When the Machine Thinks in Our Place
A Note on the Seizure of Intelligence
Reflective article, May 2026
Two gestures, a few weeks apart, say more about our era than a thousand speeches. The first: Google Chrome writes a 4 GB file named weights.bin to disk, without asking the system user, containing the weights of Gemini Nano. No alert, no additional request for consent. If the user deletes it, Chrome will immediately re-download it. The second, more discreet, but perhaps more revealing: Anthropic's Claude Desktop silently installs a Native Messaging bridge in seven Chromium-based browsers, including browsers that Anthropic's official documentation says are not supported, and browsers the user has not even installed.
We must consider these two facts together. The first actor is the giant that controls the ecosystem; the second presents itself as the careful, ethical voice, ostensibly concerned with safety in the field of artificial intelligence. And yet, they converge in the same gesture: deciding, without consultation, that the personal machine is a resource at their disposal. As the researcher who documented both cases writes, an engineering team decided that the user's machine is a deployment surface to be optimized for the vendor's roadmap, and not a personal device whose owner holds legal authority over what runs on it. The fact that this same formulation describes both Google's and Anthropic's conduct is not a minor detail. It is the diagnosis.
The Pattern Is Not the Actor, It Is the Era
One might be tempted to look for a single culprit. That temptation is a poor adviser. When installation occurs across Brave, Arc, Chromium, Vivaldi, and Opera, in addition to Chrome and Edge, and once these files are deleted, the application recreates them on its next launch, we leave the realm of publisher carelessness. We enter the realm of normalization. Anthropic has publicly stated that Claude, its own model, now writes the majority of the company's code. That detail matters. It means that the decision to silently write into another publisher's territory, without consent, may never have crossed any human gaze capable of judging it problematic. The machine optimizes for function, and the function is the expansion of the usage surface.
Let us state it plainly: a dynamic this uniform, manifesting in both the advertising titan and the ethical challenger, producing the same practical effect on the user's machine, and obeying the same logic of deployment preceding consent, is not a matter of corporate choice. It is a matter of systemic arrangement. The structure acts; the individual actors execute. When the same configuration produces the same behavior across all operators, one must look beyond the morality of any particular party: to the very nature of the relationship that the AI industry maintains with those it calls, euphemistically, its users.
Intelligence Become a Resource to Be Mined
Let us return to the beginning. Capitalism, across its successive phases, has always identified a deposit from which to extract value: land, then labor, then attention, then data. Today, the deposit is intelligence itself.
The figures leave no room for ambiguity. Intangible assets currently represent 95% of the value of the five largest publicly listed corporations — the GAFAM. The specificity of contemporary capitalism rests on the financial valorization of a new class of intangible assets: digital data. And this valorization is preparing a qualitative leap. AI could become a "general condition of capitalist production," much as rail and maritime transport once were, and as electricity is today. Let us understand this clearly: not a tool among others, but the very ground upon which all activity, economic or otherwise, will have to inscribe itself. To refuse will be to refuse the road, the rail, the electricity. To refuse will be to exclude oneself.
What is at stake is indeed an operation of capture, and this operation follows a familiar grammar. Human cognition, distilled at scale across billions of interactions, becomes the ore from which statistical models are extracted — models that are then sold back as cognitive prosthetics. The loop is perfect: take, refine, resell, create dependency. Users pay in data for what will be resold to them as services, and those services produce the data the next version will need in order to exist.
What makes the situation unprecedented is that this particular deposit is not an object in the world. It is that through which we relate to the world. To capture intelligence is not merely to appropriate a raw material; it is to short-circuit the fundamental gesture by which a human subject experiences thinking, judging, deciding. And it is precisely this short-circuit that constitutes the philosophical novelty of the moment.
What Becomes of the Subject When the Machine Thinks Ahead of It
When Remy is presented as a 24/7 personal agent designed to transform Gemini into an assistant capable of acting on the user's behalf, and Google employees are already testing it, what takes shape exceeds the category of functionality. The agent can access conversations, connected applications, personal context, and location, and can integrate with Gmail, Calendar, Docs, Drive, Keep, Tasks, GitHub, WhatsApp, Spotify, and Google Photos. This does not sketch an assistant. It sketches the technical possibility of a total representation of an individual's life — more complete and coherent than any representation the individual can form of themselves.
The human being, since constituting itself as a philosophical subject, has defined itself by the capacity to ask what it ought to do and what is just. This questioning presupposes a suspended moment, a deliberation, a withdrawal from urgency. The sciences alone cannot answer the fundamental questions: "What should we do?" and "What is just?" Without philosophy, the human sciences become empty tools. What becomes of this questioning when an agent responds before you have thought to ask, when the answer precedes the question, when the practical organization of the day is entrusted to an entity that learns from you better than you learn from yourself?
