‘Probably’ Doesn’t Mean the Same Thing to Your AI as It Does to You

Why It Matters

When AI systems miscommunicate uncertainty, decisions in high‑stakes fields can be misguided, eroding trust and raising safety concerns. Aligning linguistic expressions of risk is essential for reliable human‑AI collaboration.

Key Takeaways

• AI models misinterpret “probably” compared to human expectations
• Gap largest for hedge terms like “maybe” and “likely.”
• Gendered and language prompts shift AI probability estimates
• Misalignment threatens AI safety in healthcare and policy decisions
• Researchers suggest consistency metrics and chain‑of‑thought prompting

Pulse Analysis

The study highlights a fundamental blind spot in today’s conversational AI: the translation of vague probability language into precise numerical confidence. While LLMs excel at generating fluent text, they often average disparate usages of words like “likely” across billions of training examples, producing probability estimates that diverge from the human mental model. This discrepancy is most pronounced for hedging terms, where a model might label an 80% chance as “likely,” whereas most people would interpret the same label as closer to 65%. The researchers also uncovered that subtle changes in prompt phrasing—such as switching gender pronouns or language—can shift the AI’s internal probability mapping, exposing latent biases inherited from the data.

These findings have immediate ramifications for sectors that rely on AI‑driven risk communication. In medical diagnostics, an AI assistant describing a side effect as “unlikely” could be understating the true risk if its internal calibration differs from a clinician’s expectation, potentially influencing treatment choices. Similarly, policy analysts using AI‑generated forecasts may misread the severity of projected outcomes, leading to suboptimal resource allocation. The misalignment therefore threatens public trust and raises ethical questions about deploying LLMs without transparent uncertainty handling mechanisms.

Looking ahead, the research community is exploring solutions such as robust consistency metrics that enforce uniform word‑to‑probability mappings and chain‑of‑thought prompting that forces models to expose their reasoning. However, early results suggest that even advanced prompting does not fully bridge the gap. Continued interdisciplinary work—combining linguistics, cognitive science, and AI engineering—is needed to ensure that when an AI says “probably,” it truly means what users expect, paving the way for safer, more trustworthy AI assistants.