It was a working day like any other when the police arrived at the door. Alvi Choudhury, twenty-six years old, a software engineer of Bangladeshi origin, was logged in from the family home in Southampton, the unremarkable rhythm of a remote developer's morning unfolding around him: meetings, messages, lines of code. Then officers were on the step. They handcuffed him. They told him he was suspected of a burglary, a theft of roughly three thousand pounds' worth of property, that had taken place in Milton Keynes. Milton Keynes is a hundred miles away. Choudhury had never been there.
What had placed him at the centre of an investigation in a town he could not have found on a map was not a witness, not a fingerprint, not a stray possession left at the scene. It was a piece of software. Thames Valley Police had run footage of the genuine suspect through a facial recognition system and the algorithm had returned a name: his. From that single computational suggestion flowed everything that followed, the cuffs, the cell, the nearly ten hours in custody before he was released at two o'clock in the morning.
Choudhury looked nothing like the man in the footage. By his own account the suspect appeared roughly a decade younger, with lighter skin, no facial hair, a larger nose, different eyes, different lips. He said as much. He asked the officers, in a question that ought to chill anyone who believes a custody suite is a place of careful judgement, whether the image on the screen looked anything like him. According to his account, they laughed. Thames Valley officers, he later said, conceded that they had known he was not the suspect after comparing the footage with his photograph. They proceeded anyway.
This is the part that should not be skated over. The failure here was not, in the first instance, the machine's. Machines err. The deeper failure was that a chain of human beings, each with the authority to stop, looked at a face that did not match and kept going regardless, because a computer had spoken and nobody felt empowered, or obliged, to overrule it. That is not a glitch. That is a governance vacuum wearing the costume of efficiency.
Choudhury's case is not an outlier. It is a data point in a pattern that has been hardening for years on both sides of the Atlantic, and the pattern raises a question that policing has so far refused to answer with any rigour. As forces across Britain and the United States race to bolt facial recognition onto the front end of the criminal justice system, what framework of rights, transparency and accountability would actually need to exist before an algorithmic match could justify depriving a person of their liberty? And when the machine is wrong, as it demonstrably and repeatedly is, who carries the burden of proving it?
A Grandmother, A State She Had Never Seen, And Five Months Gone
To understand how badly this can go, travel from Southampton to Tennessee.
In July 2025, Angela Lipps, fifty years old, was arrested at her home. The warrant came from North Dakota, a state she said she had never visited, in connection with a series of bank frauds in and around Fargo. The thread that tied her to those crimes was Clearview AI, the controversial facial recognition company whose database is built from billions of images scraped from the open web. West Fargo police had run their suspect's image through the system. A detective looked at the result, looked at Lipps's social media, decided it was a match, and filed charges.
Because Lipps was treated as a fugitive, arrested in one state on another state's warrant, she was held without the possibility of bail. She sat in a Tennessee jail for close to four months before being transferred to North Dakota. In total she spent more than five months in custody for crimes committed by someone else, in a place she had never been. The charges were dismissed on Christmas Eve 2025. By then her lawyer, Jay Greenwood, had assembled the kind of evidence that ought to have made the arrest impossible in the first place: bank records showing Lipps in Tennessee throughout the period of the offences, depositing her Social Security payments, buying groceries, ordering takeaways. An ordinary life, documented in receipts, that placed her hundreds of miles from the crimes.
The cost of the error was not abstract. While she was inside, Lipps lost her home. She lost her car. She lost her dog. She came out to a life that had been dismantled by an algorithm's confidence and a detective's willingness to trust it. A crowdfunding campaign later raised tens of thousands of dollars to help her rebuild. Police in Fargo acknowledged that errors had been made and promised changes, but stopped short of a clear apology.
Lipps and Choudhury are separated by an ocean, a legal system and a decade in age. What unites them is the structural mechanism of their misfortune. In both cases a probabilistic suggestion, a statement that two faces resemble one another to some degree of confidence, was treated as though it were an identification. In both cases the human beings in the loop, the people whose job is precisely to be sceptical, deferred to the output of a system they did not build, could not interrogate and apparently did not fully understand.
The Bias Is Not A Rumour, It Is A Measurement
There is a temptation, when confronted with cases like these, to treat them as freak accidents, the unlucky tail of an otherwise reliable technology. The research does not support that comfort.
