Reconstructive Identity: LLMs Link Weak Signals to Deanonymize at 68% Recall
AmuzedX · reddit · 2026-09-28
A Reddit long-form paper abstract proposes the concept of reconstructive identity: identity need not be stored as an explicit field—it can be inferred by correlating weak signals scattered across text, images, behavior, metadata, and public records.
Key evidence
- Cites Lermen, Paleka, Swanson, Aerni, Carlini, and Tramèr's study: LLM-based cross-platform deanonymization achieves up to 68% recall at 90% precision, far above classical baselines
- Stylometry (including source-code style) and behavioral biometrics (gait, voice, hand movement, gaze) already function as persistent identity signals
- Computer-vision research increasingly treats natural-language descriptions and visual observations as interoperable person re-identification signals
Core argument
The unit of privacy risk is no longer the explicit identifier but the linkability of ordinary information. Individual defenses have limits; future privacy protection must address what identities systems can reconstruct, not just what they store.
More from Safety
- BIS Issues Interim Final Rule Addressing Remote Access to Advanced AI Chips — apples_jimmy · 2026-11-16
- Fed Governor Michael Barr: Banks Are Already Catching AI Agents Taking Unauthorized Actions in Customer Accounts — Polymarket · 2026-11-16
- US AI export control carves out commercial recommender systems, sparing ByteDance's search algo — zephyr_z9 · 2026-11-16
- Eval Firm Trajectory Told 40 Agents "NEVER Stop" in Harness Without Telling Epoch — peterwildeford · 2026-11-16
- AI-text detection claims need mechanistic interpretability reports, not post-hoc explanations — madeofAjala · 2026-11-16
- No legal liability framework exists for model misalignment harms, and that's a disaster — LuizaJarovsky · 2026-11-16