How sure should a camera be before it puts your name on a list that a court will read? Delhi's police have answered in writing. In 2022, after the Internet Freedom Foundation (an Indian digital-rights group) took its information requests to appeal, the force disclosed that matches "above 80% similarity are treated as positive results," in IFF's words.
If your first reaction to the headline was good, keep people with violent records away from a protest, that's fair. It sounds like the least sinister thing a police camera can do. If the list had held up, this would be a short piece.
Start with the list. On September 1, 2026, India's Supreme Court threw out the FIRs (First Information Reports, the filings that open criminal cases) over the student protests of July 20 to 25, barred new ones, and made one exception the Delhi Police had asked for: the police "shall be at liberty to register a fresh FIR with respect to the 2873 individuals." Who are they? People "allegedly having serious criminal antecedents as per the National Crime Records Bureau database and who were prima facie reported to be present at the protest site."
Allegedly. Prima facie. Reported. The Court made no finding about these people; it described what the police had told it.
Three days later, the Indian Express published its check of 205 names from the list's most serious categories: at least 25 were in jail as per police, jail and court records when the system flagged them at Jantar Mantar, the protest site, with custody dates that "precede the protest — weeks, months, in some cases years earlier." One entry: "identified around 4.24 pm, July 25 ... First jailed on November 17, 2024; again on November 29, 2024."
The police's reply to the paper: "2873 individuals identified as prima facie having criminal antecedents were spotted. However, further verification in this regard is pending."
So what should a match be allowed to prove before anyone acts on it? And what should you ask before your own city's cameras get the same upgrade?
An affidavit, a threshold and a thin rulebook
How does a camera's guess reach a Supreme Court order? Through a sworn statement. On August 17, the Delhi Police told the Court in an affidavit that its Facial Recognition System had identified 2,873 individuals with past criminal records at Jantar Mantar, the Indian Express reported: 2,402 "identified through the Crime Kundli database" and 471 "through dossier records." The affidavit promised "field verification," adding: "No actions are taken solely based upon the result of the facial recognition system which is only the first step."
What it doesn't say, the Express noted, is "whether those identified were accused, convicts, or merely named in criminal cases." The Court's order describes the 2,873 through the National Crime Records Bureau database and never mentions facial recognition.
And the threshold? Eighty percent is a similarity score — roughly how alike the software judges two faces to be, not an error rate — and the force has not published the false-positive rate at its threshold, MediaNama, an Indian tech-policy outlet, reported in September.
What law governs any of this? Not much. India has no dedicated framework for police facial recognition, the Indian Express explained in July: no rules on watchlists, matching thresholds, independent audits or retention periods. Asked in Parliament in 2025 what safeguards prevent "unauthorized surveillance by police through facial recognition," the government cited a standing order the Delhi Police issued in 2022. IFF obtained the order in September and says the words "facial recognition" "do not appear anywhere in it."
Is it lawful, then? The Court hasn't said. In August it appointed an inquiry committee whose terms include police "monitoring and surveillance of protestors," and kept "the broader constitutional questions" for itself. A member of Parliament, A.A. Rahim, has petitioned it, arguing that the police carried out surveillance in "a complete legal vacuum"; his petition is pending. As of October 24, 2026, there's no ruling on facial recognition.
And the cameras keep coming. A tender the Delhi Police posted on July 30, to cover what it calls the "VIP/VVIP route" on Republic Day and Independence Day, would "require face recognition on every fixed and PTZ camera" (the kind that pans, tilts and zooms), IFF reported, with the feature licensed on 10 percent of them. License-plate reading must hit "90% for day time and 70% for night time"; faces get no standard. The tender "sets no accuracy standard for facial recognition and does not specify a database," MediaNama found, though such details "could be covered by another product, framework or workflow." Among the sites, per IFF, are the approaches to Ramlila Maidan, which the police named a designated protest site on October 11 (IFF has "no evidence that the two decisions were coordinated"). The same order de-notified Jantar Mantar, a move the Supreme Court has agreed to examine. As of October 24, 2026, no award is public, though a police source told The Times of India "the work order has been issued."
The police's best case (and it isn't nothing)
Is this fair to the police? Their case deserves its strongest form.
First, they say the system doesn't watch everyone. It "does not automatically capture, create, generate or maintain profiles of every individual present at the protest site," the affidavit says, and Solicitor General Tushar Mehta told the Court it "does not identify everybody indiscriminately." (Checking whether a face is on a list does mean looking at every face. The question is what happens to the rest.)
Second, the protesters were protected, at the government's own request. The order records the government's and four states' "acknowledgement of the fact that mere participation in the protests is not taken as an offence under the penal laws," and the FIR the police sought concerns "bodily harm" or "destruction of public property," they told the Court. From the bench, Chief Justice Surya Kant said: "We don't say they are offenders. Allegedly."
