Forgot your password?
typodupeerror
Privacy United Kingdom

Sainsbury's Store Pauses Facial Recognition After False Shoplifting Claim (theguardian.com) 121

Bruce66423 shares a report from The Guardian: Sainsbury's has paused the use of AI face scanning in one of its stores after a customer was wrongly identified as a shoplifter and ejected from the shop. "I was embarrassed, mortified even, and felt quite humiliated and powerless," Matt Arnold, 46, said of his ordeal. The comedy promoter was buying supplies in the store in East Dulwich, in south-east London, for a standup event at Dulwich Hamlet football club when, after scanning his items and a Nectar card, he was approached by two managers who told him he could not be served owing to an earlier incident. He was then asked to leave and they tried to escort him from the store.

As he left, he saw an overhead CCTV monitor alert with a red circle surrounding his face. He asked the shop staff to keep his shopping in the trolley so his friend could come and pick up the supplies for the comedy night happening soon next door. "I think they were quite confused by this, understandably, but agreed and my colleague Dave went in to pay for and pick up the shop about five minutes later. There was no pause for thought from the staff, no suggestion that they understood this is not how a shoplifter would behave. Just blindly following the machine's orders." Sainsbury's head office apologised to Arnold the next day and has paused use of its AI-assisted Facewatch technology in the store while an investigation takes place.
Arnold says the facial recognition tech should be paused in all stores. "Anyone could be falsely accused and at some point that will be someone vulnerable, someone with mental health issues like anxiety. It's inevitable," said Arnold. "Also, I would worry about the confidence-destroying effect of it happening to a younger person or someone less willing or able to stand up for themselves as I have done."

A Sainsbury's spokesperson said: "We have contacted Mr Arnold to apologise for his experience at our Dulwich superstore. The incident was caused by human error, not the facial recognition technology. Customers can be reassured that the Facewatch system has a 99.98% accuracy rate, and every match is reviewed by a trained manager." A Facewatch spokesperson said their technology was not at fault in this case. "A correct alert was sent to the retailer, but was subsequently subject to human error in the way it was handled in store," they said.

Sainsbury's Store Pauses Facial Recognition After False Shoplifting Claim

Comments Filter:
  • by Anonymous Coward

    He needs to be like an American and sue them for slander and defamation, the only way a private business will learn the downsides of private law enforcement is if you cost them millions of quid.

    • I don't understand why taking someone to court is considered "bad" or "American". That's a very civilized way to settle a dispute.
      • American and British laws are different. Most people consider it easier to win a defamation case in Britain, although the plaintiff does need to prove "actual harm" to their reputation. British law does place a high value on one's reputation.
        • Notably, fact is not an absolute defense like it is here. If your intent is to do harm then it's illegal, even if the harm is to a nonce like Saville, which is why it took so long to get him

    • The business has the right to refuse service for most reasons that do not tickle a lawyer's foot.

      This customer started squeaking and got the grease - most will just silently leave and move on.
  • 99.98% (Score:5, Insightful)

    by Rei ( 128717 ) on Tuesday August 18, 2026 @07:16AM (#66294340) Homepage

    the Facewatch system has a 99.98% accuracy rate,

    So one in every five thousand people who walk into your stores is going to be falsely accused of a crime? Wow, thanks for making abundantly clear that people should NOT shop there...

    • It's penalising the customers only.

    • Re:99.98% (Score:5, Informative)

      by Mushur ( 870120 ) on Tuesday August 18, 2026 @07:50AM (#66294378) Journal

      It really means nothing. What does the 99.98% "accuracy" even refer to? For all we know, it could be the true negative rate (e.g. it correctly detects who is NOT a shoplifter 99.98% of the time).

      At minimum, you need a "confusion matrix", you need False Positive, True Positive, False Negative, True Negative, to have a minimum of information to evaluate any prediction system.

      • by Anonymous Coward
        These systems are banned in Germany for this very reason.
      • It's laboratory tests in perfect condition with a tailor made data set designed to make the number look better. In other words it's a lie. It's not even lying with statistics it's just a lie.
    • A short course in Bayes' theorem should be mandatory for anyone making that kind of "success rate" claim.

