Published: 24 February 2026

What happened: A study published in the journal NPJ Complexity has found that large language models consistently misalign with humans when using words that express uncertainty — terms like “maybe,” “probably,” and “likely.” Researchers compared how AI models and humans map these words to numerical probabilities, finding significant divergences: models tend to agree with humans on extremes like “impossible,” but diverge sharply on hedging language, with a model using “likely” to mean an 80% probability where a human reader might assume around 65%.

Why it matters: The misalignment is more than a linguistic quirk — the researchers frame it as a fundamental challenge for AI safety and human-AI interaction. In high-stakes fields like healthcare, an AI describing a side effect as “unlikely” could lead a clinician to assume lower risk than the model’s internal calculation actually represents. The gap between what the AI means and what a human infers could directly affect clinical decisions.

Wider context: The study found that AI uncertainty estimates also shift with seemingly unrelated prompt changes: switching from “he” to “she” made probability estimates more rigid, while switching from English to Chinese produced different mappings altogether — both likely reflecting biases in training data. The researchers also tested chain-of-thought prompting as a potential fix, but found that even advanced reasoning did not reliably close the gap between statistical data and verbal labels.

‘Probably’ doesn’t mean the same thing to your AI as it does to you — The Conversation

Singularity Soup Take: Calibrated uncertainty is one of the most important signals in any high-stakes decision — and if the words AI uses to convey it systematically mean different things to human readers, that’s an alignment problem hiding in plain sight.

Key Takeaways:

• Probability Gap: LLMs agree with humans on extreme terms like “impossible” but diverge on hedging words — a model might use “likely” to mean 80% probability where a human reader assumes around 65%, a consistent and measurable discrepancy.
• Training Data Averaging: The researchers suggest LLMs may be averaging over conflicting usages of uncertainty terms across training data, rather than grounding them in context the way humans do, leading to systematic drift from human interpretation.
• Bias Sensitivity: Probability estimates shifted when prompts changed gender (he/she) or language (English/Chinese), indicating that uncertainty communication in LLMs is also shaped by biases embedded in their training data.
• High-Stakes Risk: The misalignment is particularly dangerous in healthcare, policy, and scientific reporting — domains where verbal probability cues inform real decisions and where a gap between AI intent and human interpretation could cause tangible harm.