TL;DR
Sainsbury’s has switched off live facial recognition at its East Dulwich superstore after Matt Arnold, 46, was stopped at a self-service till and told to leave on the strength of a camera alert. The retailer and its supplier Facewatch both say the software worked correctly and staff mishandled the alert. This is the second such case at a Sainsbury’s branch in under a year.
What happened at the till
Arnold, a comedy promoter buying supplies for a gig at the neighbouring football ground, had already scanned his shopping and his Nectar card and was waiting for an attendant to clear an alcohol purchase. Two managers approached instead and refused to serve him, referring to an incident earlier that week. Walking out, he noticed a monitor above the entrance ringing his face in red.
Head office rang the next day to apologise. Sainsbury’s puts the fault on staff, not the system, and quotes a 99.98% accuracy rate, with a trained manager checking every match. Facewatch maintains the alert itself was accurate and only its handling went wrong, and that pausing the branch is precautionary while staff are retrained.
Not the first time
Elephant and Castle, September: the same software flagged Warren Rajah, who was made to leave — and Facewatch later confirmed he had never been on its database. That case, too, was put down to human error. Other retailers using the supplier have seen comparable incidents, at a Home Bargains branch in 2024 and, last year, a B&M in Cardiff.
Arnold’s arithmetic is the part retailers should sit with. A 0.02% margin sounds negligible until it is multiplied by the millions of weekly visits a national grocer handles. Silkie Carlo of Big Brother Watch argues errors are unavoidable when identity checks run indiscriminately at that volume.
The case for keeping it
Sainsbury’s is not backing down on the wider deployment. It points to British Retail Consortium counts of 1,600 abusive or violent incidents against shop staff every day, up from a 2019-20 daily figure of 455. Its two-store trial cut logged theft and anti-social behaviour by 46%, with more than nine in ten flagged offenders not coming back.
Looking forward
The repeated defence — the software was right, the humans were wrong — is doing a lot of work here. For any UK business deploying automated alerting, it points at the real exposure: the decision rests with a staff member under pressure who is not equipped to overrule the machine. Arnold’s question stands unanswered. If the technology is safe enough for every other branch, why pause it in this one?