AI can be astonishingly helpful—and occasionally confidently wrong. From made‑up facts to invented quotes, those “hallucinations” don’t come out of nowhere. New research points to two big drivers: reward systems that nudge models to guess, and the way we talk to them.
A study released October 3 on arXiv, titled “Mind the Gap: Linguistic Divergence and Adaptation Strategies in Human-LLM Assistant vs. Human-Human Interactions,” suggests users themselves often set the stage for unreliable answers. After examining more than 13,000 human-to-human exchanges and 1,357 real conversations with AI assistants, the authors found people communicate very differently when they message a chatbot. Inputs to AI were shorter, less grammatical, less polite, and used a narrower vocabulary—yet they carried nearly the same information as human-to-human messages. In other words, what people say doesn’t change much; how they say it does.
The researchers call this a style shift. Large language models are predominantly trained on well-structured, courteous text. When the tone suddenly becomes terse, blunt, or sloppy, the model is more likely to misread intent, fill in gaps, or fabricate details. That’s where hallucinations creep in.
The team analyzed six linguistic dimensions, including grammar, politeness, vocabulary range, and information content. They found:
– Grammar and politeness scores were more than 5% and 14% higher, respectively, in human-to-human conversations.
– Despite the rougher delivery to AI assistants, the actual information shared was nearly identical.
This matters because language models also respond to incentives. Separate work from OpenAI researchers highlights how reward mechanisms—used to train models to be helpful—can unintentionally push them to produce confident guesses. Combine that with unclear or abrasive prompts, and the risk of incorrect outputs goes up.
What can improve reliability
– Train models on broader language styles: Style-aware fine-tuning helped models better interpret user intent by at least 3% in the study.
– Be cautious with automatic paraphrasing: Rewriting user input on the fly slightly hurt performance, likely because emotional cues and context were lost.
– Encourage clearer prompts: When users write in complete sentences, use proper grammar, and maintain a polite, consistent tone, models are less likely to stray into fabrication.
Practical tips to reduce AI hallucinations right now
– State your goal in one or two clear sentences.
– Include key details and constraints; avoid vague references.
– Use complete sentences and standard punctuation.
– Keep a consistent, respectful tone.
– Ask for sources or reasoning steps if you need verification.
– If the topic is complex, break the task into smaller questions.
What builders should consider
– Adopt style-aware training and evaluation so models handle terse or unpolished inputs without overconfident guessing.
– Calibrate reward functions to discourage speculative answers when uncertainty is high.
– Offer user-facing guidance and prompt templates that encourage clarity without stripping away important context.
Bottom line: AI doesn’t hallucinate in a vacuum. Training incentives and user style both play a role. By making prompts clearer and more courteous—and by tuning models to handle diverse communication styles—we can meaningfully reduce fabricated responses and get more reliable results from modern language models.