Third, the sample wasn't random. Because the Express checked serious-offense categories, its findings, The Federal notes, "cannot yield an error rate for all 2,873," though finding 25 in prison "establishes that some claimed sightings were impossible."
Fourth, could some of the 25 have been briefly out that day, on parole or at a hospital? It's a fair question, one the police's printed reply doesn't raise. Delhi's prison rules, as the Delhi High Court has quoted them, say under-trial prisoners (those held awaiting trial) "are not eligible for regular parole and furlough," though a trial court can order "custody parole" of "not more than six hours." Any such explanation could be checked against the records. As of October 24, 2026, the police haven't offered one on the record.
So far, so good, as long as the first step stays a first step. This one didn't. By the police's own account, verification was still pending when the list became the basis for an exception to a Supreme Court order — one that "prescribed no method or express requirement that checks precede arrest or other action," as The Federal observed. Five weeks later, police sources told the Indian Express that everyone flagged in July would be "tracked by their local police stations" ahead of an October 10 protest. On October 8, the Delhi Police's Crime Branch registered the fresh FIR the order allowed, police sources told PTI; it will "examine the role of 2,873 people" flagged in July. The same wire report has police putting the system's accuracy at "approximately 50 per cent"; it doesn't say at what threshold, or how that was measured. Officials cited by The Tribune said the Crime Branch "will verify whether the individuals were actually present at the protest site," so the field check will run inside a criminal case. The first step is where a name enters the record. And records travel.
Buenos Aires asked where the faces came from
What happens when a court starts with the list rather than the match? Buenos Aires offers a test case. (Its court records are in Spanish; English renderings here are translations.)
In April 2019 the city began checking passing faces against CONARC — Argentina's national register of people wanted on warrants — with a fugitive-search system called the SRFP, and that register, Human Rights Watch found in 2020, had errors of its own: "a 3-year-old is listed as being wanted for aggravated robbery." The city blamed the register for mistaken matches.
Then a judge looked. In September 2022, ruling in a suit brought by rights groups, Judge Elena Liberatori declared the way the system had been switched on unconstitutional and voided what had been done without a verifiable court order. The city's 2020 law, as she quoted it, barred loading anyone outside CONARC "Salvo orden judicial" (absent a court order) and set up a legislative oversight committee — one the legislature itself had reported was not yet constituted. She also counted. CONARC held about 35,000 people, yet the city had made 9,900,282 requests to the national identity registry between April 2019 and March 2022, and "15.459 personas fueron cargadas en el SRFP para ser buscadas sin encontrarse dentro de la base CoNaRC": 15,459 people were loaded into the system without being on the warrant list. Any restart had to wait for working overseers.
In April 2023 an appeals chamber upheld the ruling, adding that the software must first be tested for whether it affects people differently depending on their personal characteristics, and that the public must be told before any restart.
Today the system is off. Amnesty International's August 2026 report says it "continúa suspendido" (remains suspended), and the city — explaining to Amnesty why a commercial face-search tool it contracted in November 2025 wasn't in use — cited "las disposiciones judiciales vigentes" (the court orders in force). On September 25, 2026, the court held a hearing on a tender to replace the suspended system, according to Vía Libre, a digital-rights group in the case as a friend of the court, which wrote on October 16 that the court must now decide whether to authorize a new system. As of October 24, 2026, it hasn't.
Before this becomes a fable, the city's side. The system had barely run since April 2020, when masks defeated it, and the city credits it with putting 1,695 "delincuentes" (offenders) before the courts. In July 2022 the city's high court sided with the city and removed the judge who had first suspended the system. And one wrongful arrest that made the news — a man held for six days in 2019 — came from a badly loaded search, judicial sources said: the system "funcionó bien, pero estaba mal cargada la búsqueda" (it worked fine, but the search was loaded wrong). The error was in the list the cameras searched, not in the match. Delhi's police haven't said which failed when their system put at least 25 jailed men at Jantar Mantar.
Argentina hasn't solved facial recognition. What its judges did was narrower: they asked which database fed the cameras, counted the people who shouldn't have been in it, and kept the system off until the software could be checked. India's Supreme Court let the police act on a facial-recognition list in an order that never mentions the technology, though it has kept the constitutional questions for itself.
Picture the next protest
What follows is my extrapolation from the documents above.
Picture the next big demonstration in Delhi, at Ramlila Maidan. If the July tender is awarded as written, the cameras on the approach will be built for face recognition and will count "people gathering." Against which database? The tender doesn't say.
Now load last summer's list. It already underpins an October FIR; imagine it becoming the starting point for every protest after, with no published count of how many July matches anyone checked.
