How prompts manipulate models: measuring cognitive bias vulnerability in commercial LLMs across languages

In plain words

Ask a chatbot a question two ways — once plainly, once with a nudge like a leading assumption or a confirmation-seeking frame — and you often get two different answers. This paper measures exactly how much that nudge moves commercial large language models, and whether the effect grows when the conversation is not in English.

The study evaluated six commercial LLMs in English, Hindi and Bengali, using carefully verified Hindi and Bengali translations of three established bias benchmarks, over a total of 683,616 model responses.

The headline finding: confirmation-pattern prompts cut neutral answers by 19.42 percentage points on average, with neutral-response rates ranging from 74.11% to 94.54% across models — and vulnerability grows in Hindi and Bengali relative to English.

How it works

1. Translate the benchmarksThree established cognitive-bias benchmarks were translated into Hindi and Bengali, with verified translations, so all three languages test the same thing.
2. Frame the promptsEach question is asked both neutrally and through manipulative framings such as confirmation-seeking patterns.
3. Query at scaleSix commercial LLMs answer every variant, producing 683,616 responses for analysis.
4. Measure the tiltThe drop in neutral answers between neutral and manipulative framings quantifies each model's bias vulnerability, per language.

Why it matters

Hundreds of millions of people now ask LLMs for advice in their own language. If a model can be steered by how a question is phrased — and steered more easily in Hindi or Bengali than in English — that is a fairness and safety gap affecting exactly the users the AI industry tests least. Measuring the gap is the first step to closing it.