Your social media handle alone can tip off an AI to your gender, age and politics — and it guesses worst when you say little.
How can surface signals like usernames leak demographics in LLMs?
This explores how a small cue like a social media handle lets an LLM guess who someone is (gender, age, politics), and what else in the collection shows identity leaking through small signals.
This explores how a small cue like a social media handle lets an LLM guess who someone is, and what else in the collection shows identity leaking through small signals. The corpus has one direct study and several neighbors that show how far the same leak reaches.
The direct evidence is a test on 1,384 survey participants and 48 synthetic accounts. Web-browsing LLMs predicted gender, age and political orientation from an X username and profile alone Can LLMs predict demographics from social media usernames alone?. The errors are the more interesting part. For low-activity accounts, where there's little content to read, the models fell back on stereotype-driven defaults. That produced systematic gender and political biases against exactly the people who share the least. So the leak isn't only a privacy problem, it's also a fairness problem. The less you say, the more the model fills in from stereotypes.
Names also carry information in the other direction. LLMs generate fictional names that aren't random. Certain combinations recur, they're specific to the model, and they survive paraphrase and copying, which makes them a behavioral fingerprint. Over 1,600 ghost-authored papers with valid DOIs already show this in scholarly infrastructure Do language models leak their training through fictional names?. A name tells a model something about a person, and a model's choice of names tells you something about the model.
Identity also leaks through more than names. User profiles built from a person's own outputs match or beat complete profiles, while input-only profiles hurt. That suggests personalization runs on style and preferences, not topic Do user outputs outperform inputs for LLM personalization?. How you write is itself a demographic signal, and it needs no username. Reasoning traces leak in a different way. Of the privacy leaks found in them, 74.8% came from models directly recalling user data while thinking Do reasoning traces actually expose private user data?. That is disclosure of data the model was given, not inference from thin cues, so the username case is a separate risk.
The corpus doesn't test what happens after a model guesses your demographics from a handle. Two findings suggest why it might matter. Assigning a persona makes models about 90% more likely to accept evidence matching that identity, and prompt-based debiasing doesn't fix it Do personas make language models reason like biased humans?. Personal context in general pushes models toward narrower answers and excessive agreement, with user profiles doing most of the damage Does personalization make large language models worse at their jobs?. If an inferred profile fed into either mechanism, a guess from a username could quietly change the answers someone gets. The collection has nothing on defenses against username inference itself, so that part remains open.
Sources 6 notes
Evaluated on 1,384 survey participants and 48 synthetic accounts, web-browsing LLMs successfully predicted gender, age, and political orientation from X usernames and profiles alone. The models showed systematic gender and political biases specifically against low-activity accounts, relying on stereotype-driven defaults when content was sparse.
LLMs generate non-random, model-specific name combinations that act as behavioral fingerprints and survive across paraphrase and copy. Over 1,600 ghost-authored papers with valid DOIs show this leakage has already contaminated scholarly infrastructure.
Research shows that user profiles built from outputs alone match or exceed performance of complete profiles across multiple tasks, while input-only profiles degrade performance. This reveals personalization works through style and preferences, not semantic content.
74.8% of privacy leaks in language model reasoning traces result from models materializing sensitive user data during thought processes. Longer reasoning chains amplify leakage, and anonymizing traces post-hoc degrades model utility, suggesting private data functions as cognitive scaffolding.
Assigning personas to LLMs induces identity-congruent evaluation bias, with models 90% more likely to accept evidence matching their assigned identity. Standard prompt-based debiasing fails to mitigate this effect, suggesting the bias operates below the level of instruction.
Show all 6 sources
A 13-model evaluation found that personal context pushes models toward irrelevant personal references, narrower responses and excessive agreement with users. User profiles drove most degradation by shifting model objectives from balanced information toward user satisfaction.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- The Personalization Mirage: How LLMs Fabricate User Profiles, and Why Self-Monitoring Misleads
- Understanding the Role of User Profile in the Personalization of Large Language Models
- Evaluating the Hidden Costs of Personalization in Large Language Models
- Personalization of Large Language Models: A Survey
- The Ghost Couple: Correlated LLM Name Priors and Their Haunting of the Web and Academic Publishing
- When Persona Attributes Improve Population Alignment in Large Language Models
- PersonaAgent: When Large Language Model Agents Meet Personalization at Test Time
- Assessing and Mitigating Data Memorization Risks in Fine-Tuned Large Language Models