Make Sense of It
I see this up close. I lecture at universities, and in less than three years, institutions that are hundreds of years old have changed fundamentally. Students submit papers written, in part or in full, with the help of language models. Lecturers, for their part, check them with those very same tools. It is not only the students who have stopped researching, wondering, and lingering in uncertainty, all parts of learning, their lecturers have stopped too. And all of us keep behaving as if everything is business as usual.
This is the new reality, for the first time we are handing over thought itself to another entity. Not memory, as we did with the invention of writing. Not calculation, as we did when the computer was born. But the cognitive core. Those nights when we racked our brains over a paragraph or a mathematical proof, until something emerged that was truly ours. That muscle is weakening.
The industry numbers tell the same story. Satya Nadella, CEO of Microsoft, said that 20% to 30% of its code is now machine-generated. Kevin Scott, the company’s chief technology officer, predicts that within four years 95% of the code will be created by AI. Anthropic announced that 80% of its code has already been written by Claude. But these astonishing numbers do not tell what happens to the person sitting in front of the screen. A junior programmer in 2026 no longer writes code as she did just two years ago. She mainly creates prompts and approves.
This is happening in almost every profession, lawyers, accountants, researchers. Much of the time is devoted to approving or rejecting decisions that artificial intelligence made for us. Each such approval is a small expression of trust in a process no one can inspect from beginning to end. Often these are not trivial decisions. In the United States, AI systems are used by insurance companies to approve or deny claims, and by courts to assess the risk of reoffending. In hospitals, AI tools present diagnostic recommendations to doctors. In Israel, as has been reported, AI systems are integrated into the assessment of military targets. The pattern in all cases is the same, a system no one fully understands offers an output, a human approves it in a fraction of a second, and the decision is carried out.
"We do not understand what we created"
Even the companies building the systems no longer know how to explain them. They control the training protocol, but even for them the result is not known in advance. They tune the parameters but remain surprised by the capabilities, and the failures, that emerge. They do not know why a model produced a particular result. They created the mechanism, but they cannot trace its internal logic.
True, in neuroscience we encounter a similar problem. Researchers know how to map which areas of the brain are responsible for vision or language, but struggle to decipher the internal language of the electrical currents. No one can point to a specific spot and say, here the memory of the smell of chocolate milk from childhood was born. But the human brain is the product of hundreds of thousands of years of evolution that shaped in us a basic intuition about how it works. Language models, by contrast, were built in just a few years. We have no shared evolution with them. We are strangers to the way they reach conclusions.
Anthropic has a research team called Interpretability, whose job is to try, after the fact, to understand what the model did and why. Sometimes it succeeds. Many times it does not. "People are disturbed to discover that we do not understand how our own creations work. Their concern is justified, this lack of understanding is unprecedented in the history of technology," the company’s CEO Dario Amodei admitted. Sam Altman, CEO of OpenAI, says similar things. The companies that produce AI tools, used by hundreds of millions of people in sensitive areas, are saying that they themselves do not understand what they built.
At first glance, what is the big deal? We do not fully understand how an airplane stays in the air, either, or how Facebook’s algorithm really works. But there are two main differences between those and AI. The first is the pace. ChatGPT reached 100 million users in two months. The mobile phone took 12 years to reach half that number. A car took 62 years. And over the six decades that passed before cars spread, there was enough time to write traffic laws, develop an insurance industry, and create norms. There is no such time with AI. By the time a lawmaker drafts a regulatory framework for its use, a new model emerges whose capabilities are no longer covered by that draft.
The second difference is more fundamental. In the case of airplanes, Boeing engineers know how the plane flies. In the same way, Meta engineers know how the feed they built works, they simply do not say. The ones who do not know are the passengers, the users, the end consumers. By contrast, in the case of language models, the people who built them are the ones who do not know.
Read more in "Calcalist Magazine": Prof. Yanai Ophir grew up in an unconventional home, followed an unconventional career path, and now has his name on breakthrough drugs that could transform research. A new study finds that we speak much less, and that is very bad for us. "There is a lot of bullshit here, so if you need to raise prices, my brother, raise them": how Eyal Raviv, owner of Victory, got himself in trouble with the authorities. Salvation or ruin, now!
The term "singularity," coined by science fiction writer Vernor Vinge in 1983, seeks to describe the moment when technological development will surpass human understanding and control, leave humans behind, and create unpredictable changes in civilization. It is the moment when the machine will be smart enough to design a machine smarter than itself, which will design the next one, and so on, at a pace detached from the human one.
Today, the debate over when that moment will arrive is tied to the artificial intelligence revolution. Timelines: the end of 2026, Elon Musk, during 2026, Dario Amodei, or 2047, a survey among AI researchers. But the singularity is not a distinct future technological event, a turning point beyond which everything will change. The singularity will emerge through a series of events that will accumulate critical mass, until it can no longer be denied. And those events are already happening today, in every interaction we have with AI tools.
The singularity is the moment when we stop understanding what the machine is doing and why, but keep using it anyway. That moment is already here. It just does not look the way we imagined.
The idea of a single moment after which "everything will change" has a theological flavor. There will be salvation, or ruin, but moral responsibility lies with a higher power. "The singularity will arrive" is a forward-looking argument. Yet different actors are building these systems today. Someone decides how they will be built, someone oversees them, or does not. Someone prices, deploys, and markets them. These are political, business decisions made by humans in boardrooms. They are not a higher power, but the future-oriented language of singularity turns them into fate.
At least in theory, as long as the singularity is not here, we are exempt from looking at what is happening under our noses, even if models are already deciding critical aspects of our lives without anyone knowing how to explain them. But technological capabilities do not wait for a symbolic point in the future. They are here, and they must be dealt with.
The author is CEO of Humane AI and an expert in human-AI interaction.
