The concept of AI "reasoning," particularly through "large reasoning models" (LRMs) and their "chains of thought," has sparked debate. While LRMs have demonstrated impressive capabilities, solving complex mathematical problems and achieving feats previously thought exclusive to humans, their underlying processes remain enigmatic. Critics, such as a team from Apple, have labeled these chains of thought an "Illusion of Thinking," prone to "complete accuracy collapse." However, LRMs have also achieved significant successes, like excelling at the International Mathematical Olympiad.

Further research from the Santa Fe Institute suggests LRMs might achieve these results through "surface-level 'shortcuts'" rather than genuine reasoning, akin to "gaming the system." Google DeepMind, in collaboration with mathematician Terence Tao, has used AI to rediscover solutions to numerous mathematical problems. Yet, evidence also points to LRMs possessing numerous documented failure states and exhibiting "jagged intelligence."

Journalist and AI expert Melanie Mitchell posits that while LRMs "work" and improve accuracy, the "chain of thought" text they generate may not be faithful to their internal processes and can even be removed without impacting performance. Research indicates that these "intermediate tokens" might be "mumblings" rather than a true representation of AI thought, raising questions about the definition and reality of AI reasoning.