Startups are paying people for the real-world data needed to train their robots.

This week, an AI training startup called Shift said it would clean New Yorkers’ homes for free. It has plans to expand into other cities as well, including London, and looking around my flat, I get the appeal.

In exchange for the cleaning, Shift wants footage of its cleaners at work: scrubbing dishes, wiping counters, dusting tables, mopping floors. It wants everything. Video of all the boring domestic labor we’d happily outsource if we could — and that robotics companies are racing to teach machines to do so they can sell us something to do it for us.

That’s harder than it sounds. Unlike chatbots, image generators, and other AI tools that have exploded in recent years, robots have to deal with the physical world. That means understanding space, motion, force, friction, weird shapes and materials, awkward lighting, and everything else that humans — and other organics — tend to grasp instinctively. It’s why things that are generally easy for us, like folding clothes, picking up an apple, or pouring a glass of water, have proven so maddening for roboticists to codify.

Teaching machines to do those things takes data. Lots of it. Text, images, and videos could be easily scraped from the internet at an industrial scale. And they were, often without compensating the people who made them. The physical world is harder to scrape, and harder still to scrape quietly without paying for it. This means access to high-quality data is a massive bottleneck for companies developing physical AI. It’s a lucrative opportunity, so companies like Shift are getting creative.

They’re not alone. In India, recent reporting revealed that home services platform Pronto has been using clients’ homes as a source of AI training footage for chores like cooking, cleaning, and laundry. Pronto says it only records footage if customers explicitly opt in — it’s not clear what customers get in return, other than a copy of the footage — but the practice still set off a wave of backlash in the market, with rival startups insisting they have never recorded inside homes to train AI and have no plans to do so.