Abstract

Generative AI is quickly becoming a learning companion in statistics education. Many learners already upload a textbook to ChatGPT or a similar tool and ask for examples and exercises tailored to them. What they get back depends on the training data of the model. These data over-represent dominant languages, cultures and role models, and larger models have been shown to produce more stereotypical output even when they are told not to (Siddique et al., 2024). Bias in online learning is not new. With LLMs it may become bigger and also harder to see, because the output looks personalised.

Existing research points to several groups for whom this matters. For Indigenous learners, Cook et al. (2025) state that there is no framework yet to enhance Aboriginal and Torres Strait Islander Peoples’ data literacy. For adults without an academic background, numeracy is closely linked to education and social background, and first-generation students worry most about the demands of self-directed learning, which is how most learners use LLMs. Findings on gender differences in statistical literacy are mixed, but differences in statistics anxiety and self-concept are well documented. In all these cases, the problem is often a mismatch between the learner’s context and the context the material assumes, more than a lack of ability. The same mismatch affects many people in low- and middle-income countries, where most teaching material and most training data of LLMs come from elsewhere.

We ask whether systems built for hyper-individual training can reduce this mismatch instead of amplifying it. Our contribution is based on two lines of work. The first is AIRIES, a publicly funded R&D project at STAT-UP. It combines a machine-readable competence standard based on IEEE P7015 with a knowledge graph and LLM-based generation of learning content from curated material. By design, it models alternative learning paths instead of one target profile. The second is our collaboration with the ISI Education Working Group since 2024 on an AI-supported training in ethics for statisticians. We plan to offer it especially to people who work with statistics as part of their job but have limited access to training, for example in national statistical systems, public administration or NGOs in low- and middle-income countries.

From this, we derive design principles for bias-reduced statistics education: grounding generated content in sources that the communities concerned endorse, letting learners define their own context, allowing different but equally valid learning paths, checking generated exercises systematically for bias, and co-design with the groups concerned. We also present the design of a planned evaluation with two levels. The first is an audit of generated learning material: the same content is personalised for learner profiles that differ in only one attribute, and the outputs are compared for stereotypes, difficulty and fit to context. The second compares learning outcomes, statistics anxiety and perceived relevance between generic and context-grounded personalisation, asking whether gaps between groups become smaller or larger. We invite the statistics education community to discuss how such evaluations should be done.