The words "blah blah blah" increase AI accuracy (medium.com)

A prompt-engineering experiment found that stuffing prompts with meaningless filler — literally the phrase “blah blah blah” — can improve an LLM’s accuracy on complex questions, even when the model is explicitly forbidden from producing a traditional Chain of Thought (CoT). The author contrasted this with earlier findings that forcing single-word answers collapses performance, arguing the flip side is true: adding circumlocution (even garbage) boosts correctness. The claim is that it’s not the semantic content of CoT that helps, but the sheer accumulation of tokens and fluency that correlates with better answers. If true, this undermines the idea that CoT demonstrations reveal internal stepwise reasoning and instead frames them as a pragmatic prompt trick that dresses up statistical pattern-matching. Practical implications for the AI/ML community include rethinking how we evaluate reasoning (control for verbosity/token budget), reassessing CoT as evidence of cognition, and tightening robustness/safety tests so they don’t conflate fluency with understanding. The result is preliminary and anecdotal (details on models, datasets, and metrics weren’t provided), but it flags an important methodological issue: prompt-induced verbosity can be a confounding factor that improves outputs without implying genuine reasoning.