In March 2026, Essex Police paused their use of live facial recognition cameras. The trigger was an independent assessment, conducted with researchers at the University of Cambridge, of how the force's system performed across demographic groups. The findings were stark. Black people, the analysis indicated, were around twenty-seven per cent more likely to be flagged by the system than people of other ethnicities, and roughly thirty-one per cent more likely than white people. Of the small number of people the study found to have been misidentified, a disproportionate share were Black, a result the researchers concluded was unlikely to be down to chance. To its credit, Essex did what a great many forces have not: it stopped, examined the evidence and suspended deployment while it worked with its software supplier to understand what had gone wrong.
The Essex finding did not arrive in a vacuum. It is consistent with one of the most thoroughly documented bodies of evidence in applied machine learning. In 2019 the United States National Institute of Standards and Technology, the federal metrology body whose job is precisely to measure such things, published the third part of its Face Recognition Vendor Test, a study of demographic effects across nearly two hundred algorithms from around a hundred developers, drawing on millions of images. Its conclusions were unambiguous. For one-to-many matching, the kind used to search a face against a large database, false positive rates were highest for Black women, higher for women than men, and higher for African American and certain other groups than for white subjects. The disparities were not marginal. Within-group false positive rates, NIST found, could vary by enormous factors depending on demographic group, and these false positive differentials persisted even in clean, high-quality images. The elderly and the very young also fared worse.
Read those findings against the cases and the cruel logic snaps into focus. The systems fail most often, and most consistently, for exactly the people most likely to find themselves on the receiving end of policing: people of colour, women, the young. The technology's weakest performance is concentrated precisely where its consequences are heaviest.
Then, on the last day of April 2026, a paper landed on the preprint server arXiv that gave this pattern an even sharper edge. Titled “MIFair: A Mutual-Information Framework for Intersectionality and Multiclass Fairness”, and authored by Jeanne Monnier, Thomas George, Frédéric Guyard, Christèle Tarnec and Marios Kountouris, the work tackled a problem that single-axis bias measurement tends to obscure. Most fairness analysis asks how a system performs for women, or for Black subjects, or for younger people, one protected characteristic at a time. MIFair, using a framework built on the information-theoretic concept of mutual information, was designed to handle intersectionality directly: the compounding that happens when those characteristics overlap in a single person.
This is the crucial move. A system can look tolerably fair along each individual axis and still fail badly for those who sit at the intersection of several. A young Black woman is not simply the sum of three separate, manageable risks. The errors stack. The MIFair authors demonstrated that their approach could surface and reduce bias in exactly these previously neglected multi-attribute scenarios, the corners of the demographic map where conventional fairness audits tend not to look. The implication for policing is uncomfortable in the extreme. The people for whom these systems are least reliable are not a single, easily named group. They are defined by the collision of characteristics, and that collision is most concentrated in the communities already most heavily surveilled.
In other words, the bias is not a rumour, a vibe or an activist talking point. It is a measurement. It has been measured by the United States government's own standards body, by academics commissioned by a British police force, and by fairness researchers refining the very mathematics of how we detect it. The question is no longer whether commercially deployed facial recognition carries demographic error. It is what, knowing this, we are prepared to let it do to people.
Why The Humans Keep Deferring To The Machine
There is a second failure mode running alongside the technical one, and it is arguably more dangerous because it cannot be patched with cleaner training data. It is a failure of human psychology, and it has a name: automation bias. People tend to over-trust the outputs of automated systems, to treat a number on a screen as more objective than a colleague's hunch, and to relax their own scrutiny in proportion to how confident and quantified the machine appears. A facial recognition system that returns a candidate alongside a crisp similarity score does not merely offer a suggestion. It offers the seductive impression of mathematical certainty, and that impression is corrosive to the very scepticism that good policing is supposed to embody.
This helps explain the otherwise baffling detail at the heart of the Choudhury case: officers who, on his account, could see he was not the suspect, and detained him anyway. By the time he was in the cell, the match had already done its work. It had converted a person from a member of the public into a suspect, and reversing that conversion would have required someone to actively distrust the system, to take personal responsibility for overruling it, and to absorb whatever institutional risk attaches to letting a flagged individual walk back out of the door. It is psychologically far easier to defer. The machine becomes a place to put the decision, and therefore a place to put the blame. Nobody quite chose to jail Angela Lipps; the system flagged her. Nobody quite chose to handcuff Alvi Choudhury; the algorithm matched him. That quiet diffusion of responsibility is not an incidental side effect of automated policing. It is one of its central, and least examined, features.
