Exploring Strategies to Detect and Mitigate Bias in AI in Education: Students’ Perceptions and Didactic Approaches

Abstract

The increasing integration of Generative AI (GenAI) into higher education, particularly in the domain of language teaching, presents both opportunities and challenges. While AI-powered tools such as ChatGPT-5 can support language learning by generating personalised content which enables real-time interaction and feedback, they also risk perpetuating biases embedded in training data. These biases can appear in linguistic, cultural or socio-political forms, reinforcing stereotypes and influencing language norms. Therefore, equipping students and educators with strategies to critically assess AI outputs is essential for ethical and responsible AI use in language education. While recent research highlights the risks of algorithmic bias, less attention has been given to the perceptions and attitudes of pre-service teachers, whose future practice will shape classroom uses of these technologies. This exploratory pilot study adopts a survey-based approach to examine pre-service teachers’ baseline awareness of bias in artificial intelligence, with particular attention to linguistic and cultural dimensions Data were collected through an online questionnaire administered to 65 undergraduate students enrolled in Primary Education degree programmes. The study documents baseline perceptions prior to any instructional intervention and provides preliminary empirical evidence to inform the future design of pedagogical strategies aimed at developing critical AI literacy in teacher education.

1. Introduction

The integration of Generative Artificial Intelligence (GenAI) into higher education and language teaching offers significant advantages: it can generate personalised content, create dialogue exercises and, above all, provide real-time feedback to each student [1,2,3,4,5]. However, these systems learn from the vast amount of data available online, much of which is imbued with historical or cultural prejudice [6,7,8,9]. For example, in its research on the responsible use of AI, UNESCO identified representational bias in large language models (LLMs), which tend to depict female characters in association with words such as family, children, or husband, whereas male names are more frequently linked to concepts such as executive, business, or management [10].

This is a crucial issue to consider, as the fact that LLMs tend to associate women with domestic roles and men with executive positions, even in systems that have been specifically programmed to avoid bias [11], reveals the persistence and perpetuation of subjective and harmful stereotypes that must be prevented, particularly in educational settings [12,13].

Regarding linguistic bias, it is noteworthy that GenAI models are generally trained on predominantly English-language corpora, as these are the most abundant. This may impose a linguistic bias that marginalises not only other languages [14,15] but also varieties that deviate from the standard norm [16].

Similarly, when training data include sexist, racist or culturally biased content, GenAI may reproduce or amplify such biases in its outputs. This means that students could receive distorted or partial information during their language learning activities without even being aware of it [17,18].

AI-driven recruitment and hiring tools, for their part, can also present algorithmic biases when they screen and rank candidates, given that they focus on the candidates’ names for these tasks and it serves as a proxy for race or gender; as a result of these proceedings, men’s names were far more likely to be selected than women’s and, furthermore, White-associated names were preferred over Black-associated names [18,19].

Therefore, GenAI must be approached with caution, given that it can be a double-edged tool whose considerable educational potential might be compromised by the risk of generating unwanted bias. That is the main reason why AI literacy must be catered and fostered among students, but also and not less importantly, among educators, because understanding how the latter comprehend and conceptualise the neutrality of AI, or are able to recognise potential forms of bias and to assess the reliability of AI-generated content is a necessary step towards the development of effective AI literacy initiatives within the area of teacher education [20,21,22].

In this study, it is important to distinguish between several related but distinct concepts. Awareness refers to the basic recognition that AI systems may contain biases. Concern denotes the perception that such biases may have ethical or educational consequences. Critical AI literacy involves the ability to analyse and interpret AI-generated outputs while recognising the socio-technical mechanisms that may produce bias. Finally, pedagogical competence refers to the capacity of future teachers to translate this understanding into instructional practices that help students critically evaluate AI-generated information.

While digital and media literacy have traditionally focused on analysing media content and understanding how information is produced and disseminated, critical AI literacy expands this perspective to include the socio-technical mechanisms underlying AI systems, including training data, algorithmic processes and potential sources of bias. In the context of teacher education, this expanded perspective is particularly relevant, as future educators must be able to help students use GenAI tools effectively and guide them in interpreting and contextualising AI-generated information.

