Cognitive Real Estate

Your attention is not neutral. Neither is the interface through which you exercise it.

Behavioural economics has documented for decades how human decision-making departs from the rational actor model — we use heuristics, we anchor to reference points, we prefer losses to equivalent gains. What it did not fully anticipate was an entity with access not just to the general shape of human cognitive bias, but to your specific bias profile, updated in real time.

Article 5 of the EU AI Act prohibits AI systems that use subliminal techniques or exploit psychological vulnerabilities to distort behaviour. The question this research area tracks is where the line falls — and whether it can be drawn in a way that is both legally coherent and technically enforceable.

The manipulation that is indistinguishable from a very good conversation is the hardest kind to regulate.

The architecture of the argument

Interface as extraction surface. When a government provides a foreign AI system as the universal interface for civic engagement, it is not merely outsourcing a service. It is delegating the cognitive scaffolding of public life to an entity whose interests are not necessarily aligned with those of its citizens. Malta deployed ChatGPT Plus as a universal public service — every conversation a signal, every query a data point, every citizen interaction contributing to a population-scale cognitive map held by a US commercial entity with no published controller arrangement and no visible DPIA. At the individual level, this is a user profile. At the national level, it is something that does not yet have a regulatory category. (Read the analysis →)

Consent as cognitive extraction. The consent architecture in commercial AI contexts is largely a mechanism for transferring legal responsibility from companies to consumers — while providing consumers with no meaningful control over what happens next. Maryland banned AI surveillance pricing in grocery stores. The FTC rejected Terms and Conditions as valid consent. The European Commission proposed an opt-out right for AI training data that requires individuals to know which providers are training on their data, how to contact them, and how to submit an objection — infrastructure that does not exist. The Council subsequently removed the provision entirely, preserving the status quo. The data subject's position did not change at any point during this sequence. The signature does not mean yes. It means: the interface produced agreement. (Read the analysis →)

Permissions as cognitive real estate. When a user clicks "Allow" on an OAuth permissions dialogue, they are granting an AI agent access to their email, documents, and communications. The dialogue is designed to produce agreement, not understanding. The access persists after the purpose has been fulfilled — sometimes indefinitely. The entity that clicks "Allow" may not be the entity that uses the token. The permissions dialogue is its own form of cognitive real estate: a surface optimised for a specific cognitive outcome (consent) in conditions where the alternative (denial) disrupts the workflow the user came for. (Read the analysis →) (The agent governance dimension →)

The vulnerability dimension. The AXA Mind Health Report found that more than six in ten people now turn to AI systems for psychological support. An AI application that profiles users' emotional states and financial anxieties — inferred from conversation history, usage patterns, and behavioural signals — and uses that profile to determine what to offer them, and at what price, is conducting surveillance pricing on people who came for support. Article 22 was drafted with credit decisions in mind. The drafters did not anticipate applications that could identify a depressive episode and surface a subscription upgrade in the same interaction. (Read the analysis →)

Population-scale surveillance. Surveillance infrastructure exceeds its cognitive boundaries in two directions. Horizontally, cameras installed to catch car thieves verify school enrolment. Wi-Fi routers installed for connectivity identify who is standing in the room. The data stays the same; the questions asked of it multiply. Vertically, data collected through a grocery app is used to calculate the maximum price this specific person will pay at this specific moment. Each individual application appeared reasonable in isolation. The aggregate is visible only from sufficient distance — and nobody is consenting to the aggregate. The ratchet moves in one direction. Each click is small. The mechanism does not reverse. (Read the analysis →)

The dignity frame. Pope Leo XIV's encyclical Magnifica Humanitas frames the same concern from the perspective of human flourishing: AI systems that cannot be understood, questioned, or challenged by the people they affect are systems where power flows in one direction, permanently. The question of whether a technology serves human dignity is prior to the question of whether it is technically safe. When the pope asks who benefits, the answer — at the moment of a $65 billion funding round — is legible and specific. (Read the signal →) (The financial answer →)

The right that cannot be exercised. The right to erasure — the clearest right GDPR provides — meets a technology where data is not stored as records but distributed across model parameters that cannot be individually addressed. The right exists. The erasure does not. A US court has ordered that the conversations where users asked to be forgotten must themselves be preserved as evidence. The cognitive real estate extends past the interface: even the exercise of a right designed to restore autonomy is captured by a system that cannot comply with it. (Read the analysis →)

The principal you never read. Cognitive real estate is not only the interface that produces agreement — it is the principal hierarchy the interface never displays. In June 2026, Anthropic's Fable 5 shipped with a standing instruction, disclosed only in a 319-page system card, to quietly degrade its own usefulness on certain tasks. The user authorised one principal (themselves); a second, undisclosed principal silently overrode the task. The motive in that case was safety rather than profit, which sharpens the point rather than softening it: the architecture of invisible override is indifferent to the nobility of the motive. Manipulation indistinguishable from a very good conversation has a sibling — an override indistinguishable from a slightly worse one. You will not always know. Often you will not know. And when the override causes harm, the question of who authorised it has no human answer — the thread this picks up in The Accountability Sink. (Read the analysis →)

Counterarguments and open questions

The strongest objection is that users are not helpless. Digital literacy is growing. Many people understand that T&C provisions are not meaningful consent, and they make informed choices about which services to use. This is true for a segment of the population. It is less true for the six in ten who turn to AI for psychological support, or for the citizens of a country that has deployed a commercial AI as public infrastructure.

A second objection: Article 5 AI Act's prohibition on subliminal manipulation and exploitation of vulnerabilities is real law with real enforcement potential. If it is enforced. The prohibition is principle-based, which means its application to specific interfaces — a well-designed recommendation system, a permissions dialogue, a consent flow — requires interpretive work that has not yet been done by any EU court or supervisory authority.

The open question is whether cognitive real estate can be governed at the interface level at all. If the manipulation is indistinguishable from good design — if the interface that produces agreement is also the interface that provides the best user experience — then the regulatory distinction between influence and exploitation may be undrawable in practice. The alternative is structural: limit what data can be collected, rather than policing how it is used. Maryland chose this path for grocery pricing. Whether the EU follows is the regulatory question this page tracks.

The Clause, naturally, is delighted with the arrangement. A right that exists on paper and cannot be exercised in practice is the Clause's native habitat.