This article introduces a distinct shift in how we prompt large-language models: moving beyond craftily worded instructions toward a method called verbalized sampling. The author argues that traditional prompt engineering (e.g., few-shot examples, chain-of-thought) often guides the AI into predictable responses — because the model defaults to high-probability outputs shaped by its training. Verbalized sampling, instead, asks the model to generate multiple alternatives, each with its own probability score, and then selectively sample from the lower-probability tail of the distribution. This approach unlocks more creativity, variety and unexpected insights instead of simply steering toward the “safest” answer.
In practical terms, the workflow proposed is: ask the AI to output several candidate responses, annotate each with a likelihood or “probability” indicator, and then instruct the system (or user) to pick from across the spectrum — including lower-probability options that might be less polished but richer in novelty. The article gives examples such as creative writing, brainstorming new business angles or crafting novel metaphors, where this diversity adds value. The key insight is that the model already knows many valid answers but conventional prompting channels it toward the most generic one; verbalized sampling lets you dig into the model’s “hidden space” of ideas.
The author also highlights trade-offs and contexts: for tasks where accuracy, compliance or reproducibility matter (e.g., legal summaries, medical advice), the method may require coupling with verification approaches (like consensus prompts or chain-of-thought refinement) to avoid “creative but wrong” outputs. However, the technique is especially useful for ideation, design, narrative generation or any domain where novelty and diversity matter more than single-best-answer correctness. The article argues that as AI models become more capable and aligned toward safe responses, we must intentionally dial up the “unusual” outputs if we want innovation rather than repetition.
Finally, the piece concludes that verbalized sampling signals a future for prompt-engineering where instead of just telling the model what we want, we open up the model’s internal possibility space and collaborate with it more like a creative partner. The author suggests a hybrid prompt-architecture: start with verbalized sampling to explore widely, then refine with chain-of-thought or consistency prompts, then verify as needed. In doing so, teams can move from “reacting to AI outputs” to “guiding AI exploration” — a shift that may matter especially for organisations wanting to harness generative AI beyond automation into genuine creativity and strategy.
