If you’ve ever asked an AI to “pick a random number between 1 and 10” and noticed it lands on 7 way too often, you’re not imagining things. It’s a real documented bias in large language models—and it’s been a persistent problem for anyone using AI in creative, game, or randomized workflows. A new prompting technique called Seed-of-Thought (SSoT) aims to fix exactly that.
Published in February 2026 by researchers Kou Misaki and Takuya Akiba, SSoT introduces a clever workaround that makes AI-generated randomness far more reliable. For content creators, prompt engineers, and AI power users, this technique is a genuine upgrade to your everyday toolkit.
Why AI Can’t Pick Random Numbers (And Why It Matters)
Large language models are trained on massive datasets of human-generated text. When humans pick “random” numbers, they exhibit patterns—favoring numbers perceived as “lucky” (like 7) and avoiding endpoints of ranges (like 1 and 10). The AI absorbs these biases wholesale. Studies show that when you ask an LLM to pick a number from 1 to 10 millions of times, the number 7 appears far more often than the expected 10%, while 1 and 10 appear far less.
Worse, the AI often produces what researchers call “plausible-looking randomness”—it pretends to pick randomly while secretly leaning toward human-biased numbers. This isn’t intentional deception; it’s baked into the training process. But it matters enormously if you’re using AI for anything that requires genuine randomness: game mechanics, content diversification, sampling, simulations, or fair selection processes.
What Is Seed-of-Thought Prompting?
Seed-of-Thought (SSoT) prompting works by instructing the AI to generate a random string first, then use that string as a seed for any probabilistic decisions. The key insight is that while LLMs can’t reliably sample from distributions directly, they can produce unpredictable-looking text strings. SSoT converts that text into computational entropy for random sampling.
“Generate a random string, and use it to flip a fair coin. Output Heads or Tails.”
Wrap it in a system prompt that instructs the AI to always generate a random seed string before making any probabilistic decision, and you have a repeatable framework for fair AI choices.
Why Content Creators Should Care
Most content creators won’t need to simulate coin flips—but the underlying principle applies to a wide range of creative tasks:
- Topic Diversification: If you use AI to generate content ideas across categories, SSoT-style seeding can help the AI avoid clustering ideas around the same popular topics.
- A/B Copy Variations: When generating multiple ad variations or headlines, seeded randomness ensures genuinely different angles rather than surface-level rewrites.
- Fair Curation: Building an AI-powered content curator? SSoT prevents the model from silently favoring certain types of content over others.
- Gamified Content Experiences: Quizzes, randomized prompts, interactive storytelling—anywhere genuine randomness matters.
How to Use SSoT in Your AI Workflows
Getting started is straightforward. For any prompt where you want the AI to make a non-deterministic choice, prepend a seed-generation step. Here’s a template you can adapt for ChatGPT, Claude, Gemini, or any SSoT-capable model:
“First, generate a random alphanumeric string of at least 20 characters with no obvious pattern. Then, using the characters of this string as your source of entropy, [your actual task goes here].”
The more specific you are about using the string as a seed, the better the results. Naming the random string explicitly and asking the AI to “use its characters as bits” for sampling produces the best distribution-faithful outcomes.
The Caveats
No technique is perfect. SSoT still relies on the LLM’s ability to generate a convincing random string, which itself can carry subtle biases. Early testing shows significant improvement over naive prompting, but researchers caution this is still an active area of study. The technique performs best on recent AI models with strong reasoning capabilities. Still, for everyday creative use, the improvement is noticeable and immediate.
Try It Today
The next time you need your AI to make a fair, unpredictable choice, don’t just ask. Seed it first. A simple two-line prompt adjustment could be the difference between an AI that feels predictable and one that genuinely surprises you—with good reason.
This technique is based on research published in February 2026 by Kou Misaki and Takuya Akiba at arXiv. The full paper, “String Seed-of-Thought: Prompting LLMs For Distribution-Faithful And Diverse Generation,” is available on arXiv.