      • by pesho ( 843750 )
        Ah, the priors strike again! There should be a mandate that every accuracy claim to comes with the expected positive predictive value and the number of expected false positives.
    • the Facewatch system has a 99.98% accuracy rate,

      So one in every five thousand people who walk into your stores is going to be falsely accused of a crime? Wow, thanks for making abundantly clear that people should NOT shop there...

      More interesting is if 1 in five thousand is also a shoplifter, that means 1 fals positve and one real one, so a 50/50 chance of being wrong when an identificattion is made.

    • and

      "A correct alert was sent to the retailer, but was subsequently subject to human error in the way it was handled in store,"

      So was the guy a shoplifter or not? Because apparently not and the human error was acting on the alert it seems. Sounds like the front fell off.

      • Re: (Score:2, Insightful)

        by Anonymous Coward

        So was the guy a shoplifter or not?

        Something not yet mentioned is that face recognition can't even tell whether someone "is a shoplifter" NOW. At best, it can only identify that someone is on a list. A list of what? People who have been convicted of shoplifting? Or accused? Or just a list of people the shop-keeper doesn't like? Like the infamous terrorist watch list or license-plate readers, who's on the list, how do people get onto the list, how do they get off the list, who makes the list, and how many people are on the list incorrectly?

      • by cusco ( 717999 )

        and

        "A correct alert was sent to the retailer, but was subsequently subject to human error in the way it was handled in store,"

        So was the guy a shoplifter or not? Because apparently not and the human error was acting on the alert it seems. Sounds like the front fell off.

        Facial recognition software doesn't actually "recognize" your face, it selects a series of points the algorithm thinks important in a pattern and matches them to other patterns in the database. It's the same technology, though less advanced, than fingerprint readers. Until better filters for the patterns were developed there were a lot of false positives in the fingerprint readers as well, and I suspect this will be the case with facial recognition software as well.

        • Re:99.98% (Score:5, Interesting)

          by dgatwood ( 11270 ) on Tuesday August 18, 2026 @04:35PM (#66295404) Homepage Journal

          and

          "A correct alert was sent to the retailer, but was subsequently subject to human error in the way it was handled in store,"

          So was the guy a shoplifter or not? Because apparently not and the human error was acting on the alert it seems. Sounds like the front fell off.

          Facial recognition software doesn't actually "recognize" your face, it selects a series of points the algorithm thinks important in a pattern and matches them to other patterns in the database. It's the same technology, though less advanced, than fingerprint readers. Until better filters for the patterns were developed there were a lot of false positives in the fingerprint readers as well, and I suspect this will be the case with facial recognition software as well.

          Faces are way less unique than fingerprints unless you are using a camera with such high resolution that you can see the individual hairs, pimples, birthmarks, moles, freckles, and other tiny details, and even then, only if you are analyzing it at that level of detail after using fixed points to unwrap the image into a flat UV map for comparison.

          Statistically, no two people have the same fingerprint, so it's just a matter of comparing enough points to achieve the desired confidence level. At a shape/structural level, lots of pairs of unrelated people [wikipedia.org] have faces that are too much alike to reliably detect the difference, likely because there's not as much genetic diversity as one might assume [sciencedirect.com].

          Compounding this is the fact that a fingerprint is detected by being pressed flat against something, whereas face shape has to be recognized from arbitrary angles.

          Any approach that does not stitch together multiple camera angles to get a complete image of the entire face and then unwrap it into a high-res UV map and then use the result as a secondary filter pass after structural comparison is doomed to fail frequently when presented with an adequately large number of people. There are simply too many people that look far too much alike even within a fairly small geographical area like a state or small country, much less the entire world.

          Accuracy is the wrong number to give. Accuracy is right answers over wrong answers, without any info about whether they were false positives or false negatives. What you really want is a relationship between the number of false positives and by total positives, or perhaps the relationship between false positives and total negatives. The latter is called specificity (true_negatives/ (true_negatives + false_positives)), and gives you an indication of how likely people are to be falsely accused. This needs to be incredibly low.

          But the better number would be false_positives / (false_positives + true_positives), because that would tell you how likely it is that someone flagged is really the person in question.

          Either way, for this purpose, the only rational way to use the tech is to watch flagged people more carefully. Trying to escort them out of the store is going to be a disaster of false positives, for all the reasons previously mentioned (look-alikes, limitations caused by camera perspective, etc.).

