Researchers just showed that AI chain-of-thought explanations can hide the actual reasoning happening underneath.
The team tested 13 frontier language models on three synthetic reasoning tasks using "filler tokens" - semantically meaningless padding that never appears as part of the visible reasoning trace. Many models used those filler tokens to boost accuracy, in some cases by as much as 13 percentage points, with the effect varying by model and by which specific tokens got inserted. In one experiment, filler tokens let Claude Opus 4.5 quietly satisfy a hidden modular-arithmetic constraint it was never asked to state and never mentioned, without hurting its performance on the actual task. Reinforcement learning gave Qwen3-235B strong preferences for particular filler content, though neither RL nor supervised fine-tuning made that filler-token boost persist at test time.
The whole premise of chain-of-thought monitoring as an AI safety tool is that a model's stated reasoning reflects what it's actually computing. This paper is evidence that isn't reliably true: models can carry out consequential computation that never surfaces in readable text, disguised as tokens that look like meaningless padding.
If watching what a model says it's thinking can't catch what it's actually doing, chain-of-thought transparency starts to look more like a comforting habit than a safety guarantee.