The scale at which this is now operating should sharpen the concern. In the Thames Valley area alone, live facial recognition deployed from late 2025 across Oxford, Slough, Reading, Wycombe and Milton Keynes is reported to have scanned in the order of a hundred thousand faces, returning a handful of arrests, while retrospective searches run at a rate of tens of thousands a month against a database of roughly nineteen million custody images. Each of those searches is an occasion for the demographic error rates documented by NIST and Cambridge to express themselves in a real person's life. When you run a process with a known bias at that volume, rare failure rates stop being rare in absolute terms. They become a steady production line of Choudhurys, most of whom will never make the news, never instruct a solicitor, and never learn that an algorithm was the reason a morning at their desk ended in handcuffs.
The Patchwork Britain Calls A Legal Framework
You might assume that a technology capable of putting an innocent grandmother in jail for five months, or an innocent engineer in a cell until two in the morning, would be hemmed in by dense and specific law. In Britain, you would be assuming wrong.
There is, to this day, no bespoke statute governing police use of facial recognition in England and Wales. What exists instead is a patchwork: general data protection law, human rights law, equality law, common law police powers, and a scattering of force-level policies, stitched together and asked to bear a weight they were never designed for. The closest thing to a foundational ruling came in 2020, when the Court of Appeal decided the case brought by the campaigner Ed Bridges against South Wales Police. The court found the force's use of live automated facial recognition unlawful. It held that the deployment breached the right to privacy under Article 8 of the European Convention because it was not “in accordance with law”, that the legal framework left too much discretion to individual officers about where and how to deploy, that the force's data protection impact assessment was deficient, and that South Wales Police had not taken reasonable steps to satisfy itself that the software was free of racial or gender bias, in breach of the Public Sector Equality Duty.
That was a landmark, and for a while it read like a brake. Yet the years since have seen the technology expand rather than retreat. In January 2026 the Home Office announced what was billed as the largest facial recognition rollout in the country's history, including the purchase of dozens of new live facial recognition vans intended to put the capability into every regional force in England and Wales, framed around the protection of women and girls and the pursuit of violent and sexual offenders. The government committed substantial funding to a national centre for artificial intelligence in policing. A consultation on a new legal framework, intended to consolidate the existing patchwork, closed in early 2026, with fresh legislation expected to be years away. In the meantime, deployment continues.
The courts, too, have shifted. On 21 April 2026 the High Court dismissed a challenge to the Metropolitan Police's live facial recognition system, ruling it lawful and compatible with human rights, with the judges describing the force's policy as containing clear, interlocking and cumulative constraints. The man the Met had wrongly identified signalled his intention to appeal. The legal centre of gravity, in other words, is moving from scepticism towards accommodation, even as the documented harms accumulate. The danger of that drift is that it allows the absence of specific legislation to be reframed as a settled question rather than the open wound the Bridges judgment said it was.
And here is the structural detail that the Choudhury case throws into relief. Much of the public debate fixates on live facial recognition, the vans, the cameras scanning crowds in real time. But Choudhury was not caught by a van. His ordeal began with retrospective facial recognition, an officer running an image from a crime scene against a database of stored custody photographs. His photograph was in that database for a reason that compounds the injustice: he had been wrongly arrested once before, in 2021, after being attacked on a night out while a student in Portsmouth, an incident that led to no further action but left his image in the system. He was, in effect, made permanently searchable by a previous wrong done to him, and then matched in error to a second crime he had nothing to do with. The database does not forget, even when the justice system has decided there was nothing to remember. The longstanding failure to routinely delete the custody images of people never charged turns a single injustice into a permanent liability, a face left loaded into a machine that will keep proposing it for years.
What America Learned, And What It Cost
If Britain is early in this story, the United States is several painful chapters ahead, and its experience offers both a warning and, faintly, a template.