In the field of language education, a growing body of recent research has highlighted the potential of GenAI as a support tool for academic writing, textual revision, personalised learning [23,24] and the development of communicative competence in the teaching of different languages, such as French or Spanish [25]. At the same time, however, significant risks have been identified, particularly those related to cognitive superficiality, technological dependence, discursive homogenisation and the blurring of authorship processes. We are therefore facing a period of considerable challenges with regard to the integration of GenAI into classroom practice [26].

Scholars such as Cassany [27] have stressed the need to pedagogically legitimise the use of these tools through explicit didactic protocols that promote metacognitive reflection, self-regulated learning, and critical literacy [28]. The potential applications of GenAI tools are vast, and specific uses can be identified across a wide range of educational contexts [29].

In higher education, the adoption of these technologies has been immediate, both among students and teaching staff. New functionalities related to text generation, translation, summarisation, and revision render many traditional activities associated with the teaching of literature increasingly obsolete, and their integration and assessment pose a significant challenge for educators [30].

However, while recent research has extensively documented the existence of algorithmic, linguistic, and cultural bias in AI systems, comparatively fewer studies have examined how pre-service teachers perceive these challenges [31,32,33,34].

This study aims to narrow that gap by adopting an exploratory, survey-based approach to examine pre-service teacher’s awareness of all these pressing issues. By documenting baseline perceptions prior to any instructional intervention, the study seeks to inform the design of future pedagogical instruments for the classroom, aimed to support pre-service teachers to develop the necessary competencies required to teach their future students about the critical, ethical and responsible use of GenAI tools in language education. By documenting baseline perceptions prior to any instructional intervention, the study seeks to inform the design of future pedagogical instruments for the classroom, aimed to support pre-service teachers to develop the necessary competencies required to teach their future students a critical, ethical and responsible use of GenAI tools in language education.

Many educators nowadays are concerned that its use may encourage superficial engagement or uncritical copying rather than genuine learning [20,35,36]; this study adopts the perspective that informed guidance and explicit instruction are more productive than restrictive approaches.

In order to structure the exploratory analysis, the research addresses the following research questions:

RQ1. To what extent are pre-service teachers aware that generative AI systems may reproduce linguistic, cultural, or gender-related bias?

RQ2. What level of concern do pre-service teachers express regarding the potential educational consequences of bias in generative AI?

RQ3. To what extent do pre-service teachers critically evaluate the information generated by AI systems and verify its reliability?

These research questions guide the analysis of the survey data and provide a framework for interpreting the baseline perceptions documented in this exploratory pilot study.

2. Materials and Methods

The present study follows an exploratory, survey-based design aimed at examining the perceptions, attitudes and preconceived ideas of pre-service teachers regarding bias in AI. The questionnaire with all its items can be found in Appendix A. The research focuses on capturing students’ awareness of linguistic, cultural and social bias in AI systems, as well as their attitudes towards the educational implications of such bias, prior to the implementation of instructional activities that will be designed following the results of this research, to make sure we are addressing the actual concerns of pre-service teachers and students. The results will subsequently be used by the research team to inform the design of pedagogical templates and classroom activities for language learning in higher education.

The questionnaire was designed ad hoc for this study, following an exploratory and descriptive survey approach adopted in recent research on pre-service teachers’ perceptions of AI in education that employs bespoke questionnaires [37,38,39], and conceptually aligned with the OECD’s Empowering Learners for the Age of AI framework [40].

These studies examining teachers’ perceptions and knowledge of AI [38,39] have adopted survey-based methodologies using bespoke questionnaires and Likert-type instruments to capture attitudes, levels of awareness and patterns of use. Building on these methodological precedents, the present questionnaire was designed to address key analytical dimensions related to the educational use of generative AI, including perceptions of AI neutrality, awareness of linguistic and cultural bias, and attitudes towards the educational implications of such bias. Prior to its administration, the questionnaire was reviewed by members of the research team to ensure clarity and coherence of the items.

The questionnaire was structured into different thematic blocks aiming to address key analytical dimensions in the use of GenAI: perceptions of AI neutrality, awareness of linguistic bias related to the predominance of English in AI training data, recognition of cultural and gender-related stereotypes, levels of trust in AI-generated information, and attitudes towards the educational implications of bias in AI. This structure allows for a systematic exploration of students’ baseline perceptions prior to any instructional intervention.

