GPT Guesses Between 1 and 100
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- GPT-4.1 was queried 10,000 times through the OpenAI Responses API at temperature 1.0, each call requesting a single integer between 1 and 100.
- Methodology employed a fixed system prompt and a unique UUID per request, then cleaned the responses to retain only valid integers within the target range.
- Baseline for the study is a uniform distribution, allowing a chi‑square goodness‑of‑fit test to detect deviations from true randomness.
- Limitations note that the experiment covers only this single model, is prompt‑ and temperature‑sensitive, and does not reflect the consumer ChatGPT product or other LLMs.
Why it matters: AI developers gain concrete data from 10,000 GPT‑4.1 samples, revealing token‑sampling bias that could affect the reliability of stochastic features in their apps, prompting tighter calibration or design adjustments.