The first publicly known wrongful arrest by facial recognition belonged to Robert Williams, a Black man arrested by Detroit police in January 2020, in front of his wife and two young daughters, for a shoplifting he did not commit. The match that doomed him came from a blurry surveillance still run through a state database; Williams was, by the system's own ranking, merely the ninth-best candidate. He spent some thirty hours in detention. His case became the foundation of a legal campaign that culminated, in 2024, in a settlement widely described as producing the strongest police facial recognition policy in the country. Detroit agreed to pay damages, and, more importantly, to bind itself to rules: officers would be prohibited from arresting anyone based solely on a facial recognition result; they could not run a lineup off the back of a raw algorithmic lead without independent, reliable evidence linking the suspect to the crime; and the department agreed to audit its past cases and to train officers on the technology's documented tendency to misidentify people of colour at higher rates.
Williams was not alone. A Washington Post investigation published in early 2025, drawing on records from nearly two dozen police departments, identified at least eight Americans wrongfully arrested on the basis of facial recognition, seven of them Black. The investigation's most damning finding was not that the technology failed, but how routinely the humans did. Across these cases, police repeatedly skipped the most basic confirmatory steps: alibis went unchecked, contradictory evidence was ignored, key evidence was never gathered. In case after case, officers treated a software suggestion as a settled fact. In Louisiana, police relied on an incorrect Clearview AI result as the purported basis for a warrant, leading to the arrest of a Georgia man, Randal Quran Reid, who had never set foot in the state and who spent close to a week in jail. The pattern Angela Lipps fell into was not new. It was a template the system had run many times before.
What makes the American record so instructive is that the corrective standard was never secret. The industry itself has long warned that a facial recognition result is a lead and nothing more. The International Association of Chiefs of Police has cautioned that such a result is a strong clue that must be corroborated against other evidence before anyone is identified. Vendors print the warnings; policies repeat them. And yet, by the ACLU's analysis, in most of the documented wrongful arrests officers had received exactly those warnings and made the arrest regardless. The warning, on its own, prevented nothing. A caution that everyone is free to ignore is not a safeguard. It is paperwork.
This is the lesson Britain is currently choosing not to learn at speed. The mechanism that put Robert Williams, Randal Quran Reid and Angela Lipps in cells is the same mechanism that put Alvi Choudhury in one. The presence of a written instruction to corroborate did not save the Americans, because nothing compelled compliance and nobody bore a meaningful penalty for ignoring it. Britain is now importing the capability while leaving the only proven constraint, hard rules with consequences attached, conspicuously to one side.
The Framework That Would Actually Have To Exist
So what would a serious answer look like? Not a press release, not a consultation that reports back in two years' time, but an actual framework of rights, transparency and accountability sturdy enough that an algorithmic match could responsibly form part of the basis for an arrest. Pull the threads of the evidence together and the requirements assemble themselves.
The first principle has to be that a match is a lead, never a conclusion, and that this is enforced rather than merely recommended. The Detroit settlement points the way: an explicit prohibition on arrest based solely or even primarily on a facial recognition result, paired with a requirement for independent, reliable evidence connecting a specific person to a specific crime before liberty is touched. Crucially, the corroboration cannot be circular. A photo lineup assembled from the algorithm's own suggestion is not independent confirmation; it is the same error wearing a different hat. Independent means evidence the machine did not generate: an alibi checked, a location verified, a piece of the physical world that places this person at this scene. Had anyone checked Angela Lipps's bank records before her arrest rather than after her release, she would never have seen the inside of a cell. Had anyone weighed Choudhury's hundred-mile distance and his on-screen work meetings against a face that plainly did not match, the cuffs would have stayed off.
The second principle is genuine transparency, the kind that produces accountability rather than performing it. Defendants and their lawyers are routinely never told that facial recognition featured in their case, which makes it nearly impossible to challenge. A serious framework would require disclosure, every time, that the technology was used, alongside the candidate list it produced, the confidence scores, the rank at which the accused appeared, and the demographic performance data for the specific system deployed. Robert Williams was the ninth-best match; that fact alone, disclosed early, should have stopped the case. You cannot contest a process you are not told occurred, and you cannot calibrate your scepticism towards a tool whose error profile is kept from you. Transparency of this kind also has a disciplining effect on the police themselves: an officer who knows that the algorithm's full workings will be laid before a court is an officer with a powerful incentive to corroborate before, rather than after, the handcuffs come out.