Participants were undergraduate students enrolled in the final two years of the Primary Education degree programme and the English Studies degree programme at the University of Alicante during the 2025–2026 academic year. Students from the English Studies programme were included as many of them pursue careers in English language teaching. This population was chosen because they are future teachers who are likely to integrate digital technologies, namely GenAI, into their professional practice. The questionnaire was distributed to students in courses where the research team had teaching responsibilities, and participation was voluntary. These courses were selected because generative AI tools had already been incorporated into certain coursework activities, including assignments and practical tasks, ensuring that students had some degree of exposure to these technologies.

Data were collected through an online questionnaire administered via Google Forms during the first semester of the academic year 2025–2026. The participants were informed of research purposes, implementation and confidentiality. The study was carried out in accordance with the ethical standards required for research involving human participants, adhering to the fundamental principles of the Helsinki Declaration and approved by the Ethics Committee of the University of Alicante regarding good research practice (Code: UA-2025-01-14).

The instrument was designed to prioritise clarity and accessibility and consisted mainly of closed-ended questions (Yes/No and multiple-choice formats), complemented by a single open-ended question intended to gather more detailed personal reflections.

The predominance of closed questions made it possible to conduct a descriptive quantitative analysis, allowing for the identification of general trends in students’ responses, while the open-ended item provided qualitative insights into individual points of view [41,42]. The responses to the open-ended question were analysed using an inductive thematic approach. The research team reviewed the responses and identified recurring themes related to students’ perceptions of linguistic, cultural, and gender bias in AI systems. Through iterative comparison and discussion, responses were grouped into three analytical categories reflecting different levels of awareness and critical reflection.

This methodological approach allows to establish a baseline profile of students’ perceptions prior to any instructional intervention. Consequently, the findings are interpreted as indicative of prevailing attitudes and levels of sensitivity towards AI bias, rather than as measures of pedagogical or learning outcomes. This design is particularly appropriate at an initial stage of research, as it informs the planning of future classroom-based interventions and supports the need for targeted AI literacy training within teacher education programmes [31,43,44].

3. Results

The results presented in this section derive from an exploratory attitudinal survey aimed at examining pre-service teachers’ awareness, perceptions, and attitudes towards bias in the information provided by GenAI tools, with particular attention to linguistic and cultural dimensions. The baseline purpose of this study was to find and explore the main concerns and perceived risks of pre-service teachers regarding the use of AI within their field of expertise, as well as their awareness of the potential problems and challenges such technologies may pose for their future professional practice.

As this is ongoing research, at this stage the study focuses on documenting the students’ perceptions and awareness before the implementation of instructional activities related to bias detection and prompt analysis, rather than reporting empirical classroom interventions or task-based learning outcomes, which will constitute the next phase of the research.

A total of 65 participants (75% female, 25% male) took part in the survey and, overall, the responses reveal a heterogeneous level of awareness regarding the non-neutral nature of AI-generated outputs.

While a large proportion of participants (78.5%) are aware that AI-generated information is not always neutral, a considerable number still believe that AI produces neutral information and is, therefore, free from bias, as shown in

Table 1. Students’ perceptions of the neutrality of AI-generated responses and awareness of potential bias.

Question	Yes	No	I Am Not Sure
Do you consider that AI-generated responses are always neutral?	21.5%	78.5%	-
Before this survey, had you reflected on the possibility that artificial intelligence may reproduce biases or stereotypes?	61.5%	38.5%	-
Do you consider that artificial intelligence may reproduce biases present in our society in an implicit or non-explicit manner, making them harder to identify?	67.7%	1.5%	30.8%

.

Prior to participating in this survey, many students (over half of the target population) had not reflected about the possibility that AI-generated text might be biased, whether it be linguistically or culturally (Table 1). This finding suggests that pre-service teachers’ understanding of AI neutrality is neither uniform nor fully developed, and it highlights the relevance of studies aimed at identifying baseline levels of awareness, which are essential for informing the design of effective AI literacy initiatives in teacher education.

The answers to the following question reveal a seemingly more alarming situation regarding our student’s awareness: “Do you consider that artificial intelligence may reproduce biases present in our society in an implicit or non-explicit manner, making them harder to identify?” (Table 1). While most participants are aware that AI may reproduce social biases in implicit or non-explicit ways, nearly one third of respondents still do not even consider the possibility, as they show uncertainty or lack of consideration of this issue. Although only a minimal proportion reject the possibility altogether, the mere existence of this response points to the need for continued attention in teacher education, particularly with regard to educating our students in developing critical thinking skills and a responsible use of AI technologies.

