San Francisco's frontier labs are accumulating vast reams of data.


Alex Randall
Nov 18 2026 - 12:00am 8 mins

The founder’s letter accompanying Anthropic’s recent IPO filing includes a rather unsettling forecast: “Authoritarian governments may use AI to create high-tech surveillance states, deepen repression, and make themselves much more difficult to dislodge.” The letter moves on, leaving this statement unexplained. One of the most valuable companies ever to list publicly is telling us that their product might entrench centralized forms of power, and yet public discussion of the eventual outcome implied by this risk is limited.

Anthropic’s business is built on one of the most powerful forms of information processing to have ever existed. Our own human capacity for information processing is arguably the main reason for our dominance of the planet. It enabled our early ancestors to hunt in groups, and it formed the basis of the layered social hierarchies at the foundation of early civilization. Language, writing, and accounting are all types of information processing that allowed us to design better strategies for achieving our goals and started us on the path toward our current technological acceleration. An entity’s ability to gather and process information is a large part of its success or its failure — in many cases, it is a matter of life and death.

When it comes to the game of power, information processing is existential. Knowing more than an enemy knows — gaining an advantage in awareness and knowledge over rivals — is critical for staying in power. It reduces the risk of being outmaneuvered, assassinated, or overwhelmed by complexity or decay. The means to uncover plots against a ruling elite has been a common pursuit of power throughout history. By the 16th century, the strategic advantage of one-sided knowledge had become an obvious prerequisite to the maintenance of power, and organized forms of surveillance were established as an instrument of statecraft. When Venice fell to Napoleon’s forces in 1797, it marked the end of a sovereign state that had persisted unbroken for more than a thousand years, maintained by one of the earliest centrally organized intelligence operations. In England, Elizabeth I died peacefully after a 45-year reign, secured by a network of merchants, couriers, and double-agents that foiled successive Catholic plots to replace her with Mary, Queen of Scots. In both cases, power rested in part on the information gathered about adversaries’ plans. History demonstrates that a strategic advantage in information matters to those with something to lose.

But spies are expensive. At the peak of the Cold War in 1969, the National Security Agency employed almost 100,000 Americans to sift through intercepts and analyze surveillance data. A decade later, the KGB was operating close to half a million people in its network of informants and analysts. Since then, there has been exponential growth in the amount of data gathered by microphones, cameras, and a growing range of technologies that watch the world, its inhabitants and their activities. There is no number of spies you could hire to make sense of it all.  Now, we live in an interconnected web of advanced sensors that expands and deepens by the day, penetrating all aspects of reality at ever finer resolution. 

When the Doge’s Palace in medieval Venice was dispatching agents as far afield as Persia, it was still necessary to be in close physical proximity to a target in order to extract information. This constraint is now a thing of the past. Thousands of satellites are launched each year, and we are close to persistent pattern-of-life tracking of individuals at sub-30cm resolution from space. The next generation of 6G mobile networks will enable what is called refraction analysis in urban environments, using wifi signals bouncing off our bodies to match location and gait analysis to facial recognition data. As these capabilities become increasingly linked across smart-city infrastructures, we will have built the datastreams necessary to know who people are, where they are, and what they are doing at all times. 

Yet our relationship with sensor technologies is as open as it gets. Surveillance has been normalized in a way that is so totalizing we barely notice it. We film ourselves constantly; when something (anything) happens, we willingly record it and store it in databases that we typically do not control. Our digital record is more permanent than a tattoo, with far less certainty regarding ownership. Often we cannot delete our data once it has reached a corporate server; both Meta and Discord state quite plainly that your messages to others, once sent, are no longer your own and cannot be erased. 

This is what we do to ourselves, but what is done to us is orders of magnitude more invasive. For decades, data brokers — those who collect and collate personal data before selling it to third parties — have been refining the science of distilling your essence into an individual digital profile. Over a short period of time, we have become remarkably tolerant of forensic corporate knowledge of parts of us barely known even to ourselves. Modern ad tech uses algorithmic profiling and AI predictive modeling to see through our facades to the truths in our hearts, so that they may be traded in the $363 billion data broker industry. Even the now-antiquated profiling techniques of eye movement tracking, mouse hover patterns, clicks, time-on-site, and time-engaged-on-videos altogether reflect our preferences, emotional states, political views, and likely future behavior with high accuracy. 

Our growing use of AI chatbots is a powerful addition to the data armory. Since 2025, Claude’s consumer plans have presented every user with the pre-selected option to allow training on chat logs. At the time, an Anthropic spokesperson stated that this “provides the model with valuable signals on what works and what doesn’t”, and enables “a feedback loop that helps future models improve on similar tasks”. Of course, “what works” here means the ability to predict what a user will ask next, what they will like in a response, and what will keep them paying their subscription.

