People who volunteered to test iRobot’s Roomba devices say they agreed to help improve a product, but did not realize images from their homes could be viewed by outside workers. Their complaints have raised questions about what consent means when a home robot records data, and whether a company’s agreements explain what happens after that data is collected.
What testers expected
Greg, an engineer who had spent about a decade beta-testing products, joined an iRobot test in December 2019. He understood that the preproduction Roomba J series device would move around his home, collect data to help improve its artificial intelligence, and give him a chance to provide feedback.
He says he did not understand that images could be shared through a global data supply chain, where contractors outside the United States might see and annotate them. Greg, who asked to be identified only by his first name because he signed a nondisclosure agreement, says iRobot’s handling of the data damaged his trust.
He is not alone. Nearly a dozen people who took part in iRobot data collection between 2019 and 2022 contacted MIT Technology Review after its investigation into images captured inside homes. They questioned whether the company’s practices matched its promises about data protection.
When images reached human annotators
iRobot shared data with data annotation companies, including San Francisco–based Scale AI. Contractors hired by those companies labeled images to help train the robot’s artificial-intelligence algorithms. In one 2020 project, gig workers in Venezuela were asked to label objects in home interiors, with some people’s faces visible.
Workers shared at least 15 images in social media groups, including images of a minor and a woman sitting on the toilet. After MIT Technology Review contacted iRobot about the images last fall, iRobot ended its contract with Scale AI.
iRobot CEO Colin Angle said the images were needed to train object recognition algorithms. He and the company emphasized that the data came from paid data collectors and employees who had signed consent agreements, rather than customers. But testers interviewed by the publication disputed the suggestion that volunteering meant giving up all privacy. Some said the distinction between testers and customers felt dismissive, particularly because at least one tester later bought the device he had tried.
What the agreements did and did not explain
Testers shared several kinds of agreements with MIT Technology Review, including a general Betabound agreement and a “global test agreement for development robots.” The documents said that images, video, text, or audio could be used for purposes including machine learning and object detection. They also said third-party vendors and service providers would be vetted for privacy and data security, bound by confidentiality, and governed by a Data Processing Agreement.
But testers said the documents did not make clear that human data labelers would see and analyze personal images, or explain in detail how data would be handled after collection and with whom it would be shared. iRobot said the broad language covered different testing situations. Its spokesperson also said tests involving images for annotation had specific terms in the agreement for that test.
The company said it notified people whose photos appeared in the set of 15 images. It did not answer whether it would notify everyone else who participated in its data collection, and testers who contacted MIT Technology Review said they had not received a notice.
Consent is more than a signed form
Participants in the tests were recruited through platforms such as Betabound, run by Centercode. The platform listed opportunities to test consumer technology products. Testers generally applied for specific projects and signed agreements before receiving devices. They were not paid through Betabound, but might receive gift cards or devices. For the Roomba tests, participants told MIT Technology Review they received gift cards ranging from $30 to $120 for using the vacuums multiple times a week over multiple weeks.
Privacy lawyers and scholars who reviewed the agreements described the concerns as part of a wider problem with consent. People may not read privacy policies closely, but the testers’ experience suggests that even reading a document may not make a data process understandable. A statement allowing data to be used for machine learning does not, by itself, spell out that workers will inspect images from inside a home.
The dispute leaves a practical question for companies that test connected devices: can participants make an informed choice if they are told what data may be collected but not clearly told who may see it? For testers, the answer matters because a robot’s camera can capture more than the object a company hopes to identify. Their complaints center on whether the agreement made that exposure clear—and what accountability follows when private images are mishandled.