AI's New Trade: Free House Cleaning for Surveillance Data to Train Robots

A German startup called Shift has begun offering something that sounds too good to be true: free professional house cleaning in Manhattan apartments. Two hours of deep cleaning, no charge. The catch arrives quietly, in the form of a camera-equipped cap that the worker wears throughout the job, recording every movement—how they navigate around furniture, organize a nightstand, scrub under a bed. The footage becomes the real product.

Shift, owned by Munich-based MicroAGI, has discovered a market arbitrage that turns the economics of domestic labor inside out. While the company absorbs the full cost of the cleaner's wages and materials, the video footage from those two hours is worth far more in the artificial intelligence market than the going rate for apartment cleaning in New York. Hardware manufacturers building humanoid robots will pay substantial sums for real-world footage of humans solving unpredictable problems in uncontrolled environments. They need to see how someone responds to a sink full of dishes, how they fold a fitted sheet, what angle they use to scrub under a sofa. That raw material becomes training data for physical AI systems—the kind of systems that might one day clean apartments without human intervention.

The New York experiment has proven wildly popular. According to CEO Harry Killberg, the service launch caused the platform's infrastructure to collapse under thousands of reservations. Shift plans to expand to London, Zurich, and other cities soon. But the free cleaning offer is only the visible edge of a much larger data collection operation. On the company's website, Shift advertises payments of twenty dollars per hour to a global network of contributors who film themselves during daily activities. In the first quarter of 2026 alone, the startup spent more than five million dollars on these payments.

Shift is not alone in this pursuit. DoorDash, the American delivery giant, recently launched Task, a separate app that pays its couriers not to deliver food but to record themselves performing routine household tasks. Workers strap smartphones or body cameras to their chests and film their hands while cleaning, entering buildings, or having spontaneous conversations in multiple languages. The complexity of the task determines the compensation. The movement sequences, complete with their real-world imperfections, are then sold to companies developing AI and robotics systems. The practice borders on the dystopian: workers become unwitting data factories, their labor doubly exploited—once for the task itself, and again as training material for the machines that might replace them.

This model of mass data collection through human activity is not entirely new. In 2016, Pokémon Go became a global phenomenon, and millions of players believed they were hunting virtual creatures. What they were actually doing, without realizing it, was participating in a distributed mapping project. Each time a user scanned a statue, plaza, or building facade to earn in-game rewards, they contributed to a three-dimensional representation of the real world. Over years, Niantic accumulated tens of billions of images of cities across the planet, courtesy of an army of volunteers who never thought of themselves as such. That mountain of data became more valuable than the game itself. Today, parts of it power visual positioning systems that can locate a device or robot with centimeter-level precision—far better than conventional GPS. Robotics companies now use this technology to help their machines navigate streets and sidewalks using, in a sense, the eyes of Pokémon Go players.

If Pokémon Go turned millions of people into accidental cartographers, the current fever for physical AI aims to transform them into robot teachers. The new datasets are no longer extracted from the internet, where publicly available text and images are beginning to deplete. Instead, they come from real human activity in the physical world. You might be the customer receiving free cleaning or the courier seeking extra income. Either way, you are teaching a machine how to do your job.

The concerns raised about Shift and DoorDash Task center partly on privacy, but José Luis Calvo, an AI expert and founder of startup Diverger.AI, identifies a different problem. He sees no privacy threat as long as companies honor their contracts and do not create social profiles or enable social scoring—practices prohibited in the European Union. The real issue, he argues, is that companies are hiring workers to record their labor specifically to automate and replace that labor. It is similar to the film industry recording actors to generate artificial extras, except vastly larger in scale. Meta has proposed recording all employee computer interactions to train AI systems. Uber collects video of driver routes. In China, provincial governments have built data factories where human workers repeat the same physical tasks over and over while machines observe and learn to replicate them.

The legal and regulatory response is beginning to bite. DoorDash Task has encountered unexpected resistance from labor law. The app is expanding into selected markets but has had to temporarily exclude states and cities with strict worker compensation and independent contractor classification laws—California, New York, Seattle, Colorado. Whether such work should carry premium pay or even recurring royalties remains an open question. The industry, however, has already decided this is no marginal bet. Nvidia CEO Jensen Huang recently stated bluntly that the ChatGPT moment for physical AI is approaching. Behind that claim lies a thesis: physical AI applied to robots, factories, logistics, and homes represents a potential market of up to fifty trillion dollars within years. The capital is already flowing. Jeff Bezos has backed Project Prometheus, aimed at building AI systems that understand and simulate the physical world. Travis Kalanick, Uber's former CEO, has returned with Atoms, focused on robotics for mining, transport, and food production. Around them, a new generation of startups—Physical Intelligence, Skild AI, Figure AI—are raising multimillion-dollar rounds with only years of existence and minimal revenue. The race is on, and the data collection infrastructure is being built now.