A camera sees a square filling up. A model can compare flows and warn that an exit is blocked. That may help organizers and emergency services, but proves neither violent intent nor identity.

Detect a situation, do not judge a person

Counting silhouettes, detecting smoke or collective running differs fundamentally from recognizing a face. The first functions can open an emergency corridor. The second directly affects public freedoms.

Shadows, umbrellas, clothing and dense crowds create errors. An alert is a lead to verify, never proof.

European law creates barriers

The EU AI Act tightly restricts real-time remote biometric identification in public spaces for law enforcement, with limited exceptions and authorization conditions. Technical availability is not a legal basis.

Disproportionate surveillance may discourage lawful protest. Purpose, area, duration, bias tests, public information and avenues for challenge must be documented.

Useful AI must be auditable

A defensible system shows uncertainty, records decisions and separates crowd-management data from investigative files. A trained operator reviews and may reject an alert.

AI may speed up detection of congestion or smoke; it becomes dangerous when visual correlation becomes individual suspicion.

Conclusion

The best AI protects without cataloguing: collective alerts, data minimization, human oversight and no automated adverse decision.