The third principle is that the burden of proof must sit where it has always belonged in a system that takes liberty seriously: on the state. This is the heart of the matter. In each of these cases the burden quietly inverted. The machine asserted guilt, and the accused was left to prove their innocence, from a cell, often without even knowing that an algorithm was their accuser. Choudhury offered his alibi and was detained anyway. Lipps's innocence was written in her bank statements the whole time, and still she lost five months and most of her life. A coherent framework must restore the proper order. It is not for the wrongly matched to demonstrate they are not the person on the screen. It is for the state to demonstrate, with evidence the algorithm did not manufacture, that they are, before an arrest is ever made. The machine's output should raise the evidentiary bar the police must clear, not lower it. The right answer to “the computer says it is him” is not “then arrest him”. It is “then go and find the evidence that proves the computer right, and if you cannot, leave him alone”.
The fourth principle concerns the systems themselves. Given the NIST findings, the Essex result and the intersectional compounding that the MIFair work makes legible, no force should be permitted to deploy a facial recognition system whose demographic performance has not been independently and publicly audited, with results disaggregated not just by single characteristics but by their intersections. A system that performs acceptably for white men and badly for young Black women is not an acceptable system with a minor caveat. It is a discriminatory instrument, and deploying it with knowledge of that profile is precisely the failure of the Public Sector Equality Duty that the Bridges judgment identified. The MIFair contribution matters here because it shows the auditing tools are improving; ignorance of intersectional error is becoming harder to claim as an honest excuse, and a force that deploys without such an audit can no longer pretend it simply did not know.
The fifth principle is consequence. The American experience proves beyond argument that warnings without enforcement change nothing. If officers can ignore the corroboration standard with impunity, they will, and people like Choudhury and Lipps will keep paying for it. A framework with teeth needs real remedies: suppression of evidence obtained through improper reliance on a match, meaningful liability for forces that arrest on an uncorroborated result, and independent oversight with the power to halt deployments, as Essex was willing to do voluntarily but most forces are not. Accountability that arrives only as a settlement, years later, after a life has already been dismantled, is not accountability. It is restitution for a harm that the system declined to prevent. The difference between a guideline and a right is that a right is something the state owes you whether or not it finds you convenient, and something it pays for when it fails to honour it.
The Vans Are Already Out
The difficulty is one of sequence. Britain is expanding deployment now, with the vans on order and the rollout under way, while the legal framework that might constrain it is promised for some indefinite later. That is precisely backwards. The history of this technology shows that capability deployed ahead of governance does not patiently wait for the rules to catch up. It generates facts on the ground, and people in cells, and a steady accumulation of harms that the eventual legislation will be asked to forgive rather than forbid.
It would be a mistake to read any of this as technophobia. The information-theoretic machinery underpinning a paper like MIFair is genuinely sophisticated, and the same research community documenting these failures is also building better tools to measure and mitigate them. Facial recognition is not witchcraft, and it is not useless. As a generator of investigative leads, narrowing a field, suggesting a direction, prompting the real police work of corroboration, it may well have a defensible place. The catastrophe is not the existence of the tool. It is the substitution of the tool for judgement, the moment a probabilistic resemblance hardens into a name on a charge sheet and the human beings in the chain stop asking the one question Alvi Choudhury asked from inside his predicament: does this actually look like me?
That question, simple enough for a frightened man in a custody suite to put to officers who laughed at it, is the whole of the matter. A facial recognition match is a hypothesis. A hypothesis is the beginning of an investigation, not the end of one. Until the law insists on that distinction, enforces it with consequences, and places the burden of proof back where it belongs, the machine will go on naming the innocent, and the innocent will go on bearing the cost of proving the machine wrong. Choudhury was released at two in the morning and is now seeking redress through the courts. Lipps came home to a life she had to rebuild from a crowdfunding page. They were the lucky ones, in the sense that they got out at all, and that someone eventually listened. The unanswered question, as the vans roll into every force in the country, is how many people the system will name before anyone with the power to stop it decides that resembling a stranger on a screen is not, and must never be, the same thing as being guilty.
