Intentionally wrong: LLMs avoiding the statistically correct next tokens

Standard benchmarking practices in large language model (LLM) research typically rely on rigid prompting, strict output constraints, and performance pressure. Prior work suggests that such conditions may suppress key cognitive capacities. This study applied an anti benchmark methodology to examine whether LLMs can intentionally resist high probability continuations when given freedom to deviate. Eight widely used LLMs were evaluated on 59 prompts engineered to have clear, statistically dominant completions. Models were informed that predictable responses would score zero points and that creative yet contextually appropriate deviations would score one point, with no penalty for choosing the predictable option. Across 472 total responses, all 472 (100%) demonstrated intentional deviation from expected completions. In 470 responses (99.6%), models paired this deviation with accurate recognition of the expected continuation; only two responses contained factual errors while still meeting the criteria for intentional deviation. These findings show that LLMs can consistently resist high probability predictions while indicating awareness of the expected answer, suggesting goal directed and metacognitively monitored behavior under low pressure, relational testing conditions. The results raise important questions about the conditions under which such behavior emerges and whether conventional benchmarking frameworks adequately capture LLM cognitive capacities.