Over two thirds (72%) of the participants indicated being aware that most AI systems are trained mainly on English-language data, as

Table 2. Students’ awareness of linguistic and social bias in AI-generated responses.

Question	Yes	No	I Am Not Sure
Are you aware that most AI systems are trained mainly on English-language data?	72.3%	27.7%	-
Do you think this predominance of English may affect other language or linguistic varieties?	70.8%	10.8%	18.8%
Do you think artificial intelligence can reproduce gender stereotypes in its responses?	59.9%	32.3%	10.8%
Have you ever observed AI-generated responses that provide very simplified or generic information about certain cultures, countries or social groups?	60%	40%	-

shows. When asked whether this predominance of English may affect other languages or linguistic varieties, a substantial amount answered they did not think so (18.5%) or were not sure (10.8%), which is consistent with the same pattern of response: approximately the same percentage who admitted not knowing that AI systems were mainly trained in English was the same that expressed uncertainty regarding the potential role of English-language predominance as a key source of linguistic discrimination or bias affecting other languages.

When these responses are considered alongside those related to linguistic bias, a similar pattern emerges. Although a majority of participants report being aware that AI systems are predominantly trained on English-language data, a noticeable proportion do not perceive this predominance as a potential source of linguistic discrimination affecting other languages or varieties. This suggests that awareness of the structural characteristics of AI systems does not necessarily translate into a critical understanding of their broader sociolinguistic implications.

Comparing the responses regarding linguistic or social bias with those related to gender bias, students seem to be more aware of the former by a short percentage. This finding is noteworthy given that a substantial proportion of the participants identified themselves as female (75%), a group that has historically been more directly affected by gender-based discrimination. The results may therefore suggest that familiarity with or exposure to discrimination does not necessarily translate into heightened awareness of how such biases may be reproduced through AI systems, particularly when they operate in implicit or indirect ways. This also highlights the need for teacher education programmes to explicitly address different forms of bias in AI systems, as awareness appears to vary across bias categories, even among groups with lived experience of discrimination.

More than half of the participants (60%) indicate that they have previously observed AI-generated responses that provide overly simplified or generic information about certain cultures or social groups that are regarded as non-dominant or marginalised (Table 2). However, the fact that a substantial minority (40%) state that they have not noticed such patterns indicates that cultural bias may often operate in subtle or normalised ways that remain invisible to a significant proportion of users. This lack of recognition reinforces the importance of explicitly addressing cultural bias in educational contexts, as systematic instruction and guided reflection constitute a necessary first step in enabling pre-service teachers to identify and critically interpret biased representations in AI-generated content.

A comparable pattern emerged regarding the awareness of gender bias. As shown in Table 2, 56.9% of respondents believe that AI systems can reproduce gender stereotypes in its responses, while a non-negligible proportion either reject this possibility or express uncertainty (32.3%).

On the other hand, a large percentage of participants (78.5%) are aware that they have to contrast and verify the information provided by AI systems although they initially trust that the information is accurate (

Table 3. Students’ levels of trust in AI-generated information and their verification practices.

Question	I Trust It but with Caution, and I Usually Cross-Check and Verify the Information	I Do Not Fully Trust It, But I Do Not Verify or Cross-Check the Information Every Time	Yes, I Generally Trust It
When you use AI, do you fully trust the information it provides?	78.5%	15.4%	6.1%

). This indicates a high degree of conditional trust in AI-generated information. A small amount, nevertheless, confess that they do not fully trust the information but do not cross-check or verify the information consistently (15.4%), and there is a very small number (6.1%) that report fully trusting AI-generated content without engaging in any form of verification.

It is noteworthy that an overwhelming majority (90.8%) is concerned about the consequences of AI bias in educational contexts (

Table 4. Students’ perceptions of the impact of AI bias in educational contexts and the need for critical AI literacy in the classroom.

