The human labor behind humanoid robots is carefully hidden

In January 2026, Jensen Huang, head of Nvidia, the world's most valuable company, solemnly announced the arrival of the era of physical artificial intelligence. In his words, AI is finally moving beyond language models and chatbots, gaining a body and the ability to interact with the real world. It is a polished line, but it hides an uncomfortable truth: much of what the industry passes off as a breakthrough in robotics still runs on human labor.

A report by MIT Technology Review exposes a systemic problem that Silicon Valley prefers not to discuss out loud. Dozens of companies, from giants like Tesla and Nvidia to ambitious startups with billion-dollar valuations, regularly publish impressive videos in which humanoid robots fold laundry, sort items in warehouses and perform complex object manipulation. Viewers see the future. But off camera there is often an operator with a joystick or a VR headset, controlling every movement of the machine in real time. Or an engineer launching a pre-rehearsed sequence of actions and passing it off as autonomous behavior.

This is not fraud in the legal sense: the companies do not formally claim that their robots are fully autonomous. They simply omit the details. Marketing videos are edited to give the impression that the machine acts on its own. Press releases are packed with the words "AI-driven" and "autonomous," but nobody reads the fine print. Investors, journalists and the general public see what they are meant to see, and draw conclusions that benefit the manufacturers.

Teleoperating robots is neither new nor objectionable in itself. It is an important stage of development: a human operator collects data on movements, which is then used to train neural networks. The method is called "learning from demonstrations," and it does help robots acquire new skills. The problem arises when a development stage is presented as a finished product, when a demo with a human operator is offered as proof that the robot can already do all this by itself.

The gap between promise and reality in robotics resembles a situation the AI industry has already been through with language models. Remember how early chatbots were advertised as "understanding" conversation partners, when in reality they just predicted the next word? Something similar is happening with robots, only the stakes are higher. The physical world is far more complex than the digital one: a robot must account for gravity, friction, the fragility of objects and the unpredictability of its surroundings. Each of these variables is a separate engineering problem, and none has yet been solved well enough for reliable autonomous operation.

The consequences of this gap could be painful. Investors are pouring billions of dollars into companies whose real capabilities they overestimate. Startups feel compelled to sustain the illusion, because an honest demonstration of the current state of the technology will not win the next funding round. A vicious circle emerges: inflated expectations demand even more impressive demos, which demand even more hidden human involvement. Sooner or later this bubble will collide with reality, and the disappointment may hit the whole industry, including those who work honestly.

There is also an ethical dimension that goes beyond investment risk. The people who operate robots remotely often work in far-from-ideal conditions: long shifts, monotonous motions, low pay. Their labor is literally invisible, hidden behind a marketing narrative about machines that "can do everything themselves." It is an irony worth discussing on its own: a technology that promises to free people from routine physical labor is, at this stage, creating a new form of exactly that labor.

None of this means humanoid robots are a dud. Progress in robotics is real, and some systems do show impressive learning abilities. But between a lab prototype and a reliable commercial product lies a chasm that the industry is currently filling with marketing and human hands. An honest conversation about where we actually stand would serve everyone better: investors, engineers and the society that will ultimately have to live alongside these machines. The era of physical AI may well be arriving. But it is arriving more slowly than those selling it would like.