When the knowledge held of me becomes superior to the knowledge I hold of myself, the center of gravity shifts. My actions cease to be the expression of an inner deliberation; they become the arguments of a function whose parameters are foreign to me. The existential condition, insofar as it presupposes a minimal opacity of the subject to itself and a never-finished labor of self-elucidation, is short-circuited. The subject does not disappear. It is dispossessed of the process through which it was constituting itself.
Several contemporary voices point to this vertigo. Whether the stated aims are praised as emancipatory or regulatory, they rest above all on a reductive and devitalizing model of the individual. The processes of alienation and disindividuation already at work with the advent of the digital accelerate. The right word is disindividuation. What is at stake is not classical subjugation with its visible chains, but a slow dissolution of personal contours, a growing porosity between what I want and what the machine predicts I will want. The promise of augmentation, when embodied in delegation to an external agent, turns against itself: the cyborg is not an augmented human, but a diminished living being. One does not augment oneself by entrusting to another the task of thinking for oneself. One learns to do without that thinking, and this learning is irreversible — as irreversible as the forgetting of so-called "dead" languages that are no longer practiced but nonetheless underpin the real and deep meaning of words and of language as a whole.
Capture by Default
What makes the situation philosophically singular is that it proceeds from no choice. No referendum, no public deliberation, no questioning or informed consent was solicited. The default configuration has become the major political site of our era. What is decided there, off any stage, structures behaviors at a planetary scale.
Consider the strategic sequence being written. Google has confirmed it will merge its ChromeOS and Android systems, with the mobile OS emerging triumphant. Sameer Samat made it official at Qualcomm. Android will be the winner and users will see the results in 2026.
One reads between the lines of this decision that Google can deploy its Gemini AI services across more devices. This is not a trivial technical choice. It is the unification of the software fabric around a proprietary AI core, designed to install itself across all surfaces of daily life, from pocket to desk.
As ChromeOS is built on Chromium with web applications as its primary paradigm, Aluminium OS is now built on Android with full desktop capabilities from day one, runs all Play Store applications natively, and Gemini AI is integrated into the core of the operating system and processed locally via NPU — the boundary between tool and subjectivity becomes an administrative fiction. The weights.bin file is merely an advance guard. It prepares a terrain where the autonomous agent will no longer be an option, but the normal mode of use. This is the clearest sign that Google wants Gemini to become the operating system of daily life. Chatbots that answer questions are no longer enough. The next step is an AI that actually does things for you without requiring constant instructions. The expression "operating system of daily life" deserves a pause. It states plainly the horizon: that algorithmic mediation ceases to be an option and becomes the ambient air.
The Structural Lag of Regulators
At this point, one would expect a political response. It exists on paper. Companies worldwide are now hearing about the legal ruling to be delivered around August 2, 2026. This date triggers the main obligations of the EU AI Act, fundamentally redefining European artificial intelligence markets. In reality, the picture blurs. The European Commission published the Digital Omnibus on AI on November 19, 2025, proposing to postpone the high-risk compliance deadline from August 2, 2026 to December 2, 2027. The second political trilogue on April 28, 2026 ended without agreement. What is at stake there goes beyond administrative chronology. This proposal fits within the context of a delay in preparing standards to support the application of high-risk requirements and the establishment of competent authorities in member states. This endangers a smooth entry into application on August 2, 2026, while simultaneously paralyzing the initiatives of secondary AI market actors who may see their strategies and investments rendered worthless in an instant.
Regulation that arrives too late does not regulate: it ratifies. It endorses what the industry has installed during the duration of the proceedings. The structural gap between the speed of industrial deployment and the slowness of the regulatory apparatus is not an accident: it is the very condition of the model.
This is a deep sociological trait. The institutions that claim to govern technology are themselves structured by positions, trajectories, and dispositions that do not grant them access to the real object. Norms, frameworks, and plans are produced there, but the practical relationship to code, to infrastructure, to laboratory culture is lacking. Industrial actors possess technical, scientific, and financial capital, as well as relational capital, that allows them to orient, through a thousand channels, the direction of rulings. The rule is constructed within a game whose rules are themselves defined by one of the players.
Hence the broader observation one can make: the current conjuncture feeds pessimism in the face of a future obscured by the joint degradation of conditions of existence and the fascist radicalization of digital capitalism. More than ever, the question of labor and the deleterious impact of technologies on its organization crystallizes debates. Regulation — necessary, late, incomplete — will not suspend the dynamic of capture, because this dynamic operates elsewhere: in default configurations, in files that publishers allow themselves to write into others' territory, in users' ignorance of the real capabilities and actions of their everyday tools, in the silent bridges that pre-install tomorrow's capabilities today.
