An easy-read version of this project stance document can be found here.
Summary
In its broadest sense, artificial intelligence (AI) refers to the capacity for technological tools and automated systems ‘to perform tasks commonly associated with intelligent beings’ (Britannica 2026), including emulating and potentially outperforming human cognitive processes. Conversational or chatbot AI interfaces trained on large language models (LLMs), such as ChatGPT, CoPilot, Gemini, Claude, etc., are here referred to more specifically as generative artificial intelligence (genAI),[1] and are the focus of this document.
First and foremost, Multilingual Mentoring advocates for critical literacy in relation to the use of genAI—whether or not individuals choose to engage with such tools, we have a collective responsibility to educate ourselves and to ensure that our decision-making is as informed, effective, and ethically responsible as possible. Such criticality is particularly crucial in light of research (Tully et al. 2025) finding that reduced AI literacy corresponds directly to increased AI receptivity—in other words, those who are less informed about the capabilities, limitations, and associated risks of genAI tools are more likely to use them, and to do so uncritically.
This document outlines a few key benefits and risks of using genAI so that readers may make their own informed and critical choice. We do not specifically prescribe or proscribe the use of such tools for our mentors, for language teachers, or for other stakeholders. However, we as a project team systematically adopt a policy of critical caution, and do not rely on the use of genAI in relation to our internal work processes, external communications, or for any other workstreams. Wherever genAI tools are used to support the work that we do, this will be clearly stated and cited appropriately (see the end of the present document, for example).
Benefits
We recognise that genAI tools can offer significant benefits in terms of production, efficiency, accessibility, and multilingual engagement. However, these potential advantages are only realised when such tools are used thoughtfully, critically, and with an awareness of their limitations and risks.
This sections outlines some key benefits:
Idea generation
GenAI tools can aid users in generating initial ideas, exploring alternative perspectives on existing ideas, or overcoming creative blocks. This may be helpful when planning sessions or developing resources. However, AI-generated content must always be reviewed critically and should undergo a human editing process, rather than being accepted uncritically.
Improving efficiency
GenAI tools can reduce time spent on straightforward, repetitive, or administrative tasks, potentially freeing up time for more important and valuable tasks, e.g., those that require higher-level, executive decision-making, or human creativity.
Enhancing accessibility
GenAI tools can assist with the simplification of complex texts, adaptation of materials for different reading levels, or the creation of alternative formats. With due care, they can therefore be used to make information more accessible to a wider range of learners, i.e., scaffolding.
Facilitating multilingual engagement
For projects such as Multilingual Mentoring or for language classrooms more broadly, genAI tools may assist with drafting or adapting content across languages, helping users engage with multilingual audiences and experiment with language use. It is important to be aware, however, that the quality of translated content is variable depending on the language in question. Minority languages such as Welsh are less well represented in LLM training data, meaning the quality of Welsh machine translations are less reliable. Outputs should be checked by competent speakers, particularly for accuracy, nuance, and cultural appropriateness.
Risks
Despite the advantages listed above, using genAI tools does not come without a range of ethical, legal, practical, environmental, and philosophical risks. Some of these can be mitigated against through cautious critical use, while others are unavoidable from the perspective of the end-user and cannot be lessened without large-scale sector regulation.
For the purposes of this section, we aim to make users aware of the most pressing concerns, and—where possible—how these can be accounted for:
Hallucination or confabulation
GenAI systems can produce content that appears plausible and authoritative but is factually incorrect, incomplete, outdated, or entirely fabricated. Users may not always recognise these errors, particularly when working outside their areas of expertise. As a result, AI-generated outputs should be verified against reliable sources before being used or shared.
Over-reliance
Frequent reliance on AI-generated responses may discourage independent thinking, reflection, and problem-solving. GenAI tools often prioritise appealing to and meeting users’ needs above strict accuracy. In other words, responses may be designed in such a way as to appease users’ preferences in order to increase engagement, potentially promoting dependence and limiting criticality— ‘Although affirmation may feel supportive, sycophancy can undermine users’ capacity for self-correction and responsible decision making’ (Cheng et al. 2026, p.1348; see also Shaw and Nave 2026). In order to retain intellectual independence, we caution against defaulting to the use of genAI tools for tasks that could be better accomplished autonomously. When genAI tools are used, they should be seen as supplementary aids rather than as sources of original ideas to which any task can be outsourced.
