Artificial intelligence has acquired another capability that forces us to reconsider something we have taken for granted about the Internet: anonymity.

For most of the Internet’s history, people have been able to participate under pseudonyms, screen names, and accounts that did not explicitly disclose their real identities. That never guaranteed absolute anonymity. A sufficiently determined investigator might piece together enough clues to discover who was behind an account. But doing so could require hours, days, or weeks of investigation. That cost and difficulty created a form of practical anonymity.

Artificial intelligence may be changing that assumption.

Researchers have demonstrated that large language models can analyze seemingly ordinary pieces of information scattered throughout a person’s online activity, extract identity-related clues, search for possible matches, and reason across those clues to identify the person behind a pseudonymous account.

That is an impressive technological capability.

It can also be used for good or for harm.

That makes this research an unusually clear example of a broader problem we increasingly face with artificial intelligence: the same technological capability can simultaneously produce positive and negative consequences. Our challenge is not simply to encourage or prohibit the technology. It is to understand its consequences well enough to optimize its beneficial uses while minimizing its harmful ones.

The Research

The research is described in “Large-scale online deanonymization with LLMs,” by Simon Lermen, Daniel Paleka, Joshua Swanson, Michael Aerni, Nicholas Carlini, and Florian Tramèr. The authors are affiliated with MATS, ETH Zurich, and Anthropic. The work appears in the 35th USENIX Security Symposium in 2026.

The researchers demonstrated that large language models can perform deanonymization at a scale and cost that were previously impractical. With Internet access, their agent could attempt to re-identify Hacker News users and participants in Anthropic interviews from pseudonymous profiles and conversations. They also constructed controlled experiments linking Hacker News and LinkedIn profiles, matching users across Reddit communities, and matching different portions of the same Reddit user’s history separated in time. (USENIX⁠)

The significance lies partly in what AI can use as evidence.

A person does not have to publish a name, home address, or telephone number.

Consider the fragments we leave behind over years of ordinary conversation: where we went to school; what technologies we have worked with; cities where we have lived; unusual hobbies; industries in which we have worked; events we attended; people we have known; organizations with which we have interacted; and subjects about which we possess unusual expertise.

Individually, these facts may reveal very little.

Together, their intersection may describe very few people—and sometimes only one.

How AI Changes Deanonymization

Traditional deanonymization frequently depended upon structured information or extensive manual investigation. The researchers demonstrated that LLMs can instead work directly with unstructured human language.

Their scalable approach extracts identity-relevant characteristics from text, searches large candidate populations using semantic embeddings, and then uses an LLM to reason across the strongest candidate matches to verify matches and reduce false positives. (USENIX⁠)

In simplified form:

ordinary conversation → identity clues → candidate identities → evidence correlation → probable identification

In the researchers’ controlled experiments, LLM-based methods substantially outperformed classical approaches. The researchers concluded that the cost of deanonymizing pseudonymous Internet users has fallen sharply. (USENIX⁠)

That changes the economics of anonymity.

An investigation that once might have been economically practical only for an important intelligence target, criminal investigation, or determined stalker could eventually become inexpensive enough to perform against thousands or perhaps millions of accounts.

And that is where the consequences become complicated.

One Capability, Two Directions

There is nothing inherently beneficial or harmful about recognizing that two collections of information probably describe the same person.

The consequences depend upon who performs the identification, why it is performed, what authority permits it, how reliable the identification is, what happens to the resulting information, and what subsequent actions it enables.

Consider two uses of essentially the same technology.

A stalker uses AI to discover the identity of a woman participating anonymously in an abuse-survivor community.

An authorized investigator uses similar technology, subject to appropriate legal safeguards, to identify someone operating multiple accounts to exploit children.

Technologically, the identity-correlation operation may be remarkably similar.

The consequences could hardly be more different.

That is why regulating a technology merely by examining its individual technical capabilities may prove inadequate.

The Negative Consequences

The immediate privacy concern is obvious.

