The War Against Fake Data
Physical-AI companies pay hundreds of thousands of people to film themselves working, and every payment is an incentive to fake it. Saurab Dhir built the systems that decide which recordings are real.
The iron was never plugged in. In footage sent to Mecka AI, a company that pays people to film themselves doing household and industrial work, a contributor ran an iron back and forth across a shirt. A machine catches what a human eye might miss: the wrinkles never move. The same story shows up in vacuuming clips, a person pushing a dead machine across a floor that stays as dirty as it started. “With a turned off vacuum, the mess doesn’t move,” said Saurab Dhir, the engineer whose detection systems caught the pattern. The surface being worked on has to change, and in a fake it never does.
The unplugged iron is a small crime with large stakes. Physical AI companies have quietly built one of the stranger labor markets online: hundreds of thousands of contributors across 12 countries, paid to record themselves cooking, cleaning, assembling, and repairing so robots can learn from the footage. Every payment is an incentive, and every incentive gets tested. The marketplaces that came before, rideshare, e-commerce, ad networks, learned the same lesson on the same schedule: pay people at scale, and a fraction will probe the system for the least effort that still gets paid.
Saurab Dhir sits at the checkpoint where those probes are meant to die. A co-founder of Docula, the medical-AI startup Mecka acquired this year, he leads the machine-learning side of the company’s data-integrity work. He researches the detection techniques, collects and cleans the data, and designs the model architecture; the infrastructure engineers take over at pipelining and scaling the work to hundreds of thousands of hours. When a contributor irons with a cold iron somewhere in those 12 countries, the catch traces back, sooner or later, to something Dhir built.
The recordings are almost aggressively mundane, kitchens and laundry rooms, workbenches and shop floors, ordinary hands doing ordinary jobs, and the mundanity is the value. Robots have to learn the physical world as it is, so the footage has to be real in a stricter sense than any content platform ever asked: the iron hot, the vacuum running, the object truly moved. Authenticity is the product spec, and the spec is easy to violate.
The fakes are often stranger than the honest work. The first one Dhir ever caught was a person scooping coffee beans from a bowl into a bag, over and over; he assumed it was a one-off, until the same pattern kept recurring. The hardest fakes are the ones that look like work. Loading a dishwasher, two clips can look identical, and the difference is speed and grouping: someone stretching paid time picks up one spoon at a time and places it just so, while someone actually doing the task grabs a handful and lays them out. The task gets done, the state changes, the motion is valid, and something still feels wrong without quite naming itself.
Policing has a cost, and Dhir is quick to name it. The contributor who tries too hard to follow the rules can trip the same wires as a cheat. One person, not wanting a shaky video, mounted the camera on a tripod, which hurt them, because a tripod is not how a body moves through a task. They were being conscientious, and the system read it as unnatural.
Training data is a worse place than most for fraud to hide. A fake ride or a fake review harms one transaction. A fake demonstration of a physical task, accepted into a training set, teaches every downstream robot a small lie about how the world works, and does it quietly, blended into millions of honest clips. The customer never sees it; the customer sees a robot that fails in ways no one can explain. Without the detection layer, Dhir says, invalid and malicious recordings flow into training data and degrade every model below them, and the damage compounds for weeks before anyone notices.
Detection is only half the job. Models catch bad recordings; they miss drifting judgment. Mecka’s data still passes human reviewers, and Dhir built the instrumentation that keeps their work honest: per-reviewer quality dashboards benchmarked against project averages, deliverability and loss metrics, and pipeline-health tooling that tracks the operation end to end. Reviewers are people, and people have good weeks and bad ones. A quality operation that cannot see the difference runs on faith.
What makes the role unusual is where it sits. Most companies split model performance, data quality, and operations into 2 separate teams that meet in escalation threads and speak different dialects. Dhir’s seat spans all three, and he argues the span is the point: a detection model tuned with no operational context rejects good-faith work and burns contributor trust, while operations with no model fluency ships quiet poison. The failure he describes for companies that split the functions is slow drift, not dramatic collapse, problems compounding for weeks before anyone owns them. “Saurab’s engineering work touches on a lot of important functions at the company, and his role is critical to how things come together here,” says Josh Gao, Mecka’s Chief Executive Officer..
Authenticity here is not sold on trust. Every Mecka customer can audit the videos they receive, and the company runs its own QA at every stage of the pipeline. The industry still mostly avoids the subject, because fraud statistics are competitive information and admitting the incentive exists sounds, to a nervous customer, like admitting a fraud problem. So the public conversation stays fixed on volume and price, while the real product moves underneath it.
The vendors that last will be the ones that treated authenticity as an engineering discipline before anyone made them. The unplugged iron is a parable as much as an incident. Fraud at data scale is not a person you catch but a pattern you instrument for, and the companies selling the physical world to machines are learning that their real product was never the video. It was the confidence that the video is true.