How does Amazon use feedback from video reviewers to train algorithms?

File photo of people commuting to work at an Amazon distribution center in the Staten Island borough of New York Photo credit: AP

(This article is part of Today’s Cash, The Hindu’s newsletter on emerging topics at the intersection of technology, innovation and policy. Subscribe to get it delivered to your inbox. Here,

Amazon is using offshore workers based in India and Costa Rica to review camera footage from its warehouses to train its surveillance algorithms. ledge shared.

Feedback from video reviewers is used by Amazon to train its AI camera system that monitors the stowing process in its warehouses. The camera also reportedly tracks the stover’s movements and sends footage to video reviewers whose input helps the company improve its machine learning tools.

Published in partnership with bureau of investigative journalismEmployees reviewing footage from warehouses are paid just hundreds of dollars a month for eight-hour shifts that can extend to 11 hours during busy periods, the report said. During the shift, reviewers can view 8,000 videos every day.

Reviewers who worked on monitoring the warehouse complained of physical problems including headaches, eye pain and loss of eyesight. They also complained that they were made to meet difficult goals, with tracking software logging periods of inactivity outside their designated break times.

These reviewers view footage from cameras positioned in warehouses that monitor the activities of stowers.

Video reviewers shared that their primary role was stock management, however, they could also record errors made by their colleagues overseas.

Managers also track reviewers’ performance with real-time analytics and maintain an accuracy rate of 95-99%.

Reviewers who are unable to maintain high accuracy at a fast pace risk losing their jobs.

According to a document shared by a former reviewer, reviewers are rated based on their “TAKT” time, highlighted in red is the average time it took those ranked videos in the last four spots to get through.

The report also shared that most of the workers are being filmed in warehouses by amazon cameras were unaware that they were being watched and that their footage could be sent to manual reviewers.

Amazon on its own Its system algorithmic monitoring stove is said to be 95% accurate and only the remainder of it requires manual checking. It also said that many of the allegations made in the report were based on anecdotes from a handful of individuals and painted a misleading picture that does not represent the majority of its team.

The company also shared that reviewers in India and Costa Rica were encouraged by the software they use to take short breaks during their shifts.

The report is based on interviews with 33 current and former Amazon employees, including 21 video reviewers.