How AIs See Our World

A man walks down the street at night. On the other side of a store window, an iMac moves as if it were following him with its gaze: when the man jumps, the computer tilts; if he makes a face, the machine opens the CD player. It was an Apple ad from 2002, but today the question is no longer when machines will look at us, but how they already see us. And the answer is stranger than we think: the mistake is not believing that artificial intelligence only has to "understand" our world, but not seeing that it actually reconstructs it from scratch, using codes and shortcuts that seem ridiculous or dangerous to us. The common idea is that we have to teach AIs to see like us, to reason like us. But the truth is that they perceive according to completely different logics, and are often forced to simplify reality to the point of distorting it. The real shift in perspective is that we don't just have to explain things to machines: we also have to learn to see the world from their point of view, to understand where they make mistakes, where they stumble, and why. A concrete example: when the artist Elisa Giardina Papa was working to train an AI to recognize images, she had to draw "bounding boxes" around each object, for example a woman on a sofa. But if the pattern on the woman's shirt blended with the fabric of the sofa, the machine lost the ability to distinguish the two subjects. For the algorithm, Papa says, "the image ended up in a queer, indeterminate category: neither just a woman, nor just a sofa." It was not a trivial error: it was the structural limit of a system that sees only what it can label. And the same thing happens with transparent objects or with a chair placed sideways: it doesn't take much for reality to break the mold. The artist Eryk Salvaggio, who was training an AI to recognize mushrooms in the woods, had to ignore every unusual detail — the exact opposite of human instinct, which seeks the exception. The result? AIs see the world as if it were a statistical average, not a collection of unique details. And when we try to adapt the world to them, strange things happen. The Amazon Go stores, where you could leave without going through the checkout, needed a thousand people to manually check the images and correct the AI's errors, and for every thousand purchases, human intervention was needed in seven hundred cases. In the end, the experiment was shut down. Another extreme case: the tragedy of Elaine Herzberg, who was hit by a driverless Uber car. The sensors had "seen" the woman on the road, but they had failed to classify her: sometimes as a vehicle, sometimes as an unknown object, sometimes as a cyclist. Having never labeled her as a pedestrian not on the crosswalk, the system did not brake. The problem was not the technology, but its inability to accept something that did not fit the predefined patterns. But these simplifications are not just technical. Tom Williams, a robotics scholar, describes how AIs often misunderstand people who speak with different accents or dialects — forcing them to "whiten" their voices to be recognized. And when companies try to improve data to be more inclusive, they risk creating new forms of surveillance or reproducing old prejudices, such as when Microsoft used racial categories borrowed from 19th-century pseudoscience. So what can be done? An alternative approach comes from those same systems that try to overcome the "bounding boxes". Waymo has started to represent people as digital stickmen, point-based skeletons that make it possible to understand not only if someone is crossing the street, but also in which direction they are looking, or if they are about to change direction. In this way, the computer does not just "put in a box" what it sees, but tries to interpret poses, intentions, and movements. It is a step towards an idea of empathy that is not emotional, but cognitive: we are not asking AIs to feel emotions, but to understand — at least in part — our way of being and moving. To succeed, we also need to rethink our interfaces: it is not enough that they are easy for us; they must put AIs in a position to truly grasp the complexity of the human environment. Just as we once used folder-shaped icons to help people understand computers, now we also need to build "translators" that help AIs grasp the nuances of our world, not just its simplified version. And the final twist is this: the path is not just to teach AIs to see like us, but to learn to live with their alien way of interpreting reality, knowing where they can stumble or hurt. Otherwise, we risk living in a world designed for algorithms, not for people. If you thought that "teaching AIs to see like us" was enough, get ready to change your mind: we must also learn to see like them, to prevent the world from becoming a series of boxes where no one is truly recognized. If this perspective has struck a chord with you, you can indicate it on Lara Notes with I'm In — it's the gesture that says: this vision is now part of how you think. And if you tell someone about it tonight — perhaps recalling the absurdity of Amazon Go stores or the case of Elaine Herzberg — on Lara Notes you can tag those who were with you with Shared Offline, so you know that conversation mattered. This journey into AI perceptions comes from NOEMA and saved you 14 minutes.