“The future won’t need iron bars to discipline people. It will need interfaces. When every question becomes a data point, the quietest prison is the one that feels like convenience.”

— Aditya Mohan, Founder, CEO & Philosopher-Scientist, Robometrics® Machines

Bentham’s Panopticon begins as an 18th-century proposition that feels almost like a mathematical proof: arrange bodies in a circle, place an inspector at the center, and let uncertainty do the labor. The core texts are a series of letters drafted in 1787 and published in 1791 as Panopticon; or, The Inspection-House. Bentham is explicit that the device is not a supernatural eye, but a “new mode of obtaining power of mind over mind,” engineered through design rather than brute force. The plan relies on light, angles, and habit: cells made legible; the inspector made inscrutable.

What makes the letters so enduring is how plainly they describe the psychological mechanism. Bentham describes the “centrality of the inspector’s situation” and the “contrivances for seeing without being seen,” and he praises the “apparent omnipresence” of the inspector combined with the “extreme facility of his real presence.” In other words: you do not need constant observation; you need a credible possibility of observation that the observed cannot time or verify. The Panopticon is, at heart, a probability distribution over attention.

A century and a half later, Michel Foucault’s Discipline and Punish (1975) turns the Panopticon into a portable law of social control: power that works by making people visible and measurable, until they internalize the gaze and begin to pre-edit themselves. The structure stops being a building and becomes a pattern: metrics, categories, supervision disguised as administration. Once the “inspection house” becomes an idea, it migrates into schools, offices, platforms, and eventually into the network itself.

A frontier foundation model deployed at national scale can act like that inspection house—except the tower is not masonry, it is compute. The model does not need to “watch” in the cinematic sense. It can operate by inference: compressing speech, images, clicks, and movement into embeddings; linking identities across devices; forecasting what a person is likely to read, buy, fear, or say next. Even if analysis is selective—run on samples, delayed by minutes or days, triggered only by certain risk scores—the uncertainty still reshapes behavior. People begin to write for the model, argue for the model, and live as if the model is always awake, because they cannot tell when they have become legible.

Inside the United States today, the governance of frontier models is a layered stack: private corporate controls, sectoral privacy law, consumer-protection enforcement, and constitutional limits on government compulsion. On the private side, leading labs increasingly publish formal “gating” frameworks that tie capability evaluations to required safeguards. OpenAI’s Preparedness Framework describes tracked risk categories and the use of evaluations, red-teaming, and deployment thresholds to decide whether a system can ship and under what constraints. Anthropic’s Responsible Scaling Policy similarly commits to measuring catastrophic-risk capabilities and escalating security and deployment requirements as models approach higher “safety levels.” On the public-law side, the U.S. remains a patchwork: statutes such as the Privacy Act of 1974 (governing many federal agency record systems) and the Electronic Communications Privacy Act family (wire/electronic interception rules and rules for some stored communications) sit alongside state regimes like California’s CCPA/CPRA. Constitutional safeguards matter most at the boundary where government tries to compel access: the Fourth Amendment’s warrant principle, strengthened in modern surveillance contexts by cases like Carpenter v. United States (cell-site location data), and the First Amendment’s constraints on viewpoint-based coercion and compelled speech. The practical reality is that the Constitution restrains state action, not private moderation choices—so the governance question often becomes: how easily can the state turn private data pipes into public power?

Outside the U.S., the shape of “AI panopticon risk” is being addressed more directly through national and regional governance, though with different emphases. In the European Union, the AI Act creates explicit obligations for providers of general-purpose AI models, and adds extra duties for models deemed to pose “systemic risk” (risk assessment, incident reporting, cybersecurity, and more). France adds a second lens through its data regulator, CNIL, which has been publishing practical recommendations on building AI systems in compliance with GDPR—treating data protection as an engineering discipline, not a checkbox. The United Kingdom has leaned on institutional capacity: the government’s AI Safety Institute (now framed around safety and security research and model testing) operates as a technical counterweight to industry claims, while the UK GDPR and Data Protection Act 2018 provide the privacy baseline. India’s approach is rapidly maturing through its Digital Personal Data Protection Act, 2023 and subsequent rules and government guidance—pushing purpose limits, notice/consent duties, and breach response expectations into mainstream compliance, even as AI-specific regulation remains more fragmented.

The next phase of governance will decide whether frontier models become civic infrastructure or a permanent inspection regime. A credible future stack looks less like one grand law and more like hard constraints that can be audited: strict data minimization by default; compartmentalization so no single breach yields a full portrait of society; tamper-evident access logs; independent model evaluations and incident reporting; and clear “no-go” zones for biometric mass surveillance and political targeting. Most of all, it requires designing systems where the watchers can be watched—where the inspection house does not get to be invisible. Without that reciprocity, the frontier model becomes the perfect Panopticon: not because it sees everything, but because it convinces everyone they are always already seen.