People sometimes have legitimate reasons to separate their real-world identities from their Internet identities. Whistleblowers, political dissidents, abuse victims, employees discussing workplace problems, medical-support participants, journalists’ sources, and ordinary citizens discussing sensitive personal subjects may depend upon pseudonymity.

Automated deanonymization could weaken that protection.

It could facilitate stalking and harassment. It could expose whistleblowers. Governments without adequate civil-liberty protections could use it to identify critics. Criminal organizations could use it against investigators or witnesses. Employers or other organizations could attempt to associate anonymous criticism with employees. Data brokers could potentially enrich existing profiles with previously disconnected activities.

There is another danger: being wrong.

A system operating at 90 percent precision is not correct 100 percent of the time. At population scale, even a comparatively small error rate could falsely associate substantial numbers of people with statements, organizations, or activities with which they had no connection.

The consequence therefore extends beyond loss of privacy. It can include reputational damage, employment consequences, physical danger, and erroneous investigative attention.

The Positive Consequences

Now reverse the problem.

The same capability could help identify networks of fraudulent accounts, coordinated scams, and people repeatedly creating new identities after legitimate platform enforcement.

It could help cybersecurity investigators correlate aliases appearing across malware forums, code repositories, previous attacks, and other sources.

It could help detect impersonation and identity theft.

It could help platforms determine that dozens of supposedly independent accounts are actually controlled by the same person or coordinated organization.

It could help authorized investigators find missing or exploited people.

It could assist properly authorized investigations of people using multiple identities to exploit children.

And it could create an important defensive capability. Instead of asking, “Who is this anonymous person?”, an AI system could ask, “Is this person actually the person he or she claims to be?”

That transforms identity correlation from a deanonymization mechanism into a potential form of identity assurance.

Once again, the underlying AI capability has not fundamentally changed.

Its purpose, authority, and consequences have.

Two Different Consequence Pathways

This example becomes even more revealing when examined through the Taxonomy of Consequences developed in the SPARTANSFIRST Technology Blog and the appendix to Hunter of Shadows (currently being edited for this fall).

Positive and negative consequences do not necessarily spread in the same way.

The positive pathway is:

Discovery / Invention / Innovation / Reform / Investment → Impact Radius → Diffusion → Emergence Horizon → Synthesis → Compounding Advancement

A new capability first creates an Impact Radius: the immediate area in which beneficial consequences can occur.

Those benefits can then undergo Diffusion. Other people, organizations, industries, researchers, and entrepreneurs adopt the capability and find additional applications.

Eventually that diffusion can approach an Emergence Horizon, where applications and combinations become possible that were not necessarily apparent when the original innovation occurred.

Those developments can undergo Synthesis with other technologies and knowledge, producing Compounding Advancement and potentially beginning additional cycles of innovation.

AI identity recognition illustrates the process.

Its initial Impact Radius might include identifying fraudulent accounts or discovering that apparently independent accounts belong to the same criminal organization.

Diffusion could carry the capability into cybersecurity, fraud prevention, detection of impersonation, child protection, identity assurance, and other legitimate applications.

At the Emergence Horizon, identity correlation might be synthesized with authentication, behavioral security, provenance technologies, privacy-preserving computation, and agentic AI in ways we have not yet anticipated.

The negative pathway is different:

Negative Trigger Event → Blast Radius → Propagation → Cascade Horizon → Interactive Feedback → Potential Systemic Collapse

A harmful use or erroneous identification creates a Blast Radius representing its immediate adverse consequences.

Those consequences can then undergo Propagation through interconnected people, databases, organizations, technologies, and decision systems.

Consider an AI system that improperly or incorrectly identifies the person behind a pseudonymous account.

The immediate Blast Radius could include exposure of that individual’s identity.

The identification might then propagate. An employer could associate the individual with anonymous criticism. A data broker could incorporate the inferred identity into a commercial profile. Other databases could reproduce it. Automated systems could subsequently use the information in employment, insurance, credit, security, or other decisions.

The original identification might take seconds while its consequences could propagate for years.

Interactive feedback could make the problem still worse. One database could appear to corroborate another even though both ultimately originated from the same AI-generated inference. Repeated automated use could amplify an original error until uncertain information acquires the appearance of established fact.

