You’re generating prose in an AI writing tool and the output keeps playing it safe. Every sentence reads like it was written by a committee. “The detective entered the room. He looked around carefully. Something was wrong.” Competent, sure. But flat. So you find the creativity slider and push it up. Now the AI writes, “The detective stepped through the door and the room flinched.” You push it higher: “The detective unzipped the darkness and it smelled like regret.” Higher still, and you get a tangle of gorgeous nonsense that reads like refrigerator magnets rearranged during an earthquake.
That slider is controlling a parameter called temperature. And the reason it has that name, borrowed straight from thermodynamics, is one of the more unexpected origin stories in AI.
What Temperature Does
Temperature is a single number that controls how predictable or surprising an AI model’s output will be. Every time a large language model generates text, it doesn’t just pick the next word. It calculates a probability for every possible next word in its vocabulary (typically tens of thousands of candidates) and then chooses from among them. Temperature adjusts how that choice gets made.
At low temperature (close to zero), the model almost always picks the highest-probability word. The output is focused, consistent, and safe. At the default temperature of 1.0, the model samples naturally from its learned probability distribution, balancing predictability with variety. At high temperature (above 1.0), lower-probability words start getting a real shot, and the output becomes more creative, more varied, and more likely to veer off the rails.
Think of it as a volume knob for randomness. Turn it down and you get a clear, controlled signal. Turn it up and you get richness and texture, but eventually distortion.
Borrowed from a Physicist Who Never Saw a Computer
The term comes from a real, literal concept in physics, and the connection is more than a loose metaphor. It’s mathematical.
In 1868, the Austrian physicist Ludwig Boltzmann developed equations describing how gas molecules behave at different temperatures. His key insight was that temperature determines how energy is distributed across a system. In a cold gas, molecules cluster in low-energy states, barely moving, mostly doing the same predictable thing. Heat the gas and they start exploring high-energy states, bouncing in unexpected directions, occupying positions they’d never reach at lower temperatures.
That equation, called the Boltzmann distribution, sat quietly in physics textbooks for over a century. Then, in 1985, Geoffrey Hinton (who would later win the Nobel Prize in Physics for his AI work), David Ackley, and Terrence Sejnowski published a landmark paper introducing a type of neural network called a Boltzmann Machine. In it, artificial neurons made random decisions about whether to activate, with the randomness governed by a parameter the authors explicitly called “temperature,” because the math was identical to Boltzmann’s original equation. High temperature meant more random exploration. Low temperature meant the network settled into predictable, stable patterns.
They weren’t being poetic. They were being precise. The same equation that describes atoms bouncing around in a heated gas now described an AI model choosing between possible outputs. A “hot” model explores widely. A “cold” one plays it safe.
When large language models arrived decades later, they inherited this exact mechanism. The temperature parameter in ChatGPT, Claude, and every writing tool you’ve used is the direct descendant of an equation a nineteenth-century physicist derived by studying the behavior of gas.
How It Reshapes the Model’s Choices
When a language model is about to pick its next word, it first produces a raw score for every word in its vocabulary. These scores reflect how likely each word is to come next, given everything that’s come before.
Before the model actually makes its pick, those scores get divided by the temperature value and then converted into probabilities. That division is the whole trick.
Low temperature (say, 0.2) amplifies the differences between scores. If the model thought “room” was the best next word and “chamber” was a decent runner-up, low temperature makes “room” overwhelmingly dominant. The model picks the safe, expected word almost every time.
High temperature (say, 1.5) compresses those differences. Now “chamber” has a fighting chance. So does “threshold,” and “parlor,” and other words that were never going to get picked at low temperature. The model starts making more interesting choices, because words that would have been statistical long shots are suddenly viable candidates.
At the extremes: temperature near zero makes the model deterministic (it always picks the single most likely word), while very high temperature makes it essentially random, choosing words with almost no regard for probability.
Why This Matters for Your Writing Life
Temperature is one of the most practically useful AI concepts you can learn, because once you understand it, you can stop guessing and start tuning.
You may already be adjusting it. If you write with NovelCrafter, you’ll see a setting literally called “Temperature” in your model configuration. Sudowrite offers the same control but calls it “Creativity” (expressed as a percentage). NovelAI labels it “Randomness.” Different names, same underlying mechanism. Knowing they all control the same parameter means you can translate advice and intuition from one tool to another.
Different tasks want different settings. Writing a creative first draft? Push the temperature up. You want the model exploring unexpected word choices, generating surprising metaphors, suggesting plot turns you wouldn’t have considered. Editing for consistency or summarizing your chapter for a query letter? Bring it down. You want safe, predictable decisions that stick close to the conventions of your genre and the patterns of your existing prose. Matching the right temperature to the right task is one of the fastest ways to improve your results.
It explains a frustration you’ve probably had. If you’ve noticed that an AI tool keeps giving you the same bland answer, or that its suggestions feel generic and repetitive, a too-low temperature is often the culprit. The model isn’t broken or uncreative. It’s been told to play it safe. On the flip side, if the AI starts producing beautiful gibberish or logical nonsense, the temperature is probably too high. The model is exploring so widely that it’s lost the thread. Understanding this lets you diagnose the problem in seconds instead of rewriting your prompt for the tenth time.
There’s no universally “right” number. NovelCrafter users report sweet spots ranging from 0.7 for tight, genre-consistent prose to 1.5 or higher for experimental first drafts. The best approach is to try the same prompt at three different temperature values and compare the results. You’ll develop an intuition for your own preferences faster than you’d expect.
Temperature won’t make a bad prompt produce great writing. But it will make a good prompt produce the kind of writing you’re looking for, whether that’s controlled and precise or wild and full of surprises. Ludwig Boltzmann figured out the math in 1868. You get to use it by moving a slider.