A Laundromat Owner Logged 214 Camera Alerts. Most Were Harmless

The spreadsheet began with an argument near the dryers.

A customer who slept in her car had brought her clothes, bedding and other belongings into the laundromat. She kept the bags beside her while she waited. An automated camera alert marked the area for possible loitering, and the attendant, seeing the notification on a tablet, asked the customer to move along if she was not actively washing.

She pointed to a machine that was still running. The attendant apologized, but the exchange had already drawn attention from other customers. The woman gathered some of her bags and stood outside until the cycle ended.

The laundromat’s owner reviewed the video later. Nothing in it looked dangerous. The customer had spent $18, stayed near her belongings and waited for her laundry, which was what customers were supposed to do.

He opened a spreadsheet on his laptop. The columns recorded the month, the alert category, the staff response and what the footage showed. The first row ended with a plain assessment: paying customer, no safety issue.

For six weeks, he logged every automated warning generated during staffed hours. There were 214.

What the camera treated as suspicious

The owner had installed the system eleven months earlier after repeated thefts from unattended machines and an argument that ended with a broken door. The cameras already recorded the store. The added service analyzed movement and sent alerts when it detected conduct associated with trespassing, prolonged presence or blocked access.

The monthly charge was $179. The owner understood that the tool could call attention to scenes an employee might miss while cleaning filters or helping someone with a payment machine. He did not expect it to decide that a person posed a threat. In practice, the distinction became thin, because an alert arrived with a marked clip and a category suggesting that staff should look.

That prompt changed the room. An attendant who might otherwise have ignored someone sitting quietly now had a reason to approach, and once an employee crossed the laundromat to question a customer, the interaction could become embarrassing or tense even when the employee spoke calmly.

The system did not identify anyone as unhoused. It had no reliable way to know a person’s housing status. It marked visible patterns: staying in one area, returning after stepping outside, lying across chairs or keeping bundled belongings nearby.

Those patterns overlapped with ordinary laundromat use. Wash and dry cycles require waiting. A person washing everything they own may carry more bags than a customer who has a closet at home. Someone without a safe place to leave a blanket will keep it close.

The camera registered the conduct without its context, while the alert category supplied a conclusion before a worker had reviewed the scene.

By the end of the second week, the spreadsheet showed that the same customers appeared repeatedly. Some came to wash clothes. Others bought detergent, charged a phone while waiting or sat near a relative’s machine. Several were known to the attendants, but the system treated each new stretch of inactivity as a new event.

One man was flagged after he finished drying his clothes and waited indoors for a ride. His bags remained against the wall, clear of the aisle. An attendant asked whether he needed help, then asked again after another notification appeared. The man left before his ride arrived.

The owner classified that alert as harmless. He also noted the staff time it consumed, because reviewing clips and making contact pulled attendants away from spills, jammed machines and customers who had requested help.

The safety test

The owner did not assume every alert was wrong. He wanted to know whether the warnings identified situations that staff would otherwise miss, so he compared each clip with employee notes and the full recording rather than relying on the short segment selected by the system.

Of the 214 alerts, 183 showed ordinary customer behavior or people waiting without causing a problem. In 22 cases, the footage did not provide enough context to make a firm judgment. Nine alerts helped staff notice conduct that needed attention, including an obstructed aisle and an escalating dispute near the entrance.

Even that result was difficult to interpret. Employees had already heard the dispute before its alert arrived. The blocked aisle involved laundry carts and bags, but the system highlighted the person standing beside them rather than the obstruction itself. The notifications pointed toward real conditions, although they did not always add useful information or describe the risk accurately.

The spreadsheet also recorded 28 staff approaches prompted mainly by an alert. Eleven ended with a customer leaving before finishing laundry or before a family member returned. None of those eleven encounters involved violence, theft or property damage in the footage the owner reviewed.

