance powers. A standalone bill, debated openly, is an improvement. Parliament should now add what the design implies but does not yet guarantee: published per-use-case performance thresholds, with deployments retired when accuracy falls short; independent audits and public transparency reporting on every prefectoral authorisation; a hard-coded list of permitted events amendable only by primary legislation, not decree; and a sunset-and-review clause so "permanent" still means periodically re-justified.
France is right to insist that algorithmic cameras need not mean facial recognition, and right to keep a human in the loop. But a proportionate surveillance law has to clear an evidentiary bar, not just a procedural one. On the cameras' own demonstrated record, Verzelen's bill has not yet cleared it — and that, not the spectre of faceless biometric dystopia, is the fight worth having.