Test Shows LLMs Can't Fake Random Vocabulary: Shuffling Fails for Words
ctjlewis · x · 2026-07-31
Following up on the flaws of LLM randomness, the author tested ways to evade AI text detectors like Pangram.
Experiments show that while the trick of generating random numbers and then shuffling them can fool detectors, the same shuffling technique does not work for words. This is because the specific word choice in LLM generation inherently reveals strong model signatures, and merely rearranging the word order cannot mask its built-in expression habits.
More from Research
- LMSYS Releases Inkling-small: 12B Active Parameters Lowers Multimodal RL Barrier — simonguozirui · 2026-07-31
- CuspAI Co-founder on Using AI to Imagine and Create New Materials — jonfildes · 2026-07-31
- Project HQ Announces Capture of Over 1 Trillion Tokens — jacob_posel · 2026-07-31
- MolmoAct Robotics Project Announces Open Source Release with Community Contributions — DJiafei · 2026-07-31
- Senior SWE-Bench Update: Three-Way Tie at First Place with Reduced Variance — ajratner · 2026-07-31
- Top ML Conferences Receive 100k Papers Yearly, But Can We Name 10 Greats? — andrewgwils · 2026-07-31