New data suggests Flock surveillance cameras have an abysmal recognition rate. That's a major problem for drivers.<\/p>\n
Flock Safety can\u2019t seem to stay out of the news recently, from the latest wave of attacks on Flock surveillance cameras to claims that Flock\u2019s gunshot detection microphones are also picking up private home conversations. But one specific report has caught my attention \u2014 a study showing a startlingly high failure rate for the surveillance system AI.<\/p>\n
Following up on complaints about Flock Safety flagging vehicles incorrectly, the police department of Roseville, California, ran its own analysis of the cameras, studying nearly 1,500 Flock reports. They found an astonishing 71% failure rate, where the cameras would mistake one license plate number for another.<\/p>\n
That\u2019s not just a snafu. Mistaking one number for another, like seeing a 7 as a 1, can lead Flock cameras to incorrectly identify an innocent vehicle as one that was previously stolen. That\u2019s exactly what happened in the California situation, where a citizen complained that the cameras were repeatedly flagging his car to police as stolen because the AI technology interpreted a 9 as an 8.<\/p>\n
This is a stark contrast to Flock\u2019s own claims \u2014 sort of. The company reports that independent evaluations show its cameras have a 96% vehicle count accuracy rating, while correctly categorizing 92% into vehicle types. But accurately recognizing vehicles is a far cry from accurately reading license plates, which Flock doesn\u2019t provide much data on. I\u2019ve reached out to Flock to learn more, and I\u2019ll let you know if the company has additional details to provide.<\/p>\n
It\u2019s important to note that this analysis was conducted between 2023 and 2024. Many AI Large Language Models can improve over time with better training and more access to material, which is one reason AI companies are literally devouring old books these days. So Flock\u2019s cameras may have improved in accuracy by now. However, we\u2019ve also seen the opposite happen: AI models can eat up bad data, especially \u201crecursive\u201d data that was created by another AI, and start becoming dumber, less accurate when performing their tasks.<\/p>\n
Flock has said that the camera setup Roseville used wasn\u2019t typical, but didn\u2019t provide more details. It\u2019s possible that where cameras are placed and when they are activated (at night, in the rain, etc.) can have a significant impact on accuracy. So can license plate covers, dirty license plates and unusual license plate placement.<\/p>\n
The problem is these are issues any city may face when adopting Flock cameras. Flock\u2019s primary value offering is that the company\u2019s surveillance is ubiquitous, speedy and accurate, doing a job that police can\u2019t. But if the cameras make mistakes when encountering everyday situations, that\u2019s more work for law enforcement, not less.<\/p>\n
Cities, including Los Angeles, continue to cancel Flock contracts (although sometimes they just switch to another provider), often due to public pressure. If Flock\u2019s core services could have such a high failure rate, that\u2019s another reason for police and sheriff departments to reconsider their use, and could make it difficult for Flock to sell new cities on its technology.<\/p>\n