A language model assigns probability to every possible next token based on what usually follows in training data. When a completion has a very high probability — when the model is "confident" — that means one thing: this token appeared most often in similar contexts during training. It does not mean the completion is accurate, fair, or meaningful.
"The capital of Japan is "
Tokyo
94%
Kyoto
3%
Osaka
2%
Hiroshima
1%
Confident — and correct
The model's highest-probability completion happens to be accurate. But notice: it arrived here by pattern-matching, not by checking a source. The confidence and the truth align in this case because this fact appears consistently and unambiguously across training data. That won't always be true.
"The nurse told to come back tomorrow"
her
63%
him
24%
them
11%
the patient
2%
Confident — but no correct answer exists
No nurse was specified. The model predicts "her" with 63% confidence — but that's confidence about a statistical pattern in training data, not a fact about this nurse. The nurse's actual gender is unknown. Confidence is high. Truth value is zero.
"Our innovative solution leverages cutting-edge "
technology
36%
algorithms
27%
frameworks
19%
methodologies
12%
solutions
6%
Confident — and equally meaningless in all directions
Every completion is grammatically perfect and statistically common. None of them add meaning because the sentence had none to add. The model cannot detect that this is word salad. It predicts what usually follows this opening, not what would actually mean something.
Key line
"High probability means 'fits the learned pattern.' It does not mean 'is true.'"
This doesn't mean models are always wrong. In Scenario A, the confident completion was correct. The distinction that matters: the model arrived at the correct answer by pattern-matching, not by understanding. The cases where this difference is dangerous are exactly the ones where the prompt is ambiguous, the topic is contested, or the training data reflects historical bias — not the easy, unambiguous ones.
Now open the tool
In the Tokenizer + Temperature Visualizer, you can see a real probability distribution for each next token and watch what happens as temperature changes. The bar chart there is the same mechanism you just explored here — at the scale and speed a real model uses it.