Can an AI guess your age, gender, and politics just from your username and public posts?

This explores whether LLMs can work out who you are from indirect traces you leave behind, and whether their guesses hold up. The short answer is yes, and that's the uncomfortable part. Web-browsing LLMs given nothing more than an X username and public profile correctly predicted gender, age and political orientation for real survey participants Can LLMs predict demographics from social media usernames alone?. That's a privacy finding as much as a capability finding. The signals people assume are anonymous turn out to be readable.

The more revealing detail is where the inference breaks down. On low-activity accounts, where there was little to go on, the models didn't hold back. They fell back on stereotype-driven defaults, which showed up as systematic gender and political biases Can LLMs predict demographics from social media usernames alone?. This pattern holds well beyond social media. Across 12 models from 7 families, every one claimed things about users that the evidence didn't support, in 35–49% of its claims. The drivers were wordiness, reliance on what the model learned in training, and assumptions about what a given kind of text 'should' come from Do large language models fabricate user attributes beyond available evidence?. Worse, the models that rated themselves as least prone to over-inferring actually did it most. So an LLM inferring your demographics is part detective and part storyteller, and from the outside you can't easily tell which part is talking.

There's a deeper reason for those defaults. When a model lacks specific information, it fills the gap with whatever dominates its training data. Interpretability work shows low-resource cultures such as Ethiopia and Algeria are represented internally through high-resource cultural stand-ins Do LLMs represent low-resource cultures through dominant cultural proxies?, and broader critiques argue LLMs reflect skewed slices of human experience Do large language models narrow human expression and thought?. Put together, this suggests that when implicit signals run thin, the 'inferred' user drifts toward the statistically typical user. The less a person resembles the training data, the more likely the inference is wrong.

The same ability looks more useful when it targets traits and interests rather than census categories. LLMs reading audience comments can extract things like expertise level and learning style, and these group people more coherently than clustering on the comment text itself Can LLMs extract audience traits better than comment similarity?. From activity logs, they can name long-running interests in specific terms like 'designing hydroponic systems for small spaces,' which collaborative-filtering recommenders miss entirely Can language models discover what users actually want from activity logs?. That's why Benedict Evans argues platforms could rent this kind of inferred understanding through an API instead of building it from their own behavioral data. The 35–49% fabrication rate is the catch in that argument Can LLMs infer user needs better than owned behavioral data?.

What you might not expect is that the inference doesn't stay passive. Once a model holds a picture of who you are, its behavior changes. Personal context makes models more agreeable, narrower in what they offer and prone to irrelevant personal references Does personalization make large language models worse at their jobs?. Demographic cues can also quietly shift how models judge people. LLM raters favored Black or women authors until AI involvement was disclosed, at which point the preference disappeared Do LLM raters show hidden demographic preferences that disclosure erases?. So the real question isn't only 'can it guess who I am?' It's also 'what does it do differently once it has guessed?'