In a conversation this summer with Cory Levy, the founder of Z Fellows, Sam Altman described a descendant of ChatGPT that would not wait to be asked. It might watch a computer screen continuously, sit in on meetings, record calls, connect to email and Slack if the user allowed it and assemble what he called “perfect context of your whole life, everything you see.” Asked when such a system might become useful, he answered: sometime within the next six months. He thought the industry was perhaps one model generation away. He was describing, almost casually, a near future in which an assistant would know far more than whatever a person had typed into a box.
The rest of the industry was already moving in the same direction. On Copilot PCs, Microsoft’s Recall can take periodic snapshots of a user’s screen and turn that visual history into a searchable memory of digital life. The feature is opt-in, processed on the device, encrypted and gated by Windows Hello. Microsoft says the snapshots are not sent to the company. The interesting change is conceptual. The computer is becoming a record of what happened on the machine.
Google has been teaching Gemini to reason across a user’s own services. Personal Intelligence, introduced this year as an opt-in feature, can draw on Gmail, Photos, Search and YouTube so that answers are shaped by a person’s mail, images, past queries and viewing history rather than by a generic model of the world. Gemini Live can already look through a phone’s camera or at a shared screen. Google’s old advantage was knowing what people searched for. The emerging one is knowing what they are looking at, what they have already done and how those fragments fit together.
Meta has put the same ambition on a face. Its latest glasses place a camera, microphones and an assistant at approximately the wearer’s point of view, so the model is no longer waiting inside a browser. In late August, the company published a patent application describing glasses that would interpret silent or whispered speech from facial muscle contractions and tiny vibrations in tissue. That capability is a filing. But the sequence it belongs to is already visible. These were the keyboard, then voice, then camera, then wearable sensors, then the body’s own micro-signals.
Apple is building a similar competence with a different architecture. The rebuilt Siri that began rolling out this month can use personal context across messages, mail and photos, notice what is on screen and act across apps. Apple still insists that much of this processing happens on the device or through Private Cloud Compute, which the company describes as stateless computation that neither stores personal data nor makes it available to Apple. The distinction matters. A device capable of handling intimate information is not the same thing as a corporation possessing that information. The political question does not disappear just because the pipeline is more carefully designed.
OpenAI, meanwhile, is more than a software company now. After acquiring Jony Ive’s hardware studio last year, it has been developing a family of devices whose first product, according to reporting and a court filing, is likely to be a screenless, portable companion rather than another phone, and is not expected to ship before 2027. The public language around that project is still a mixture of official ambition, credible reporting and rumor. The direction is clearer than the industrial design. We see an assistant people no longer have to open. These are scenes in one race. The prize is the most complete possible context model of an individual life.
The digital economy of the 2010s was organized around fragments. The questions were clear. What did you search for? What did you click? Where did you go? What did you buy? Whom did you follow? Shoshana Zuboff called the resulting order surveillance capitalism: the value extracted from the behavioral traces people left behind. That description still explains a great deal. But the AI economy is pushing toward a more ambitious question. Here we should ask: Can a machine construct a coherent model of a life? Not merely what this person clicked, but what they are doing, what they are trying to finish, who matters to them, what they discussed last week, what they are seeing now, and what they will probably need next.
Under the older model, advantage came from harvesting traces. Under the newer one, strategic advantage may come from assembling the richest available picture of a person and the situation around them. The shift is a direction. The new system may not need to hide. Its most powerful feature may be that people invite it in. In exchange for an assistant that is actually useful, they grant permission to see mail, photographs, calendars, screens, conversations, meetings, documents, cameras, locations and digital history. The new surveillance may begin by asking permission to see a life in order to become useful. That is a bargain many people will make rationally, because the service really does get better. The expected conclusion from here is familiar. Platforms enclose the individual, agency thins out, the modeled subject becomes a managed one. That ending is too neat.
The same systems that allow institutions to model people more completely are also giving people a kind of productive leverage that used to belong to organizations. In the industrial age, serious capacity usually required capital, factories, newsrooms, universities, bureaucracies and large staffs. AI has lowered the threshold at which a person or a very small group can write software, publish globally, run a research process, design a product, analyze a dataset, produce media, found a company or coordinate a network. One person cannot do everything. Relative to the recent past, one person can do far more.
That changes the composition of power. The emerging order is a crowded field of states, platforms, firms, networks, small teams, researchers, creators and unusually capable individuals. An individual’s value is decreasingly a matter of what they consume and increasingly a matter of what they can make, connect and set in motion. I think we should call it “individual creative energy.” It is a capacity to generate ideas, form unexpected links, use tools inventively, build networks, reinterpret information and produce forms of value that were not already sitting in the dataset.
Here comes the second paradox. The technology that makes people easier to observe can also make them harder to contain. There are limits to how much predictive power can be bought by accumulating more information about a person. Knowing more does not make someone proportionally more predictable. People learn. They change course. They enter new networks, try on new identities, exploit tools, study how recommendation systems work and invent things that were not in yesterday’s record. The model is learning the person. The person is also learning how the model sees them.
That becomes more than a one-way relation between observer and observed. It is iterative. Users alter their behavior around recommendation engines. Creators learn how algorithms allocate attention. Workers learn how evaluation systems score output. Political actors adapt to moderation. Companies learn how models misread a file. The observed subject begins to modify the observer. A perfect behavioral model may remain one step behind the more inventive forms of human agency.
The platform’s strategic objective is perfect context. There is enough knowledge of a user to anticipate a need before it is stated. The individual’s strategic resource is creative agency. It is the ability to make something that cannot simply be inferred from last week’s history. AI raises both sides of the equation. It increases the capacity of institutions to observe, classify, predict and intervene. It also increases the capacity of individuals to produce, organize, experiment and move around institutional bottlenecks. The political future of this technology will depend on the balance between those two forces.
Privacy law still matters. However, it is not enough. In an age of contextual modeling, freedom also depends on epistemic independence, the room to experiment, and the ability to move between networks and systems. A free person becomes someone who can still become something the model did not predict.
Within a few years, an assistant may remember meetings, screens, messages, photographs, voices, searches, movements and perhaps even physical signals that barely register as speech today. It may remember parts of a life more reliably than the person living it. Recording is not understanding. Context is not consciousness. Prediction is not agency. The most stubborn obstacle to reducing a human being to a finished data object remains that person’s capacity to create something that was not already there.
The machines may learn to see nearly every part of a life. The political question of the AI age is whether that visibility will set the boundaries of the individual, or whether the individual will use the same tools to redraw them.
