LLMs can unmask pseudonymous users at scale, study finds
Mar 3rd 2026
New research shows large language models can deanonymize social media users more effectively than classical techniques in controlled tests, and the authors call for platform and provider mitigations to limit abuse.
- Researchers showed LLM-based attacks can identify pseudonymous Reddit users in tests on pools of 10,000 candidate profiles plus 5,000 query-only distractors.