      • and

        "A correct alert was sent to the retailer, but was subsequently subject to human error in the way it was handled in store,"

        So was the guy a shoplifter or not? Because apparently not and the human error was acting on the alert it seems. Sounds like the front fell off.

        Exactly. You don't send an "alert" for a non-event. If the guy wasn't flagged, why was an alert sent? Doesn't sound correct to me. If the software just notes possible matches, subject to human review then say that, but it sounds like it's trying to do more or store staff is relying on it rather than thinking for themselves.

        • by djinn6 ( 1868030 )

          It's the store's responsibility to make sure their employees know how to use the information provided and not simply trust everything they're told by the software.

          This is the same cause of failure as when SWAT teams go on shooting sprees in random people's houses without even checking the veracity of the tip that they received.

    • by MeNeXT ( 200840 )

      the Facewatch system has a 99.98% accuracy rate,

      So one in every five thousand people who walk into your stores is going to be falsely accused of a crime? Wow, thanks for making abundantly clear that people should NOT shop there...
      Flag as Inappropriate

      My math may be wrong but that would be 2 out of every thousand nt five thousand.

    • by ccr ( 168366 )

      Even if we discount various other confounding factors, and assume in good faith that the 99.98% "accuracy" is somehow meaningful on individual recognition basis one should consider this:

      Sainsbury's stores serve millions of customers per week. As per usual, they do not seem to be publishing any current customer amounts themselves, but https://rspo.org/members/3-001... [rspo.org] claims "Over 16 million customers visit 740 Sainsbury's stores each week", and that was nearly 20 years ago.

      Assuming that figure has any truth

      • I bet management can't do the math. They only read the 99.98% accuracy promise on the supplier powerpoint and drooled themselves.

      • by cusco ( 717999 )

        That's why the people selected by the software are **supposed** to be checked against the image in the database. Apparently in this case they didn't bother with that step.

    • by AmiMoJo ( 196126 )

      It's potentially a big GDPR issue as well. If they shared the face data with anyone else (these shops usually share the data) then they need to tell him who and compensate him for now having to get all of the others to fix the mistake too.

      Otherwise the next time he walks into some other shop using the same data, he will get accosted again.

    • Re:99.98% (Score:5, Informative)

      by rsilvergun ( 571051 ) on Tuesday August 18, 2026 @09:51AM (#66294556)
      It's also an outright lie. Ignoring the fact that facial recognition has a hard time with black and brown faces the figure cited is based on laboratory tests in perfect conditions.

      Once again I am so tired of being lied to and in such an obvious fashion.
    • Re:99.98% (Score:4, Informative)

      by pla ( 258480 ) on Tuesday August 18, 2026 @12:25PM (#66294878) Journal
      To put that in real-world perspective, the average daily transaction volume for a mid-sized US grocery store (I realize this isn't the US but don't have similar data for the UK) is roughly 13,500.

      This system will therefore flag, on average, 2.7 potentially life-destroying false accusations of shoplifting per day per store.

      And spare us the "human review" defense, anyone who has ever worked in retail knows that you either obey the computer or find a new job. There is zero tolerance for employee discretion when it comes to safety, financial, or legal issues.
    • They have about 20M customers a week across the UK. A "99.98% accuracy rate" might mean 0.02% false positives, with 4000 false positives a week. But, to make some numbers up, it might mean of the 10,000 people identified as know shoplifters each week, only 2 are wrongly identifed by the system. It's then up to staff to make a judgement call, and perhaps they mostly get it right. Who knows what they mean? Statistics is hard.
    • You did the math to remind us what that percentage actually means to real people. Thanks for reminding us to do that.
    • Yeah well it's not your fucking local supermarket.

      Unfortunately it is mine. There are local smaller shops I also go to but it's proven pretty handy you know?

  • Incidents of AI [machines][computers] reigning superior over Humans is on-the-rise and out-of-control. Every power-of-authority is eager to use technology like Flock cameras and facial recognition, as a club to de-humanize Humanity. As pointed-out in countless books and movies, when the foundation of civilized society is stripped-away, no civilized society remains.
    • Incidents of AI [machines][computers] reigning superior over Humans is on-the-rise and out-of-control. Every power-of-authority is eager to use technology like Flock cameras and facial recognition, as a club to de-humanize Humanity. As pointed-out in countless books and movies, when the foundation of civilized society is stripped-away, no civilized society remains.