Bias
GenAI tools are trained using large and often unfiltered corpus data sets. As reflections of human history, culture, and society, these sources inevitably contain a range of biases (e.g., racial bias, gender bias, etc.), which can shape how the training models in question ‘learn’, in turn inflecting their outputs. In short, they may reproduce stereotypes, underrepresent certain communities, or generate outputs that disadvantage particular groups. Careful critical evaluation on the part of the human user is therefore essential to ensure that such biases are not reproduced and amplified in any work produced using genAI tools, and that AI-supported work aligns with institutional policies of equity, inclusion, and respect.
Data privacy
Information entered into genAI systems may be stored, processed, or used in ways that are not always transparent, posing risks to privacy, safeguarding, and data protection. Depending on the specific confidentiality capabilities of the tools being used, inputting given data may be equivalent to sharing them with a third party and may therefore contravene GDPR regulations. It is particularly important to be mindful of the potential ethical and legal risks involved when working with personal, confidential, or sensitive data, for example relating to school pupils.
Plagiarism and authorship
Firstly, for the sake of academic integrity and honesty, any content produced with the use of genAI tools should be cited accordingly (see the end of the present document, for example). Similarly, if uploading or otherwise using copyrighted images via genAI tools, permission must first be sought to do so.
More broadly, a key issue in the mechanics of genAI tools is that the LLMs powering them are trained on vast datasets, much of which is not obtained through transparent processes. Questions therefore remain regarding ownership, copyright, attribution, and the use of copyrighted materials in the development and operation of genAI systems. Users may unintentionally create or share content that raises legal, ethical, or professional concerns relating to authorship and intellectual property. This is particularly relevant for educational practitioners using tools to generate songs or other artwork, for example, as these typically rely on the exploitation of human artists who have not given consent for their own work to be used as training data.
Environmental impact
Training and operating large genAI models entails substantial computational resources, vast energy and water consumption, and disruption caused to local communities—disproportionately in low-income areas—by the construction and operation of large data centres. While the scale of such impact is difficult to estimate precisely since relevant information is not always transparent, users should be aware that the environmental costs of genAI technologies are significant. These costs are often hidden from users, making it difficult to assess their environmental footprint. We therefore advise against using genAI tools unnecessarily, and recommend that users consult research on the environmental impact of such resources (see, for example, Hosseini et al. 2025; Moravec et al. 2025; Zewe 2025).
Bibliography and useful resources
Artificial intelligence (AI) | Definition, Examples, Types, Applications, Companies, & Facts | Britannica. 2026. Available at: https://www.britannica.com/technology/artificial-intelligence.
Chidlow, D. 2026. AI terminology is poorly defined and oft misused. Available at: https://vale.rocks/posts/ai-terminology.
Generative AI & education guidance hub. [no date]. Available at: https://www.imperial.ac.uk/about/leadership-and-strategy/provost/vice-provost-education/education/resources/ai-education-hub/.
Hendrycks, D. et al. 2025. A definition of AGI. [Working paper].
King’s guidance on generative AI for teaching, assessment and feedback. [no date]. Available at: https://www.kcl.ac.uk/about/strategy/learning-and-teaching/ai-guidance.
Zewe, A. 2025. Explained: Generative AI’s environmental impact. Available at: https://news.mit.edu/2025/explained-generative-ai-environmental-impact-0117.
Transparency statement
This document was produced with limited assistance from Microsoft Copilot, a genAI tool licensed under Cardiff University’s tenancy. Its use formed part of the project’s commitment to engaging critically with generative AI technologies. AI-generated suggestions were evaluated alongside other sources and professional judgement, and responsibility for the final content rests with the human author.
For queries or further information, please contact mflmentoring@cardiff.ac.uk.
[1] It is worth noting the widely acknowledged inconsistencies in nomenclature applied to all AI technologies in recent years, encompassing terms such as genAI, LLM, AGI vs ASI, agentic AI, etc.—’What was a deeply technical community with generally agreed-upon terminology was largely burst open with the release of ChatGPT‘ (Chidlow 2026; see also Hendrycks et al. 2025).