Part 9 - What is Temperature in LLMs? Why Every QA Must Know It | AI Testing

Temperature in LLMs explained for QA engineers: why an AI gives different answers to the same question, and what testers must ask the developer before writing a test plan. Ask an AI “What is the capital of France?” a few times and you may get “Paris”, then “Paris, a beautiful city in Europe”, then something longer. Is that a bug? Not always. That variation is controlled by a model parameter called temperature, and it changes how you should test AI features. In Part 9 of the AI Testing series, we explain temperature with a simple gas-stove knob analogy and connect it to real testing decisions. It builds on Part 7 (Hallucination) and Part 8 (Knowledge Cutoff), so watch those first if you haven’t. 📌 What you’ll learn: ✅ What temperature is: a knob that controls variability/randomness in the AI’s output ✅ The gas stove analogy (low flame vs high flame) ✅ Low temperature (around 0): predictable, consistent answers ✅ High temperature (closer to 1 or higher): more varied answers, with the exact range depending on the model/API ✅ How it works with token-selection probabilities (70/20/10 vs 40/35/25) ✅ The locator example: “best way to fix a locator issue?” gets the same answer at low temperature and different but valid answers at high temperature ✅ Why high temperature doesn’t make the AI smarter, only more varied ✅ Why you won’t find temperature in the UI of ChatGPT, Claude or Gemini, since it’s set through the API by the developer ✅ QA perspective: the question to ask your developer, “What temperature is set for this feature?” ✅ Factual features (refund policy) should stay consistent, while creative features (funny welcome message) are expected to vary 💡 Key takeaway: Don’t assume anything from the UI. Confirm the temperature setting with the developer, because it decides whether a changing answer is a bug or expected behavior. 🔔 Subscribe for the full AI Testing series. 👍 Like the video if it helped, and 💬 comment: have you ever seen an AI give different answers to the same question? #LLMTemperature #AITesting #LLMTesting