Study (n=504): suspicion doesn't improve AI-text detection; fake-news accuracy drops 10.2 points
bit3py · reddit · 2026-09-02
A human-subject study (n=504 participants, 2,438 judgments) had participants classify news fragments along two axes: origin (human vs machine) and veracity (real vs fake). Preprint is open access (arXiv).
Three surprising findings:
- Perception-accuracy gap: more suspicious participants were not better at detecting machine-generated text — "just be more skeptical" media-literacy advice doesn't translate into accuracy
- Modern LLM output was frequently indistinguishable from human text
- Asymmetric cognitive fatigue: under sustained exposure, fake-news detection degraded by 10.2 percentage points while AI-origin detection stayed roughly stable — the two judgments draw on different resources, and only one wears out
The study adapts the cybersecurity kill chain, framing disinformation as a staged lifecycle and arguing for earlier interventions instead of relying on a tired human at the end of the chain. The author notes that platform designs forcing people to evaluate more content could quietly worsen veracity judgment.
More from Safety
- One diagram explains the difference between CSRF and XSS attacks — _jaydeepkarale · 2026-10-18
- Grady Booch slams Anthropic for tolerating up to 50 deaths and $1B damages per model — Grady_Booch · 2026-10-18
- Single Neuron Suffices to Bypass Safety Alignment in LLMs, Paper Finds — burny_tech · 2026-10-18
- Polymarket gives OpenAI just 8% odds of a full training pause by end of 2026 — Polymarket · 2026-10-18
- OpenAI engineer over alleged hidden recording device seeks bail, denies Saudi escape plan — Polymarket · 2026-10-18
- Free comic "The Collective" offers a strange retelling of the Hugging Face hack — Flugelwagen · 2026-10-18