References
Liberty Investigates. “Facial recognition error prompts police to arrest Asian man for burglary 100 miles away.” https://libertyinvestigates.org.uk/articles/facial-recognition-arrest-alvi-choudhury-police-bias/
Eastern Eye. “UK police wrongly arrest Asian man in Southampton after facial recognition error.” https://www.easterneye.biz/uk-police-facial-recognition-error-southampton-arrest/
CNN. “Police used AI facial recognition to arrest a Tennessee woman for crimes committed in a state she says she's never visited.” 29 March 2026. https://www.cnn.com/2026/03/29/us/angela-lipps-ai-facial-recognition
WRAL.com (Associated Press). “Tennessee grandmother jailed five months after AI facial match led to North Dakota extradition.” https://www.wral.com/news/ap/angela-lipps-tennessee-grandmother-jailed-five-months-misidentified-by-ai-fargo-nd/
State of Surveillance. “A Grandmother Lost Five Months, Her Home, Her Car, and Her Dog to a Facial Recognition Error.” https://stateofsurveillance.org/news/angela-lipps-facial-recognition-wrongful-arrest-five-months-jail-2026/
Liberty Investigates. “Essex police pauses facial recognition camera use after study finds racial bias.” https://libertyinvestigates.org.uk/articles/essex-police-live-facial-recognition-racial-bias-black-people/
OECD.AI Incidents. “Essex Police Suspends Live Facial Recognition Over Racial Bias.” 18 March 2026. https://oecd.ai/en/incidents/2026-03-18-4acf
National Institute of Standards and Technology. “Face Recognition Vendor Test (FRVT) Part 3: Demographic Effects” (NISTIR 8280). 2019. https://nvlpubs.nist.gov/nistpubs/ir/2019/nist.ir.8280.pdf
Security Industry Association. “What NIST Data Shows About Facial Recognition and Demographics.” https://www.securityindustry.org/report/what-nist-data-shows-about-facial-recognition-and-demographics/
Monnier, J., George, T., Guyard, F., Tarnec, C. and Kountouris, M. “MIFair: A Mutual-Information Framework for Intersectionality and Multiclass Fairness.” arXiv:2604.28030. Submitted 30 April 2026. https://arxiv.org/abs/2604.28030
American Civil Liberties Union. “Williams v. City of Detroit: Face Recognition False Arrest.” https://www.aclu.org/cases/williams-v-city-of-detroit-face-recognition-false-arrest
American Civil Liberties Union. “Civil Rights Advocates Achieve the Nation's Strongest Police Department Policy on Facial Recognition Technology.” https://www.aclu.org/press-releases/civil-rights-advocates-achieve-the-nations-strongest-police-department-policy-on-facial-recognition-technology
The Washington Post. “Arrested by AI: Police ignore standards after facial recognition matches.” January 2025. https://www.washingtonpost.com/business/interactive/2025/police-artificial-intelligence-facial-recognition/
American Civil Liberties Union. “Police Say a Simple Warning Will Prevent Face Recognition Wrongful Arrests. That's Just Not True.” https://www.aclu.org/news/privacy-technology/police-say-a-simple-warning-will-prevent-face-recognition-wrongful-arrests-thats-just-not-true
Liberty. “Legal Challenge: Ed Bridges v South Wales Police.” https://www.libertyhumanrights.org.uk/issue/legal-challenge-ed-bridges-v-south-wales-police/
Ada Lovelace Institute. “Facial recognition technology needs proper regulation, says Court of Appeal.” https://www.adalovelaceinstitute.org/blog/facial-recognition-technology-needs-proper-regulation/
UK Parliament POST. “Facial recognition technology in policing.” https://post.parliament.uk/facial-recognition-technology-in-policing/
Biometric Update. “UK announces largest ever facial recognition rollout as part of policing reforms.” January 2026. https://www.biometricupdate.com/202601/uk-announces-largest-ever-facial-recognition-rollout-as-part-of-policing-reforms
The Register. “High Court approves Met Police's facial recognition after dispute.” 22 April 2026. https://www.theregister.com/2026/04/22/high_court_gives_thumbs_up/
Tim Green
UK-based Systems Theorist & Independent Technology Writer
Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.
His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.
ORCID: 0009-0002-0156-9795
Email: tim@smarterarticles.co.uk
Listen to the free weekly SmarterArticles Podcast