The distinction between the pathways matters:

Benefits diffuse. Harms propagate.

Positive consequences move outward from an Impact Radius through Diffusion toward an Emergence Horizon, where new possibilities can arise and combine through Synthesis into Compounding Advancement.

Negative consequences move outward from a Blast Radius through Propagation toward a Cascade Horizon, potentially producing Interactive Feedback and, in extreme circumstances, systemic consequences.

The same AI capability can initiate both pathways simultaneously.

The Responsibility of Internet Platforms

Social-media companies and other Internet platforms therefore confront a difficult policy question:

Under what circumstances should AI be permitted to infer the real-world identity behind a pseudonymous account?

The first problem is that detecting an individual deanonymization attempt may itself be difficult.

The steps involved can look remarkably similar to normal platform activity: reading posts, searching for names or subjects, examining profiles, following links, comparing information, and moving from one piece of information to another.

An agentic AI system could perform the reasoning between those actions at machine speed, determining its next investigative step in milliseconds. But it does not have to execute the resulting actions at machine speed.

It could deliberately slow itself down.

Instead of conducting an investigation in an hour, an automated system could distribute the same activity across several days. Searches and profile visits could therefore resemble the ordinary behavior of a human user. Human investigators cannot simply accelerate themselves to machine reasoning speed. Automated systems, however, can operate at machine speed internally while deliberately presenting their external activity at human speed.

That asymmetry makes individual deanonymization attempts particularly difficult to distinguish from legitimate platform use.
 Large-scale activity presents a different problem. A platform may be much more capable of recognizing an account, application, or coordinated group systematically examining hundreds or thousands of users. Volume, repetition, common behavioral patterns, and coordinated access can provide signals that do not exist in a single investigation.

Platforms therefore should not assume that one technological control will solve both problems.

Policies will need to distinguish among legitimate security analysis, fraud prevention, authorized investigations, account-integrity functions, research, commercial identity enrichment, harassment, and unauthorized attempts to expose someone’s identity.

They may also need controls addressing scale, automation, purpose, authority, and subsequent use—not merely whether someone performed an otherwise ordinary search.

The Challenge for Privacy Law

Government faces the same problem.

AI deanonymization illustrates why privacy policy increasingly must consider not only information an organization collects, but also information an AI system can infer by combining otherwise public or apparently innocuous information.

Yesterday’s privacy question might have been:

What personal information did you collect?

Tomorrow’s privacy question may also have to be:

What personal information did your AI derive?

Those are fundamentally different questions.

The Growing Gap Between Technology and Regulation

This example exposes a larger structural problem.

The mechanisms through which government traditionally develops legislation and regulation move deliberately. Legislatures hold hearings, obtain testimony, negotiate competing interests, draft legislation, revise it, enact statutes, and frequently delegate implementation to administrative agencies.

There are good reasons for deliberation. Rules affecting individual liberty, innovation, commerce, security, and government authority should not be created casually.

But technological capability is advancing much faster than these mechanisms were designed to accommodate.

The result is an increasing gap between the state of technological capability and government’s ability to understand it, evaluate its consequences, and determine an appropriate regulatory response.

The problem becomes more difficult as the technologies themselves become more sophisticated. Legislatures include people from many professional backgrounds, with lawyers and businesspeople traditionally strongly represented. They cannot reasonably be expected to possess deep expertise simultaneously in artificial intelligence, cybersecurity, biotechnology, quantum computing, semiconductor technology, telecommunications, energy systems, and every other rapidly developing field.

That is not a criticism of legislators.

It is a structural problem.

Perhaps we therefore need a better mechanism for continuously providing legislators with timely, technically competent, multidisciplinary information about emerging technologies and their consequences.

The objective would not be to tell elected representatives what decisions to make. It would be to provide them with sufficient technical understanding to make informed decisions themselves.

This Question Did Not Begin With AI

There is an established intellectual foundation for thinking about this problem.

