OPEN LINK ↗
// 1h agoRESEARCH PAPER
AI-Text Detectors Break Across LLM Generations
A new paper finds that detectors trained on older LLM versions caught over 99% of scientific-text rewrites before a model-generation boundary, but only 3.8% afterward. The study analyzed 4,000 PNAS abstracts rewritten by 23 models from three vendors.
// ANALYSIS
AI-text detection accuracy can collapse silently as models evolve, making static detector benchmarks dangerously misleading.
- –Detectors may learn model-specific vocabulary and stylistic fingerprints rather than general signs of machine-generated writing
- –Updating for every new model can preserve recall, but broad coverage may sharply increase false positives
- –A commercial detector also missed most rewrites from a newer model while rarely flagging human text
- –Research-integrity workflows should revalidate detectors after model releases and treat scores as evidence, not proof
// TAGS
llm-turnoverllmevaluationbenchmarkresearchethics
DISCOVERED
1h ago
2026-10-10
PUBLISHED
2h ago
2026-10-10
RELEVANCE
8/ 10
AUTHOR
rohanpaul_ai