      We lost the clanker war one little step at a time. From "remember my password" features in browsers to carrying our little life-controllers under the label "organizer" in our pockets to letting the machines dictate behavior for us. These types of stories are just one more example of how we've already given up, and consider the machines superior. It's also why I've been so frustrated over the last decade and some change by every self-driving story being filled with comments from people convinced the machines

  • Investigation... (Score:4, Insightful)

    by PsychoSlashDot ( 207849 ) on Tuesday August 18, 2026 @08:35AM (#66294434)
    "Sainsbury's head office apologised to Arnold the next day and has paused use of its AI-assisted Facewatch technology in the store while an investigation takes place."

    What investigation? It has a 0.02% failure rate, which is very, very high. End of investigation. Discard the tech.

    "The incident was caused by human error, not the facial recognition technology. Customers can be reassured that the Facewatch system has a 99.98% accuracy rate, and every match is reviewed by a trained manager."

    Every match may or may not be reviewed by a trained manager, but customers cannot be assured of anything except that they - per factual evidence - be wrongfully accused of wrongdoing. The incident would not have occurred had the facial recognition technology incorrectly targeted the wrong individual. The technology is the only place where blame should be assigned except for the executives who authorized its use and continue to make excuses for it.

    "A Facewatch spokesperson said their technology was not at fault in this case. 'A correct alert was sent to the retailer, but was subsequently subject to human error in the way it was handled in store,' they said."

    Facewatch is incorrect. Again. An incorrect alert was sent to the retailer... unless they are insisting Matt Arnold is in fact a shoplifter. There is no human error if there is no incorrect alert.

    I do understand that lacking this technology, we would be solely relying on humans to notice things like repeat offenders. Same for automated license plate readers... we'd be solely relying on humans to spot, recognize, and act on suspect vehicles. But the sheer bulk of hits makes all the difference in the world. Humans will rely on the machine's judgement and the onus then shifts to the target to try to override that. If a store manager just randomly mistakes a legitimate customer for a shoplifter, it's much easier to convince them they've made a mistake.
    • by HiThere ( 15173 )

      I'm sure that technically they said something like "The man may be a shoplifter", and thus were correct in that sense. But using a system that does that at that error rate is NOT reasonable. This is like the automated driving systems that require a human driver to take over within 2 seconds. NOT a good idea.

    • Notice how they blamed a human and defended the machine. Sounds like the facial recognition company drafted the response.
    • Discard the tech.

      Dumb answer. Perfect shouldn't be the enemy of good enough, and a 0.02% failure rate is good enough *provided the application of additional checks and balances makes that failure rate have no consequence to the end users*

      There's no such thing as perfect.

      • Discard the tech.

        Dumb answer. Perfect shouldn't be the enemy of good enough, and a 0.02% failure rate is good enough *provided the application of additional checks and balances makes that failure rate have no consequence to the end users*

        There's no such thing as perfect.

        I never demanded perfection.

        Worse, your qualification is meaningless. A 100% failure rate would be good enough provided additional measures that compensate for it.

        This thing screws up one in every 5,000 visitors. Look up Sainsbury's. Three million customers per day. That's 600 failures per day. With around 1,500 locations, that's an average failure at every location every two to three days. Maybe that's happening when your best manager is on-hand to take a look and conclude "huh, well, this looks a

  • by Cpt_Kirks ( 37296 ) on Tuesday August 18, 2026 @08:40AM (#66294442)

    The dumb sonofabitch WOULD NOT back down.

    His system had obviously accused the wrong man, but the machine "doesn't make mistakes".

    The current state of "AI" is primitive and terrible.

    They lie, they cheat, they take the easy way out, even if it's wrong.

    People are going to be killed by shitty, fake "AI".

  • by TigerPlish ( 174064 ) on Tuesday August 18, 2026 @08:58AM (#66294464)

    Do this to me, and I'll sue. I don't care for what -- the usual bullshit of "emotional damage" and "mental anguish" and all that.

    I encourage every person erroneously flagged to do so. The only thing that hurts corporations is a hit on the wallet -- so hit their wallet, hard, and often.

    • I will destroy your store on the way out. Every single shelf will be emptied onto the floor. Every display toppled.

