LLM Creativity Test Finds Models Stilted

The test is a sharp provocation, but its value depends on whether the prompts, scoring, and model settings are reproducible.

• –Creativity benchmarks can measure novelty, usefulness, fluency, or diversity—and produce very different rankings depending on the definition.
• –Generic prompts often reward familiar patterns, exposing models’ tendency toward safe, statistically common answers.
• –Temperature, system prompts, sampling count, and model selection can materially change results.
• –A credible follow-up should publish the exact test, outputs, scoring rubric, and human baseline.
• –The broader takeaway is practical: use LLMs to expand creative search, then rely on humans to select and refine genuinely original ideas.