AI 'Reasoning' Works — But Its Chain-of-Thought May Be Meaningless Filler

Large reasoning models (LRMs) have racked up genuine wins: an OpenAI model one-shot a famous open math problem in May 2026, LRMs took gold at the International Mathematical Olympiad, and a DeepMind–Terence Tao collaboration improved solutions to 67 problems across several fields. Yet a parallel stream of research keeps puncturing the idea that these systems reason the way they appear to. The article traces the resulting whiplash and lands on a sharper question than ‘can AI reason?’ — namely, what the visible chain-of-thought actually has to do with the answer.

Melanie Mitchell frames the tension in three parts: the technique measurably improves accuracy; the emitted ‘thinking’ text isn’t necessarily faithful to what happens inside the model; and much of that text isn’t even load-bearing and can be deleted. Supporting work piles up. Kambhampati’s lab showed swapping correct reasoning traces for wrong or irrelevant ones didn’t hurt performance, and that models trained only on correct traces still emit invalid ones. An NYU paper found strings of dots — ‘meaningless filler tokens’ — can stand in for a human-readable chain of thought. A Northeastern/Berkeley study found 30–60% of thinking steps in frontier open-source LRMs had minimal causal impact on the output.

The upshot is that reasoning traces — the feature meant to separate LRMs from ordinary next-token LLMs — are often neither faithful accounts of the model’s internals nor causally necessary to its answers. Researchers warn against reading them as auditable logs of machine ‘thought’; Kambhampati’s ICML 2026 position paper bluntly urges the field to ‘stop anthropomorphizing intermediate tokens.’ This matters well beyond philosophy: if chain-of-thought is unreliable as a window into model behavior, it’s a weak basis for interpretability, safety auditing, and trust in high-stakes deployments — even as the systems keep delivering real results.