Pseudonymity After Practical Obscurity
The central question of this series is not whether pseudonyms still have value. They do. The harder question is what they can promise when LLMs make it cheaper to extract meaning from text, compare that meaning with auxiliary data, and decide when a likely match is strong enough to return. The answer is a calibrated one: pseudonymity remains a meaningful layer of separation, but it is not a guarantee that persistent semantic traces cannot be linked. A sound judgment must connect the signal, the adversary, the evidence, the access setting, and the consequences of being wrong.
The New Privacy Equation Is About Cost and Linkability
The most useful model is a chain rather than a label. A person writes under a pseudonym. The writing contains ordinary observations—interests, institutions, places, events, habits, relationships, and ways of describing experience. Those observations become more distinctive when combined, repeated, linked across time, or compared with a second source. A matching system extracts a representation of the profile, searches a candidate universe, reasons over plausible alternatives, and calibrates whether to guess or abstain. Practical obscurity is the cost barrier that used to make this chain difficult to run repeatedly.
LLMs change the economics of the chain. They can summarize arbitrary posts into identity-relevant micro-data, retrieve semantically related candidates rather than relying only on exact strings, compare multiple candidates using agreements and contradictions, and produce a confidence signal for a threshold decision. None of these stages creates identity from nothing. The source text and the surrounding information supply the material. The change is that work which once required a skilled investigator’s sustained attention can be divided into reusable, partially automated tasks.
This distinction keeps the claim proportionate. The evidence does not show that every pseudonymous user can be identified, that every model has the same ability, or that LLMs recognize clues no human could ever notice. It shows that selected unstructured-text matching tasks can be performed at a scale and cost that weakens an older assumption: that deanonymization is theoretically possible but practically too expensive to matter broadly. A correct match need not be common to be consequential, especially when the process can be repeated across a large population or directed at people for whom exposure carries unusual harm.
The chain also explains why removing direct identifiers is not the end of privacy analysis. Names, handles, and personal URLs can provide an immediate lookup, so deleting them may remove an obvious route. But quasi-identifiers and high-dimensional profiles can remain. A text corpus may preserve the combination of a technical specialty, a life event, a location history, a recurring interest, and a distinctive style. A stable pseudonym may preserve continuity. A later release may supply the missing employer or date. An outside party may not need the original mapping key if the record can be linked through another combination.
The risk has at least two forms. One is identity disclosure: connecting a pseudonymous profile to a known person. The other is attribute or inferential disclosure: making a location, occupation, health concern, relationship, or other fact more likely to be inferred, whether or not a legal name is attached. NIST’s framework matters here because it defines disclosure broadly and cautions that information can become identifying through association. A person can lose meaningful privacy without appearing in a public “match found” announcement.
The model is asymmetric in another way. A profile can be difficult for the public to identify and easier for an organization that already possesses related data. A platform operator may have account, device, timing, or moderation relationships that are unavailable to an ordinary reader. A researcher may need only aggregate statistics, while a motivated investigator may be searching for one named individual. “Anonymous on the internet” is therefore not a complete threat model.