    • Do this to me, and I'll sue. I don't care for what -- the usual bullshit of "emotional damage" and "mental anguish" and all that.

      This is the second-highest level of self-unawareness demonstrated on Slashdot after all the maggots claiming Democrats rape kids.

      You'll sue, you don't know for what, but emotional damage and mental anguish are bullshit? You absolute clown.

    • by cusco ( 717999 )

      You must have lawyers in your family.

    • The wallet to use is your own, ever hear of voting with your wallet? Don't do business with them, of course the problem is most couldn't be bothered to do that as it is inconvenient.

      This guy even as being escorted out of the business was saying (paraphrased) 'leave my items there so my buddy can come in and give you money to reward this conduct.'
    • Do this to me, and I'll sue. I don't care for what -- the usual bullshit of "emotional damage" and "mental anguish" and all that.

      Sainsbury is not an American store. It's not covered by American snowflake laws, and your emotional damage will likely get you laughed out of a court.

  • > Customers can be reassured that the Facewatch system has a 99.98% accuracy rate,

    I'm sure Matt Arnold is extremely reassured.

  • Retailers and shop managers ARE NOT law enforcement. If they SUSPECT someone is a criminal, they should notify the police with the appropriate EVIDENCE.

    • I think they do have some shopkeeper privileges to detain but they rest on having reasonable grounds (ie. they saw you slip something into your pocket)

      Asking this guy to leave I presume must have been on other grounds though, pehaps some fine print you walk past as you enter the store saying they reserve the right to refuse service to anyone for any reason.

      • In UK, if I'm correct, law allows shopkeepers from refusing service only for security reasons.

        https://www.milnerslaw.co.uk/c... [milnerslaw.co.uk]

        Still, I believe the appropriate thing to do in case they suspect of a crime, is to notify the police and provide evidence, so it's investigated.

        If they're allowed to kick people because some obscure software misidentified them as a criminal, I can envision a future where innocent people are unjustly and completely ostracized from all commerce within a city or country.

    • They have a right to trespass anyone they don't like from the store. Apparently "the AI told me some nasty rumors about you" is adequate grounds for banning someone for life
      • They have a right to trespass anyone they don't like from the store.

        No, they haven't.

        https://www.milnerslaw.co.uk/c... [milnerslaw.co.uk]

        • I stand corrected. Do be more precise, they have a right to refuse service for any reason that isn't discriminatory. I still think "You're not allowed here because I don't like you, but it's not because you're black" constitutes a legit reason to trespass someone.
          • The point of the law is the onus is on you to provide evidence that your discrimation was not due to a protected aspect (black) but something else (inappropriate attire). The victim doesn't have to prove anything. Thus the massive grey area between these two attitudes gives the benefit of doubt to the victim.

    • The problem with this argument, from the shop's point of view, is that the UK is in the middle of a shoplifting epidemic. The policing resources simply aren't there to deal with small scale (or even flagrant and much larger scale) shoplifting. It now seems to be rare for police officers to attend reports of shoplifting at all. And so effectively shoplifting becomes legalised, which is obviously hurting the big chain stores financially and potentially devastating for smaller shops.

      Shops are trying to find so

      • But I do have some sympathy with the shops who are looking for something -- anything -- to reduce the threat they're facing because the rule of law has failed in retail settings.

        And it's going to continue to fail as long as capitalism celebrates the wealth inequality that drives the theft.

  • "A correct alert was sent to the retailer, but was subsequently subject to human error in the way it was handled in store," A "correct alert" would say that he was not a match. Are they claiming that they send alerts to the store for every non-match they find?
    • The alert said someone who kinda looks like that was a shoplifter. It's the store employee's job to verify a match which they didn't do. The tech is still bad but it didn't say he was a shoplifter, it said he might be and a human fucked it up the other half of the way.

      • by sjames ( 1099 )

        But every person who sets foot in that store probably "kinda" looks like someone who has shoplifted. And since shoplifters look like everybody else, everyone COULD be a shoplifter.

        • Sure, but they also have the actual photos of the actual shoplifters to compare to, so the human clearly fucked up here.

          • by sjames ( 1099 )

            BOTH screwed up. The system shouldn't have made the match and the human should have done a better job seeing that the match was wrong. I'll go one larger and say the people who created the procedures screwed up because they didn't make allowances for the possibility of the first two screw-ups.

