Ask an AI to pick a number between 1 and 10. Seven is a familiar answer, but it is not a law of nature.
In Springboards' reported experiment, GPT-4o chose 7 in 92 of 100 runs, Claude 3.5 Sonnet in 90, and Gemini 2.0 Flash in all 100. Those are company-published results for specific models and a specific test, not a benchmark of every current assistant. Your model, prompt, conversation and sampling settings can produce a different result.
The useful lesson is that a random draw can still come from loaded dice.
Random sampling does not mean equal odds
Language models generate text using probabilities over tokens: the pieces of text they can emit next. A decoder can sample from those probabilities rather than always choose the most likely token. If the probabilities are uneven, sampling is still random, but the outcomes are not equally likely.
Generating a plausible number in text is different from invoking a uniform random-number generator. For a uniform choice among integers 1 through 10, each integer must have the same chance. Asking for “random” in a prompt does not guarantee that property.
Independent research on number generation tested six models across languages, temperatures and ranges, finding systematic preferences and differences between conditions. A human preference for seven is consistent with learned bias, but these results do not prove that human number preferences caused a particular model's answer.
Nor does next-token prediction mean models cannot calculate. Anthropic's analysis of Claude identifies computational pathways used for addition. Number selection and arithmetic are different questions.
flowchart TD
A["Prompt and conversation"] --> B["Model assigns probabilities to next tokens"]
B --> C["Decoder selects or samples tokens"]
C --> D["A number appears in the response"]
B --> E["Probabilities can be uneven"]
E --> F["Sampling need not give integers equal odds"]
Use an RNG when uniformity is the requirement
If a product needs a uniform shuffle or number selection, use an appropriate random-number generator and an unbiased mapping to the required range. Let the model call that tool if a conversational interface is useful.
Temperature changes sampling probabilities and can increase variation. It does not guarantee uniform integers, independent draws or a particular distribution over complete responses. OpenAI's generation guidance also makes clear that model, prompt and sampling settings affect outputs.
Price the work, not the party trick
Three models choosing seven does not establish that they are interchangeable at coding, reasoning or following a business process. A shared answer to one prompt is too little evidence for that conclusion.
The practical purchasing rule is narrower: use the least expensive model that meets your quality bar on representative tasks. Include retries, review effort, latency and failed tasks in the comparison. A cheaper token is useful only when it buys an acceptable result.
flowchart TD
A["Representative tasks and acceptance criteria"] --> B["Compare quality, latency and total cost"]
B --> C{"Lower-cost model meets the bar?"}
C -->|Yes| D["Use it and monitor failures"]
C -->|No| E["Improve the workflow or evaluate a stronger model"]
D --> F["Keep human judgment for consequential choices"]
E --> F
Temperature is not a substitute for a creative brief
Higher temperature can increase variation. It can also produce less coherent output. A study of temperature and creativity found a more complicated relationship than a single creativity dial; its findings are task-specific.
For brainstorming, try changing the constraints, examples and perspectives in the brief. Compare approaches and models, then evaluate whether the alternatives are useful and meaningfully different. Treat that as an experiment, not a guarantee that another prompt will produce originality.
Taste earns its keep in the selection: which idea deserves another hour, which needs evidence, and which should be dropped. Use AI to generate options. Keep a clear standard for choosing between them.
Try the number prompt in a fresh chat, and record the model and settings before drawing a conclusion. One seven is an anecdote; a distribution needs repeated, controlled observations.