Employees carried the immediate risk. A notification did not approach anyone. A worker did, often with limited information and no certainty about whether the person had broken a rule. One attendant told the owner that ignoring the tablet felt unsafe, while acting on it could provoke a confrontation that would not otherwise happen.

That conflict mattered more than the system’s accuracy rate. The service had moved judgment into a sequence: the camera selected a person, the software labeled the scene, and an employee inherited the decision along with the physical task of enforcing it. Management still expected staff to use judgment, but the alert made inaction feel like a choice that might later require an explanation.

The owner found no evidence that the system had been designed to target unhoused people. He also could not determine from the dashboard how the categories had been tested, which examples shaped them or whether performance had been measured in laundromats. The support material described general detection goals. It did not provide enough information to explain why bags and waiting produced so many warnings at his store.

That is the central limit of his test. The spreadsheet measured what happened in one business over six weeks. It did not reveal the model’s training data, establish how the same settings performed elsewhere or prove that every customer interpreted an approach the same way.

It did show a pattern the owner could act on. People carrying most of their belongings were more likely to generate repeated alerts, and those alerts increased their contact with staff even when the full footage showed no safety problem.

Privacy without a clear challenge process

A notice at the entrance told customers that video recording was in use. It did not explain that software analyzed how long they remained in an area or that selected clips could trigger staff intervention.

For customers, the practical issue was not only that a camera captured them. The system converted visible behavior into a suspicion signal, and the people affected had no direct way to see the alert, correct its context or know whether the label remained attached to stored footage.

The owner could review clips through the dashboard. He could not tell customers with confidence how long every derived alert record would remain available under the service’s changing storage settings. He sent questions to the support inbox and received general information about retention and account controls, but not a detailed account of what happened to analytical data after a clip disappeared from his view.

He began mentioning the automated analysis when a customer challenged an approach. That offered some explanation, though it did not undo the encounter. The woman who had waited outside for her laundry returned twice during the test period. Each visit produced another lingering alert.

Her rows in the spreadsheet became a useful check on the system’s claim to novelty. The camera did not learn from the owner’s earlier determination that she was a paying customer whose bags were connected to her housing situation. It observed a similar pattern and raised the same concern again.

After six weeks, the owner disabled the lingering category during staffed hours. He kept ordinary recording and a narrower set of alerts tied to access points when the laundromat was closed. Staff could still approach someone whose conduct created a specific problem, but sitting with bags was no longer enough to place a prompt on the tablet.

The change did not settle every safety concern. Employees still had to judge disputes, blocked walkways and people seeking shelter during bad weather. The owner also kept paying for parts of a system whose workings he could not fully inspect.

He continued the spreadsheet for another month. The column for alert-driven approaches fell to zero, while staff recorded no increase in theft or property damage during that period. The result was limited, but it was more concrete than the category on the dashboard.

Questions people ask

Can a security camera mistake homelessness for suspicious behavior?

A camera does not need to identify housing status to affect unhoused people unevenly. In this laundromat, the system treated lingering and bundled belongings as risk signals. Those behaviors were more common among customers who lacked a safe place to wait or store their possessions, so they received more scrutiny.

Do automated camera alerts make workers safer?

Some alerts can direct attention to a real hazard, but this owner’s log showed that most did not add useful safety information. Workers also had to carry out the response. In 28 alert-driven approaches, employees entered encounters that sometimes became tense even though the reviewed footage showed no theft, violence or damage.

Are customers told when video is being analyzed?

The entrance notice disclosed video recording, but it did not explain that software assessed behavior and generated suspicion alerts. The owner could describe the system after a confrontation, yet customers had no direct view of the clips or labels and no clear process for adding context to an alert.

What changed after the owner turned off lingering alerts?

During the following month, staff recorded no alert-driven approaches and no increase in theft or property damage. The owner kept recording video and retained narrower closed-hours warnings. On the spreadsheet, the woman who slept in her car appeared twice more as a paying customer, with no new alert beside either visit.