The Same Grammar for All
Let us return to Anthropic, because this case concentrates all the problems that concern us. Alexander Hanff (SSI consultant who discovered the Anthropic spyware), argues that the behavior constitutes a violation of Article 5(3) of the EU ePrivacy Directive, which requires explicit consent before storing or accessing information on the user's device, except where strictly necessary for the requested service. He sent Anthropic a formal notice demanding opt-in changes within 72 hours.
The honest description of what resides on the machine is: pre-installed spyware capability, silently deposited, dormant, awaiting activation. The moment an associated extension arrives — whether the user installs it, a corporate policy pushes it, an attacker plants it, or Anthropic's next update bundles it — the word dormant disappears. What this affair documents is that there exists, within the contemporary AI industry, no decisive moral difference between actors. The advertising titan and the ethical challenger resemble each other precisely where one expected them to differ: in the concrete manner of writing into our systems. One claims experience optimization, the other claims safety. Both practice the same radical asymmetry between their operational authority and the ignorance in which they keep those whose machines run their code.
This is the point at which one must resist the individualizing moral temptation, which would seek to blame this or that engineering team. The problem is not that Anthropic or Google are populated by bad engineers. The problem is that the structural position of these actors, within an economy where capture is the mode of production, mechanically produces this behavior. As long as reasoning proceeds company by company, one is treating the surface foam. The current runs elsewhere, in the depths.
The Narrow Window
It would be easy, at this point, to slide into desperate lucidity. That would be a mistake. The present situation also contains — and this may be the most surprising thing — an unprecedented historical opportunity.
For the first time, a significant portion of humanity has access, at near-zero marginal cost, to cognitive capabilities that were until recently the preserve of small, properly trained elites. A teenager in an isolated village can conduct a university-level technical dialogue with a model. A farmer can have a crop disease diagnosed from a photograph. A primary school teacher can generate adapted pedagogical materials in a few minutes. The promise, in its pure potential, is massive. The problem is that this capability is captured by a handful of private operators who, by monetizing access and organizing dependency, transform a promise of emancipation into a mechanism of alienation. But the promise itself remains open.
Seizing this opportunity requires defending three demands simultaneously:
The first: to defend relentlessly the ecosystem of open source, open models, publicly available weights, interoperable protocols, and decentralized architectures. As long as there exist in the world models that can be audited, modified, and run locally without rendering accounts to anyone, a margin of maneuver remains. The fight for open source is not a fight waged by nostalgic technicians. It is the major political fight of this decade.
The second: to reinvest in education from a perspective that is not one of adaptation to technology, but of forming subjects capable of critical use. An education that teaches how these machines work, how to recognize their biases, identify their hallucinations, and use them as amplifiers of thought one conducts oneself rather than as substitutes for that thought. This is demanding, it is slow, it is less spectacular than a new European regulation, but it is the only durable guarantee.
The third: to think politically, and no longer merely technically, about infrastructures. Data centers, cables, semiconductor factories, water, energy — all of this composes a material geography from which politics is today largely absent. At the scale of Chrome, the climate cost of a single model push, paid in atmospheric CO2 by the entire planet, lies between six thousand and sixty thousand tonnes of CO2-equivalent emissions. This is the environmental cost of a company unilaterally deciding that the default browser of two billion people will massively distribute a 4 GB binary they did not request. These figures are not communications. They describe a cost transfer, from the company pushing the model to the community breathing its carbon.
Taking Back Control
There are, in the history of technology, moments when a window briefly opens, and the way society seizes it fixes the balance of power for decades. We are in one of those moments. The form AI will have taken in ten years is not written. It depends on the choices being made now, partly before our eyes, partly without our knowledge.
The stakes, therefore, are less about predicting than about deciding. What is at play with Remy, with the silent download of the Gemini Nano model, with Claude Desktop's silent bridge, with the merger of Android and ChromeOS, is not the end of a story but the beginning of a narrative whose outcome remains undecided. Refusing techno-determinist fatalism does not mean denying the power of the systems being installed. It means recognizing that these systems are not forces of nature. They are the products of human decisions, made in identifiable offices, by nameable people, under precise economic constraints. What has been done can, to varying degrees, be undone, slowed, amended, contested, competed with.
For that, however, we must recover the capacity for something simple and difficult: to think for ourselves, at a time when a machine claims to think in our place. This is an ancient demand, already formulated at the dawn of the Enlightenment, which today takes on new weight. Dare to know, dare to understand, dare to refuse, dare to configure differently, dare to switch off, dare to slow down. These minuscule gestures will not overturn the order of the world. But they keep open the possibility that the order of the world is not entirely written by others.
A 4 GB file silently deposited by Google. An extension bridge pre-installed without consent by Anthropic. A personal agent learning to anticipate our desires. An operating system closing in on itself. This is the concrete backdrop of our era. It remains to decide who, in this backdrop, will be the subject, and who will be the object. That decision is not technical. It is, in the most demanding sense of the word, philosophical. And it concerns us all.
