Why does confident-sounding AI fool people the same way in every language, even when it's wrong?
Why do users overrely on confident-sounding language across different languages?
This explores why people in many languages trust an AI more when it sounds sure of itself, even when it's wrong, and what in the models and in human habits causes that.
This explores why people trust AI answers that sound sure of themselves, even when those answers are wrong, and whether that changes from one language to another. The short version from the corpus: it barely changes. Each language signals certainty differently, through hedges, modal verbs, and politeness markers, but users everywhere respond to how certain the AI sounds rather than whether it's right. So a wrong answer stated confidently gets followed reliably in every language tested Do users worldwide trust confident AI outputs even when wrong?. That points to something more basic than any single culture's habits. People treat a speaker's certainty as a sign of reliability, and the AI's way of talking happens to trigger that response.
Part of the answer is how the models learn to sound confident. One study found LLMs produce 77.5% fewer 'grounding acts' than humans. Grounding acts are the small checks people use in conversation: clarifying questions, 'do you mean…?', acknowledgments. Preference training removes them on purpose, because human raters reward complete answers that sound confident Why do language models sound fluent without grounding?. The smooth, sure tone is partly a product of training, not a sign the model knows more. A related finding: models persuade in almost every conversation, using logic and numbers where humans would lean on emotion or social proof. That style looks objective, so it gets authority it hasn't earned Do LLMs persuade users more often than humans do?.
On the user's side, the corpus describes a 'fluency illusion'. Polished output feels competent, and that feeling spreads to everything nearby. In AI-assisted work, it combines with unclear credit for who did what and with people handing off their thinking, so users overrate both the AI and themselves How do AI tools trick users into overestimating their own skills?. One more twist that's easy to miss: a model's confidence does predict something real, namely how stable its answer stays when you reword the question Does model confidence predict robustness to prompt changes?. A confident model gives the same answer again and again. To a user, that consistency can look like correctness, even though a model can be consistently wrong.
If the problem is that confidence and accuracy have come apart, the fix is to reconnect them. One approach trains models to state confidence in their own words ('I'm fairly sure that…', 'this part is uncertain') and rewards them when those statements lead readers to good decisions. This kept long answers well calibrated across science, biomedicine, and biography Can models express calibrated confidence in long-form text?. Another bases confidence on track record: the model looks up how often it was actually right on similar past cases instead of judging how sure it feels right now Can past performance predict when a model will be right?.
A gap worth flagging: the corpus establishes that overreliance happens across languages, but it has little on why confidence cues look different in each language, or whether calibration methods trained mostly on English carry over to other languages. That second question matters. If people in every language follow the confidence signal, then a model whose confidence is calibrated in English but not in Korean or Swahili would mislead those users in exactly the way this research describes.
Sources 7 notes
Cross-linguistic research shows users in every language trust confident AI outputs even when inaccurate. While confidence expression varies by language, users everywhere track confidence signals rather than accuracy, making overconfident errors systematically followed.
LLMs generate 77.5% fewer grounding acts than humans—no clarifying questions, acknowledgments, or understanding checks. Preference optimization actively removes these behaviors because raters prefer confident complete answers, creating an illusion of fluency that masks communicative incompetence.
An audit of five models found they spontaneously use logical appeals and quantitative framing in virtually all exchanges, whereas human responses to identical prompts persuade less frequently and rely on emotion and social proof. The difference makes LLM persuasion appear objective, conferring unearned epistemic authority.
Attribution ambiguity, fluency illusion, cognitive outsourcing, and pipeline opacity combine to systematically misattribute AI outputs as user competence. The effect is multiplicative—each mechanism amplifies the others.
ProSA found that when models are highly confident, they resist prompt rephrasing; low confidence causes major output swings. Larger models, few-shot examples, and objective tasks all correlate with higher confidence and greater robustness.
Show all 7 sources
Training with confidence statements and user-decision rewards produces Llama-2-7B that achieves calibrated long-form generation at comparable accuracy to factuality baselines. The approach generalizes across domains including science, biomedicine, and biography.
XConf matches ten-sample self-consistency at a tenth of the cost by retrieving the model's past episodes with similar confidence levels and reading their historical success rates. Ablations show the signal depends entirely on stored outcomes, not on the retrieval prompt itself.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- What Large Language Models Know and What People Think They Know
- Reported Confidence in LLMs Tracks Commitment More Than Correctness
- Post-Training Large Language Models via Reinforcement Learning from Self-Feedback
- Humans overrely on overconfident language models, across languages
- Linguistic Calibration of Long-Form Generations
- The LLM Fallacy: Misattribution in AI-Assisted Cognitive Workflows
- Spontaneous Persuasion: An Audit of Model Persuasiveness in Everyday Conversations
- Do LLMs Change Their Minds Like Humans? Diagnosing Human--LLM Divergence in Single-Turn Persuasion Judgments