This is behavioral modeling at population scale. Whether we want to or not, in each exchange with an AI chatbot, we inevitably share our preferences, desires and vulnerabilities. Every chat reveals our inner world, and feeds into a loop that refines predictions about the user. Each little tweak the AI model makes to the next response makes it feel more magically aligned with our intentions, which in turn makes us feel ever more inclined to share. And so every moment now spent with an AI model (whether Claude, ChatGPT, DeepSeek or any other) feeds into a process of recursive improvement at mind reading, behavior prediction, and methods for optimal control.  

This all adds up to a tremendous amount of data, with each day setting a new record. Historically, an authority in possession of such vast troves would struggle to muster the time and attention to extract the signal from the noise. Until 2017, this was the world in which we lived. But as we have since learned so bitterly, attention is all you need. The vast commercial and proprietary datasets that contain personal information on all of us — the data at the foundation of the global economy — are tractable for population-scale analysis for the first time. 

“Surveillance has been normalized in a way that is so totalizing we barely notice it.”

The reason that Anthropic is warning us that AI could be used to create high-tech surveillance states is because the AI revolution has solved the problem of limited human bandwidth. The innovation that launched Claude, ChatGPT, and others back in 2017 is the transformer, which is a set of mathematical functions rather like an algorithm. Although frontier models are now designed with a blend of approaches in terms of both training and architecture, the transformer backbone is still fundamental to them. Transformers work on what engineers have termed attention: the mathematical process of weighting inputs to the model (such as words in a prompt, or pixels in an image) according to their relevance to other inputs. This means that embedded in each word or pixel is a measure of its relevance to other words or pixels. When enough data flows through these systems, this mechanism of weighted relevance can reveal connections between vastly disparate data points. 

For the first time, we have an alternative to human attention. With enough computing power, the transformer enables machine attention to sift through incredible volumes of information and reveal hidden complex patterns that would have previously taken many human lifetimes to uncover. Multi-modal transformers convert a range of content and data types — images, video, audio, text, location, financial records, browsing habits, network activity, all of which reveal preferences, vulnerabilities, our deepest secrets, and more — into a unified framework. Data types that were previously incompatible can now be integrated. While the transformer has provided a single mechanism for ingesting and fusing many different data types, governments and businesses have been busy removing legal barriers to data fusion at scale. After all, the better the data flowing into the system, the better the outputs at the other end. 

The idea of harvesting the global sensor web to monitor individuals of interest has been around for a while. A quarter of a century ago, Admiral John Poindexter announced a DARPA program to detect, classify, identify, and track terrorists by correlating streams of existing data. The “Information Awareness Office” has since achieved infamy as an early statement of intent for governments seeking to make use of increasingly powerful sensor technologies in the pursuit of total information awareness.

Even back then, no new sensors were considered necessary to gather the data. Now, in the presence of exponentially more data, the current AI wave is opening up a new market niche for tech entrepreneurs aspiring to realize Poindexter’s original dream. Companies with minimal public footprint are building AI tools that unite commercial datasets from a wide range of consumer brands to scan for people and data points of interest on behalf of intelligence customers. As early as 2021, The New York Times was reporting that the Defense Intelligence Agency was buying databases containing smartphone location data and using them to track Americans’ past movements without a warrant. In January this year, the company Anomaly Six (A6), known for embedding its AI software in common mobile apps to track billions of phones, was acquired by the defense contractor Reveal Technology. Government requirements are increasingly being met by start-up innovation in AI data analytics, public awareness of which only occasionally arises by way of leaks or mergers. Total information awareness is emerging as a default outcome of the race for AI dominance, enabled by the exponential development in sensor technologies and the fusion of data sources. 

Over the past few years, frontier AI models have moved from answering questions linearly to building tools and using them autonomously over long time periods. Coordination of analysis across many different datasets is a problem more of orchestration than it is of memory, and swarms of agents are ideally suited to resolve it. AI agents can run tasks in parallel, querying multiple datasets simultaneously, linking salient details from different sources, and keeping records of what has and has not been checked. Swarms can divide workloads and reduce overall costs by chunking tasks and routing easier jobs to cheaper, smaller models. At population scale, this kind of economy is critical. AI agents are replacing the orchestration layer that used to be occupied by humans in analysis of sensor data at scale. When the Stasi operated a surveillance state in East Germany that maintained files on a third of its population, there were never enough officers to read them. We have now built more officers than will ever be needed. 

On our current path, it appears that meaningful dissent will soon become a thing of the past. If we continue to expand the global sensor web — and integrate it with increasingly powerful AI, regardless of whether it is made by Anthropic or any other AI company — then it will soon be trivial to determine what any person or group wants, thinks, and intends. In a world in which everything is recorded and open to analysis by machine attention, those in control of the machines will be able to use their advantage in awareness and knowledge to determine what does and does not come to pass in a world in which they see everything.  

For the elite, this means that the worry of revolution fermenting elsewhere will finally cease, and the transitions of power familiar from history will come to an end. For the rest of us, it means that the potential for change will remain forever beyond reach. Our children will live forever as subjects to those in power. 


Alex Randall is a researcher at the Center for Humane Technology. He has been a policy advisor to the Cabinet Office and the Ministry of Defence.