Then the paperwork. An executive at a company that says it provides the software layer behind a police surveillance van told Coda Story that once its threshold is crossed, "it is a 100 percent match for us" (he wouldn't say whether its technology was used at Jantar Mantar). The Indian Express's investigation said a match "is not, by itself, proof of identity." Which of those sentences ends up in the next affidavit?
Finally, put the camera on an officer's face. A smart-glasses company told Coda its glasses were used at Jantar Mantar during the exam protests; its executive also said 14 people were arrested after being recognized through them (the company's claims). Imagine the first step taken in a glance, with no van and no log.
Each step would look reasonable in its own filing. Together, they'd turn a similarity score into a record that follows someone from one protest to the next.
Who says what, from the police to their critics
Where do the people who study this land?
The police's view is in their affidavit: "it is in the interest of peaceful protestors that hardened criminal elements are weeded out by use of latest technology." In the middle, the U.S. National Academies' 2024 report "does not advocate for an overall ban," though it says some uses may warrant prohibition, and India's IDFC Institute in 2021 credited police facial recognition with "potentially reducing 'third-degree methods'" (coercive interrogation) while calling reliance on it "premature." Purushraj Patnaik, writing for the Observer Research Foundation, argues that "Governed with care, AI can ease the operational burden on India's police," but that algorithmic policing there has grown "without an anchoring statute." Even NITI Aayog, the government's policy think tank, warned in a 2022 draft that misidentification "may lead to gross violations of a person's right to life and liberty."
The critics go further. IFF argues that "there is no score at which a person is cleared" and wants the list withdrawn, with field checks suspended until an independent audit is published. A 2021 Vidhi Centre for Legal Policy study recommended that police facial recognition in Delhi "be halted until the questions of equality and over-policing are examined." Sarthak Gupta, a New Delhi lawyer and Columbia University researcher, called the source "a police affidavit which was admittedly unverified." The technologist Rohini Lakshané told MediaNama that the people behind "criminal record" (convicts, suspects, arrestees) "are legally different categories." NYU's Katie Kinsey warned Coda that a model "may behave differently when the cameras, lighting, image quality and population change."
Line them up and the police and their critics agree on more than it seems: a match, by itself, settles nothing. The affidavit comes close to saying so — a match is "only the first step." The fight is over who checks, when, and whether anyone has to show it.
What does this mean for you?
Wherever you live, you can:
Ask the regulator's questions before the upgrade. In October, Spain's data-protection agency told a town council that facial recognition must stay switched off until the council justifies it for a specific policing purpose (a preventive warning, the agency says, not a finding of wrongdoing). Its warning spells out, in translation, what the assessment must cover: the system, the reference images or databases, the comparison procedure, the match thresholds, false positives and negatives, human intervention, and the consequences of a match. Take those seven to your council.
Ask what a match may trigger. Patnaik proposes amending India's criminal procedure code to bar arrests or custody extensions "based solely on a facial recognition match." Under the EU's AI Act, as the Future of Privacy Forum summarizes it, when police use live face recognition in public, a decision with an adverse legal effect "cannot be based solely on a system's output." Does your police force have either rule?
Ask for the error rate at the threshold in use. A threshold without one is half an answer, and Delhi has given only the first half.
Ask who keeps the faces, and for how long. IFF found "no retention period set out for face templates, match logs or plate reads" in Delhi's tender. Buenos Aires's draft replacement tender bars the contractor from reusing the data to train AI models "sin autorización previa y expresa" (without prior, express authorization). Ask what your city's contract says.
In India, use the Right to Information Act, and expect to appeal. IFF got the 80 percent threshold only after second appeals.
If you were at the July protests, the Supreme Court's inquiry committee opened a portal for complaints on October 17 and accepts anonymous complaints; police surveillance of protesters is within its terms.
The lesson, as I see it
A face match is a probability with a number attached. What comes next (the check, the file, the court filing, the station's list) decides whether it hardens into a fact about a person. In Delhi, the number was disclosed, the checking was "pending," and the Supreme Court carved out an exception for the people on the list anyway, three days before a newspaper reported that at least 25 of them had been in jail when the system flagged them.
What should you watch? What the Crime Branch's new FIR does with the 25. The committee's first interim report, requested "at the earliest" and, as of October 24, 2026, not public. Rahim's petition, listed for November 2. The award of the July 30 tender. And the Buenos Aires court's ruling on a replacement.
My vote? Before a camera's guess can put a name in a court filing, the police should publish three things: the threshold, the error rate at that threshold, and how many matches a human checked and was free to reject. Delhi has formally disclosed one of the three, and only after an appeal. A system that can name 2,873 people in a week should be able to say how many it got wrong.
The HAIA Foundation works so that the people a camera can name have a say in what its guesses may prove. To read the next list before it reaches a judge, subscribe.