AI tools can unmask anonymous accounts

Researchers from ETH Zurich, Anthropic, and the Machine Learning Alignment and Theory Scholars program developed an automated AI system capable of reidentifying anonymized online accounts by searching for patterns in writing style, posting frequency, and biographical details across vast datasets. This system significantly outperforms traditional computational deanonymization methods, achieving up to 68 percent accuracy with 90 percent precision in testing datasets drawn from public sources like Reddit and LinkedIn. The study cautions that while the ability to identify users pseudonymously is not new, the end-to-end automation and low cost (under $2,000 for the experiment) drastically lower the barrier to entry for such surveillance, posing risks to journalists and activists. However, experts note that the technology is not perfect, citing the continued anonymity of Satoshi Nakamoto as evidence, and emphasize that basic privacy precautions remain effective. The researchers urge AI labs to build safeguards and social media platforms to restrict mass data extraction.

(Source：The Verge)