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

https://theconversation.com/probably-doesnt-mean-the-same-thing-to-your-ai-as-it-does-to-you-275626

Publish Date: 2026-02-24 08:46:00

Source Domain: theconversation.com

Certainly! Here is a summary of the article with the requested key points using an unordered list:

Study Focus: The study highlights how large language model AIs like the ones behind chatbots do not interpret phrases indicating uncertainty (“probable,” “likely”) in the same way humans do.
Key Terms: The researchers focused on “estimative probability” words such as “maybe,” “probably,” and “almost certain,” finding that these words are mapped to different numerical percentages by AI and humans.
Divergence in Communication: While AI tends to match humans in interpreting extreme probabilities like “impossible,” it diverges significantly for hedge words like “maybe.” For example, AI often correlates “likely” with an 80% probability, whereas humans typically understand it closer to 65%.
Contextual Interpretation: Humans use contextual cues and personal experiences to infer meaning, while AI averages conflicting uses from its training data, leading to misalignments.
Bias Sensitivity: The study found that AI’s estimates tend to reflect gendered and cultural biases in the language used for prompts. For instance, changing pronouns in prompts affected the AI’s rigidity, and language changes altered the probability estimates.
Importance for AI Safety: This study underscores a fundamental challenge for AI safety and trust in human-AI interaction, especially in critical sectors like healthcare and government due to potential misunderstandings in communicated risks.
Current Research: There is further research into understanding large language models’ linguistic patterns and exploring methods like chain-of-thought prompting to reduce these misalignments.
Future Goals: Future AI developments aim to create models that not only predict words but genuinely understand and consistently convey uncertainty, ensuring reliability and trust in AI-assisted tasks.