In 1948, mathematician Norbert Wiener published Cybernetics: Or Control and Communication in the Animal and the Machine. Cybernetics examined communication, feedback, and control in complex systems and established ideas that subsequently influenced engineering, computing, biology, management, and the social sciences.

Those ideas have also been applied to regulation and governance.

Karen Yeung’s 2018 “Algorithmic Regulation: A Critical Interrogation” examines algorithmic regulation explicitly through a cybernetic framework. She describes regulatory systems in terms of standard setting, information gathering and monitoring, and mechanisms for behavioral modification, while developing a taxonomy of different algorithmic regulatory arrangements and examining their implications for legitimacy, surveillance, and governance. (Wiley Online Library⁠)

More recently, Andrej Zwitter’s 2024 “Cybernetic Governance: Implications of Technology Convergence on Governance Convergence” brings together earlier work from several disciplines and directly addresses the problem that concerns us here. Zwitter argues that technological advances and technological convergence are outpacing regulatory efforts and making existing ethical and legal frameworks increasingly difficult to apply. He develops the idea of cybernetic governance by bringing together governance theory with cybernetics, including feedback and control and Ashby’s Law of Requisite Variety. (Springer⁠)

The important point is that applying feedback, cybernetics, and systems thinking to government regulation is not a new idea. There is already a substantial intellectual history behind it.

The question worth reconsidering is whether the extraordinary acceleration and increasing sophistication of technological change now make those ideas considerably more important.

The Larger Question

The mechanisms that historically enabled government to understand technological change and create appropriate regulation move comparatively slowly. Technological capabilities are developing increasingly rapidly.

The gap between those rates of change is widening.

We may therefore need a new structure for informing legislators—one that continuously brings together technologists, scientists, entrepreneurs, economists, lawyers, security specialists, affected communities, and other relevant expertise.

The objective should be optimized regulation: preserving sufficient freedom for beneficial innovation and its Diffusion toward the Emergence Horizon while creating appropriate safeguards that limit harmful consequences and interrupt their Propagation toward the Cascade Horizon.
 Cybernetics offers an intriguing way of thinking about the problem because regulation itself can be understood as a feedback process:

observe → understand → decide → act → measure consequences → learn → adjust

Wiener began developing the intellectual foundations for thinking about communication, feedback, and control more than three-quarters of a century ago. Researchers since then have extended those ideas into regulation, algorithms, and governance.

Artificial intelligence now gives us a reason to return to that body of work.

The question is no longer merely:

How should government regulate a particular technology?

A deeper question is emerging:

How should we design a regulatory system capable of learning and adapting when the technologies it governs are changing faster than the traditional regulatory process?

That is a question I intend to explore much more deeply in the coming months.

References

1. Simon Lermen, Daniel Paleka, Joshua Swanson, Michael Aerni, Nicholas Carlini, and Florian Tramèr, “Large-scale online deanonymization with LLMs,” 35th USENIX Security Symposium, 2026. (USENIX⁠)
    USENIX paper and presentation⁠
2. Karen Yeung, “Algorithmic Regulation: A Critical Interrogation,” Regulation & Governance, Vol. 12, No. 4, 2018, pp. 505–523. The article analyzes algorithmic regulation using a cybernetic regulatory model and synthesizes issues arising from regulatory-governance, legal, surveillance, and critical-data scholarship. (Wiley Online Library⁠)
    Yeung article⁠
3. Andrej Zwitter, “Cybernetic Governance: Implications of Technology Convergence on Governance Convergence,” Ethics and Information Technology, Vol. 26, Article 24, 2024. The paper develops cybernetic governance from earlier work across governance and cybernetics and specifically examines the problem of technological development outpacing existing regulatory approaches. (Springer⁠)
    Zwitter article⁠

4. Norbert Wiener, Cybernetics: Or Control and Communication in the Animal and the Machine, first published 1948; second edition 1961. MIT Press. Wiener established cybernetics as the study of communication, feedback, and control in complex biological, mechanical, and social systems, providing the foundational concepts underlying later applications of cybernetic thinking to regulation and governance. MIT Press open-access edition⁠