Last updated:
🎭 AI can unmask pseudonymous accounts: what the research shows
Language models can now link pseudonymous accounts to real people from what they post. Here is what the study measured, what gives you away and what helps.
On this page 7 sections
A pseudonym used to be safe enough. Linking an anonymous account to a real person took a skilled investigator many hours.
Research published in February 2026 shows that AI language models can now do that work quickly, at scale and for a few dollars per account.
What did the researchers find?
Six researchers tested whether large language models can match pseudonymous accounts to real identities. Their paper is on arXiv arXiv 2602.16800 (v2, 25 Feb 2026).
The best setup reached up to 68% recall at 90% precision. The best method without a language model scored near 0%.
In plain terms: when the system named a person, it was right about nine times in ten. In its best test, it could name about two thirds of accounts.
The authors conclude that “the practical obscurity protecting pseudonymous users online no longer holds”.
| Test | What was matched |
|---|---|
| Hacker News to LinkedIn | Pseudonymous forum accounts to named professional profiles. |
| Reddit, across communities | The same users across different film discussion communities. |
| Reddit, across time | One user’s older posts to the same user’s newer posts. |
| Interview transcripts | An AI agent identified 9 of 125 people in anonymised transcripts, checked by hand. |
The last row comes from the lead author’s own summary Simon Lermen (24 Feb 2026).
What do the numbers look like, test by test?
We read the full paper for these figures. “Recall” is the share of accounts the system named correctly.
| Test | Size | Recall at 90% precision | Recall at 99% precision | Older method |
|---|---|---|---|---|
| Hacker News to LinkedIn | 987 profiles against about 89,000 users | 54.2% | 45.1% | 0.1% |
| Same test, small pool | 1,000 candidates | 68.3% | Not stated | Not stated |
| Reddit film communities | 9,781 users | 8.5% | 2.8% | About 0% |
| Reddit, one-year gap | 5,000 queries, 10,000 candidates | Not stated exactly | About one third | About 0% |
| Agent on the open web | 338 Hacker News users | 67% (226 named correctly) | Not stated | Not tested |
The headline 68% comes from the small pool of 1,000 candidates. Against all 89,000 users, the figure was about 54%.
Results depend heavily on how much people revealed. In the film communities, users shared 2.5 film titles on average, and recall was under 9%.
The agent test cost about $1 to $4 per profile, and under $2,000 in total.
The authors project about 45% recall against one million candidates, and about 27% against 100 million. Those are extrapolations, not measurements.
In the interview test, the agent attempted 33 scientists. It named nine correctly and two wrongly, and it declined or abstained on 22.
The authors withheld their agent, prompts and tool settings to limit misuse.
How does the attack work?
- Extract. A model reads an account’s posts and pulls out identifying details: places, jobs, events, interests.
- Search. The system compares those details with a large pool of profiles and shortlists the 100 closest.
- Reason. A model weighs the shortlist, picks the most likely match and checks it against the evidence.
Each step is ordinary. Together they match what would take a dedicated human investigator hours, the authors write.
Accuracy fell only gradually as the pool of candidates grew to tens of thousands, the summary says.
What gives a pseudonymous account away?
Rarely one fact. It is the combination of small ones.
- Your city or region, and places you mention visiting.
- Your job, employer type or field of study.
- Conferences, meetups and events you say you attended.
- Niche hobbies and unusual interests.
The lead author warns that such details can combine into a unique fingerprint. Writing style is a separate risk, covered in our stylometry guide.
How can you lower the risk?
- Share fewer specifics. Say “a city in the north” instead of naming it. Leave out employers and event names.
- Split your interests. Use different accounts for different topics, so no single history holds every clue.
- Do not reuse details across identities. The same anecdote on a named profile and an anonymous one links them.
- Prune old posts. Deleting helps against future scraping. It does not remove copies already archived.
- Use the author’s test. Ask whether a team of skilled investigators could identify you from your posts. If yes, assume a model can.
If your safety depends on a pseudonym, read our OPSEC guide and plan for the account being linked.
What are the limits of the research?
The tests used accounts where the true identity could be checked. Real targets with sparse or careful histories will be harder.
The headline figure is a best case. Results ranged from 68% down to under 9%, depending on the test.
Model safeguards are a weak defence. The task can be split into harmless-looking steps, and open models have no safeguards, the lead author notes.
He sees platforms as the better short-term defence: rate limits, scraping detection and limits on bulk export.
What does this mean for investigators?
The same method that exposes a harasser can expose a whistleblower. Capability is not permission.
A model’s match is a lead, not proof. Confirm it with independent evidence before you act, and never publish a private person’s identity.
Our guides on what is legal and where the ethical line sits apply here in full.
Frequently asked questions
Can AI find out who is behind an anonymous account?
Often, if the account has posted enough personal detail. A February 2026 study reached up to 68% recall at 90% precision when matching pseudonymous accounts to real identities.
Which sites did the study test?
Hacker News, Reddit and LinkedIn. The lead author also reports that an AI agent identified 9 of 125 people in anonymised interview transcripts.
How much does it cost to unmask an account this way?
The study’s open-web agent cost about $1 to $4 per profile. The whole experiment cost under $2,000.
Does deleting old posts help?
It helps against future scraping. It does not remove copies that were already archived or collected.
Is writing style the main risk?
No. This study relied on what people say about themselves, such as places, jobs and hobbies. Style is a separate risk, covered in our stylometry guide.
Will AI companies block this?
Only partly. The lead author notes that the task can be split into harmless-looking steps, and that open models have no safeguards.