The strongest case for the police
The case for facial recognition at a large protest is not frivolous. Crowds in central Delhi can turn violent, and officers cannot scan thousands of faces by eye. A system that flags people wanted for serious crimes might help find fugitives and keep a demonstration safe. Delhi Police has also told the Supreme Court that it works: in an affidavit filed on 17 August, it said the system identified 2,873 people with criminal antecedents at the Jantar Mantar protest site between 20 and 26 July 2026, according to the Internet Freedom Foundation (IFF).
The trouble is not that police use technology. It is that the paper trail does not support what they did with it.
What the RTI reply actually shows
IFF published the document that Delhi Police and the Home Ministry have pointed to as the legal basis for the system. It is Standing Order No. Tech. & PI/04/2022, dated 9 June 2022, on the city's C4i command centre. Delhi Police released it to IFF on 23 September 2026, and IFF published it on 2 October. According to IFF's reading, the order covers how 112 calls are escalated, which CCTV feeds reach the command centre (Jantar Mantar among them), 204 number-plate cameras at 35 sites and a list of virtual-patrolling duties. It never mentions facial recognition. The words biometric, database, watchlist, threshold, retention, deletion, privacy and audit do not appear. Its final clause says it exists for the "internal smooth functioning" of the department.
That sits awkwardly with how the order has been described publicly. The News Minute reported that, according to the Ministry of Home Affairs, the 9 June 2022 order authorises facial recognition for "suspected individuals or known criminals." If IFF's account of the text is right, the ministry's description is not what the document says. This is worth checking against the order itself, which readers can now do.
The wider RTI record is thin. IFF says that across 161 answers in 17 letters, only 12 contained any information; no reply disclosed a vendor contract, match threshold, accuracy audit, retention period or privacy impact assessment. A 24 July IFF post adds that a Delhi Police 2022 RTI reply treats an 80% similarity score as a positive match, and that a CCTV Cell report dated 3 February 2022 reportedly said no specific rule governs facial recognition.
Why a standing order is not 'law'
In Justice K.S. Puttaswamy v. Union of India (2017), the Supreme Court held that privacy is a fundamental right. As The News Minute and IFF both summarise it, any restriction must meet three tests: legality, a legitimate aim and proportionality. The first test comes before the others. If no law authorises the intrusion, the court never needs to ask whether it was proportionate.
An internal order that binds only the force that wrote it, and that says nothing on faces, databases or retention, is a weak candidate for "law." IFF also notes that the Criminal Procedure (Identification) Act, 2022 deals with photographs of convicted, arrested or detained persons. It does not authorise scanning a crowd. The Digital Personal Data Protection Act, 2023 offers little help either; The News Minute describes it as a general framework with broad government exemptions that does not specifically regulate biometric surveillance in public spaces.
The Supreme Court's 2018 Aadhaar judgment applied this same Puttaswamy framework to a large state data system. That judgment shows the Court expects an identifiable legal footing and safeguards for such systems, not a department's own say-so.
Why accuracy matters more than the headline number
The 2,873 figure is the police's own claim, and the one independent check we have is unflattering. IFF's 4 September statement reports that the Indian Express reviewed 205 people accused of murder, rape and attempted murder among those the system flagged, and found at least 25 who were in Tihar, Mandoli or Rohini jail when the software placed them at the protest. That is about 12% of that sample, and each of those matches was physically impossible.
The sample is not random, and 25 is a floor. Still, it shows that a flag from this system is a lead, not evidence. That matters because on 1 September the Supreme Court, at the Union's request, quashed FIRs against the student protesters but allowed police to register fresh and specific FIRs against the 2,873 people, per IFF. A list of unverified matches is now the starting point for criminal cases.
A proportionate way forward
This publication favours technology adoption, including by the state, and opposes bans where narrower rules would do. The remedy here is not to forbid facial recognition. It is to make its use lawful and checkable:
- Pass a statute that names the permitted purposes, and excludes peaceful assemblies as a default surveillance target.
- Require published thresholds and independent accuracy audits before any system is used on the public, with impossible-match rates reported.
- Set retention and deletion limits and a prohibition on building lists of protesters.
- Treat a match as a lead only, never as grounds for an FIR without human verification.
- Mandate a privacy impact assessment and publish it, as IFF says none was done for the Safe City project.
Good rules help police too. A system that can show its error rate and its legal footing can survive a court challenge and earns public trust. One that rests on an order silent on faces cannot. Until Parliament or the courts supply the missing law, the Jantar Mantar deployment fails Puttaswamy's first test, whatever the system caught.