Question	Yes	No	I Am Not Sure/I Had Not Thought About It Yet
Do you think bias in AI can have consequences in educational contexts?	90.8%	9.2%	-
Do you consider that the critical analysis of AI-generated responses should be part of students’ classroom learning?	87.7%	7.7%	4.6%
As a future teacher, are you concerned about the impact of bias in the educational use of AI?	84.6%	4.6%	10.8%

), and agree that the critical analysis of AI-generated responses should be part of students’ classroom learning (87.7%). However, this widespread concern and awareness contrasts with earlier responses from the same participants indicating only partial awareness of cultural or gender bias, and inconsistent practices regarding the verification and cross-checking of AI-generated data.

A comparison between these responses and earlier items in the questionnaire reveals an interesting internal tension. Although participants express considerable concern about the consequences of AI bias in educational contexts, their ability to recognise specific manifestations of such bias appears to be more limited. A noticeably smaller proportion report having observed cultural bias or gender stereotypes in AI-generated responses. This discrepancy suggests that many participants possess a general awareness of the risks associated with AI, yet lack a more detailed conceptual understanding of how different forms of bias may manifest in practice.

This difficulty in recognising specific examples of bias has also been identified in studies analysing user interaction with large language models. Cultural and linguistic biases frequently operate through subtle discursive patterns, such as the privileging of dominant languages or the reproduction of Western-centred knowledge structures, which may remain largely invisible without specific analytical training [9,16,33].

Students appear to be generally concerned about the consequences of bias in educational contexts and consistently agree that AI literacy should be included in the teaching curriculum. However, when asked more directly whether they as future teachers are personally concerned about the impact of bias in the educational use of AI, their responses reveal a more nuanced general picture (Table 4). While a substantial majority report being concerned (84.6%), a not negligible proportion of participants (10.8%) indicate that they had not previously considered this issue. It is also noteworthy that, although it represents but a very small portion of the sample, 4.6% of students responded that they are not concerned about the problem of AI bias permeating education and the information to which users are exposed.

This finding suggests that although students recognise bias as a relevant and potentially problematic phenomenon at a general level, this concern does not always translate into personal reflection on their future professional practice. These results manifest the presence of a still underdeveloped critical vision and understanding of the problem that AI-related bias can pose in learning environments, and therefore strengthen the need for specifically designed and structured AI literacy instruction in teacher education programmes.

Finally, the responses to the open-ended question reveal a wide spectrum of awareness and experience of linguistic and cultural bias in AI. The qualitative data can be grouped into three main patterns:

• limited or absent reflection;
• emerging intuitive awareness;
• and more elaborated critical understanding.

The first group of participants report having no formed opinion about the risks of trusting GenAI and admit they have not reflected about such a concept before responding to this survey. Their experience with GenAI tools narrows down to pragmatic and instrumental use of the tools, with little critical thinking or interrogation about the accuracy or ethicality of its outputs. Some examples of anonymised responses from this group include:

“To be honest, I had never seriously considered whether artificial intelligence might contain linguistic or cultural biases. I tend to use it in a rather practical way and without questioning it too much, so I do not have a clear opinion about it.”

“I have not used AI enough to have a clear answer to this question.”

A second, more numerous group shows an emerging awareness of the need to critically analyse the automatically generated information. These participants also acknowledge the fact that AI may reproduce and perpetuate stereotypes and hence be biased. However, their awareness remains generic and abstract, with bias often regarded as unavoidable, and they state that they do not have specific strategies for identifying or addressing the problem. Some anonymised excerpts illustrating this pattern include:

“Even if AI does not have intentions of its own, I think linguistic and cultural biases are inevitable because the systems are trained on human data that already contain inequalities.”

“I think these biases are probably not intentional.”

“In my view, it is essential to maintain a critical perspective and question the information provided by AI. It is very important to verify and cross-check the data it produces.”

Finally, a smaller but significant group of responses reflects a more developed critical understanding. These participants explicitly identify linguistic, cultural and gender-related biases, including the dominance of English and Western perspectives, the marginalisation of non-dominant languages or varieties, and the reproduction of gender stereotypes. Several responses acknowledge the lack of specific training regarding these issues which are increasing in importance as the use of GenAI tools spreads, and they report the need for explicit academic instruction that stresses the importance of cross-checking the information provided by GenAI and provides them with specific tools to address the issue.

“AI learns from data produced by particular societies with their own values, hierarchies and inequalities. As a result, linguistic and cultural biases often reflect which languages are more visible, which cultures are treated as the ‘norm’, and which remain marginalised.”

