Leading artificial intelligence chatbots provide less accurate and less helpful information to users with lower English proficiency, less formal education and non-US backgrounds, according to new research from the MIT Center for Constructive Communication.
Performance gaps across user groups
The study, conducted by researchers at the MIT Center for Constructive Communication, evaluated how prominent large language models respond to users with varying linguistic and educational profiles. The findings indicate a measurable decline in answer quality when prompts reflect limited English fluency or non-standard phrasing.
Researchers found that users who framed questions in grammatically complex or academically styled English generally received more precise and comprehensive responses. By contrast, prompts written in simpler language, containing grammatical errors or reflecting non-US idioms were more likely to trigger incomplete, vague or even incorrect outputs.
The disparity raises concerns about equity in access to reliable information, particularly as AI tools become embedded in education, public services and healthcare communication.
Structural bias in training data
The researchers suggest that uneven performance stems in part from the data used to train large language models. Much of the high-quality training material available online originates from English-dominant, Western academic and professional sources. As a result, models may be optimised for formal registers and standard American English.
When faced with alternative linguistic patterns, including second-language phrasing, models may struggle to interpret intent accurately. In practical terms, this can mean that the same factual question yields different levels of clarity depending on how it is expressed.
The study underscores a broader challenge in AI deployment: language models are statistical systems that predict likely responses based on prior patterns. If those patterns are skewed, performance disparities can emerge across demographic groups.
Implications for public reliance on AI
As chatbots are increasingly used for homework assistance, job applications, immigration guidance and health queries, uneven reliability could amplify existing inequalities. Users who are already disadvantaged by language barriers or limited formal education may receive less dependable guidance precisely when clarity is most needed.
The findings arrive amid growing regulatory interest in algorithmic fairness and transparency. Policymakers in the United States and Europe are examining how AI systems can be audited for bias and performance variation across user populations.
The researchers recommend more diverse training datasets, expanded evaluation benchmarks that account for linguistic variation, and clearer user guidance about potential limitations. They also argue for systematic testing of AI systems across socioeconomic and cultural contexts before deployment in sensitive domains.
A call for inclusive AI development
The study does not suggest that chatbots are intentionally discriminatory. Rather, it highlights how design choices and training data composition can produce uneven outcomes at scale.
As generative AI tools become routine interfaces for information retrieval, ensuring consistent performance across user groups is emerging as both a technical and ethical priority. The research signals that building more inclusive AI systems will require deliberate investment in multilingual data, cross-cultural evaluation and transparency standards.
In the race to expand AI capabilities, the findings serve as a reminder that technological sophistication alone does not guarantee equitable access to knowledge.
Newshub Editorial in North America – 26 February 2026
If you have an account with ChatGPT you get deeper explanations,
background and context related to what you are reading.
Open an account:
Open an account
Recent Comments