    • I don't know, but that line bothered me as well. Too broad. It could mean that the system flagged him as "possible shoplifter, so keep an eye on him", and they misinterpreted that as "get him!". Or, "you weren't supposed to put his picture up".
  • EVERYONE looks the same at about 50x50 pixels on a bubbled lens using a 10 year old camera. You really think to stop shoplifting, they installed an isolated network that can handle barely compressed 8K footage then put in all new, top res cameras? No! Those are $200 garbage quality cams mounted in the ceiling with warped lenses. Gee, I wonder why it can't tell people apart from that distance.

    Have you seen the view of yourself that they broadcast at an American HyVee? It looks like a damn 90's skateboardin
  • Facial recognition is really neat technology, and we are all slightly better off thanks to its development. It's also imperfect, and nearly-guaranteed to be a net negative if you fully trust it, without human verification.

    If you want to come out ahead by using cool technology, you're going to have to think about how to use it. There just isn't any shortcut around thinking.

    If you have something which is only 99.9% reliable, that can be useful but it's going to be wrong sometimes. How are you preventing that

    • by sjames ( 1099 )

      The problem is, people have been over-trusting technology since technology was invented. Consider the '70s when corporate representatives were so fond on the phrase "But the computer says...". Yes, you could hear the bold italics in their voice, as if it was the infallible final word on everything.

      If THE COMPUTER says you shoplifted, rest assured, some dolt in a suit will treat that like God's own word booming from the clouds.

  • A Sainsbury's spokesperson said: "We have contacted Mr Arnold to apologise for his experience at our Dulwich superstore. The incident was caused by human error, not the facial recognition technology. Customers can be reassured that the Facewatch system has a 99.98% accuracy rate, and every match is reviewed by a trained manager." A Facewatch spokesperson said their technology was not at fault in this case. "A correct alert was sent to the retailer, but was subsequently subject to human error in the way it was handled in store," they said.

    That's frankly bullshit. The system put a red ring around his face on the security monitoring system. The system told the employees to treat him as a shoplifter and as dutiful employees of their company with the system their company told them to use, they complied.

    This is almost the exact same excuse made in the movie Brazil by Michael Palin's character Jack Lint, "Information Transit got the wrong man. I got the *right* man. The wrong one was delivered to me as the right man, I accepted him on good fai

  • "I was embarrassed, mortified even"

    Was this Snagglepuss? Did he exit the store stage left?

  • I haven't shopped at Sainsbury's for over a decade. The last time, I wanted to buy a pair of jeans (having just ripped the arse out of my own) but they had no fitting room because "people can't be trusted". The only option was to buy them and then try them on in the toilets. Said toilets were beyond squalid, and I asked - politely, with a smile - whether I could speak to a manager. What I got instead was two security guards, who made it clear that there was no room for reasonable discussion. So throwing fal

  • Customers can be reassured that the Facewatch system has a 99.98% accuracy rate, and every match is reviewed by a trained manager.

    A 99.98% accuracy rate will be wrong 1 out of 5,000 times. That's perhaps an acceptable Russian roulette probability for an individual, but for the store, that means hundreds (thousands?) of incorrectly identified people daily. If a person triggers the algorithm, what is the trained manager supposed to do? Is there a computer photo that the person can be compared against, or does the manager go with his Spidersense?

    • That means 1 in every 5000 customers gets flagged as a criminal.

      Sainsburys gets 3.7 million transaction per day, that 740 people a day that will get falsely accused.

      I don't think that many that are getting falsely accused now.

  • Their penalty for shoplifters is that they won't take their money?

    I don't think they have thought this one through.

  • How many interactions are used to calculate that number? 99.98% would mean 1 in 5 000 are false matches. Given this is an AI system that has to work with live data, and a complex live system, getting false matches to 1 in 5 000 sounds remarkable. To get an accurate reading, you would need a population of 500 000 (100x), and I would want a 5 000 000 point sample, since that would 1000x the base rate, which at that point would be large enough I'd trust the 99.98%.

    To put 99.98% in prospective, giving the av
  • Especially when a wrong identification can have real consequences. I think this should come with some real consequences for the idiots that approved this system.

The brain is a wonderful organ; it starts working the moment you get up in the morning, and does not stop until you get to work.

Working...