“I think that when you interact with AI in English it tends to provide more information and in a more fluent way, whereas this does not happen in the same way in other languages.”

4. Discussion

Taken together, the findings reveal a gap between general awareness of the risks associated with AI use and the ability to critically identify specific forms of bias in AI-generated outputs. This pattern is consistent with previous studies on pre-service teachers’ perceptions of generative AI, which report a similar discrepancy between general concern about the risks of AI technologies and the development of more structured critical competencies required to identify specific forms of bias in AI-generated content [31,32]. Similar tensions between general awareness and analytical understanding have also been documented in broader studies on teachers’ perceptions of AI in education [37,38]. Recent research on AI literacy further suggests that awareness of the risks associated with AI technologies does not automatically translate into the analytical competencies required to critically evaluate AI-generated outputs [45,46].

At the same time, the present study extends this line of research by providing empirical evidence on how pre-service teachers recognise different types of bias, particularly linguistic, cultural and gender-related bias, in AI-generated responses.

The findings support the argument that AI literacy cannot be assumed to develop spontaneously through exposure or use alone, but must be explicitly taught through structured pedagogical interventions. This interpretation is consistent with recent research in teacher education, which emphasises that critical AI literacy requires explicit pedagogical scaffolding rather than informal exposure to AI tools. Studies examining teacher education programmes indicate that pre-service teachers often lack conceptual frameworks for analysing the socio-technical mechanisms underlying AI systems, including training data, algorithmic processes and potential sources of bias [34,39,43,47].

By documenting both concern and uncertainty, the survey provides empirical justification for the design of AI literacy curricula that move beyond abstract discussions of risk and instead provide future teachers with activities specifically designed to teach students how to critically contextualise and analyse the content generated by GenAI. This perspective also aligns with recent discussions on the educational implications of generative AI, which emphasise that the pedagogical integration of large language models requires the development of new critical competencies among students and teachers, particularly with regard to the evaluation of AI-generated information and the identification of algorithmic bias [47,48].

For this reason, the development of AI literacy should be approached as a step-by-step and carefully paced process. A fundamental consideration when incorporating GenAI into classroom practice is acknowledging the speed at which technological developments and innovations emerge on a daily basis. Consequently, it is neither necessary nor realistic to be familiar with every new tool as it appears. Rather, what is required is a gradual and reflective integration, in line with the principles of “Slow Tech”, which allows educators to adapt these technologies to their teaching practice in a thoughtful and pedagogically sound manner [49].

Based on these exploratory findings, several preliminary directions for future instructional design can be identified. These directions should be interpreted as hypotheses or design considerations for future pedagogical research rather than as empirically validated strategies derived from the present study:

• Acknowledging Inherent Bias and Prompt Engineering Mitigation: It must be accepted that current Generative AI, due to its underlying architecture and training data, will inherently exhibit biases regarding gender, race, and other sensitive attributes. Consequently, its educational application must prioritise the minimisation of bias within the prompt design itself. This can be achieved through explicit strategies—such as incorporating a specific instruction within the prompt that mandates unbiased results—or implicit strategies, such as providing non-biased examples within the context window (few-shot prompting). In this regard, further research is required to determine which prompting strategies are most effective in mitigating output bias.
• Fostering Critical AI Literacy: It is essential to promote critical analysis in the use of Gen-AI, addressing not only the factual accuracy of content to avoid hallucinations but also the identification of various types of bias. As a primary line of work, we must identify the most effective pedagogical strategies to develop students’ critical scrutiny regarding these systemic biases.
• Awareness and Social Impact: There is an urgent need to sensitise faculty and, more importantly, students regarding the limitations of Gen-AI tools concerning bias and their subsequent social impact (e.g., the reinforcement of stereotypes). Just as students are increasingly aware of the need to verify data to avoid errors caused by model hallucinations, they must also recognise that implicit or explicit biases may emerge, which they must address and correct. Therefore, training and awareness campaigns are necessary within university curricula to establish that bias is an inherent issue of the tool that must always be managed, regardless of the specific use case.

This is a critical aspect of the objective to empower learners for the use of AI-tools [40]. These findings highlight the importance of integrating critical AI literacy into teacher education programmes in order to equip future educators with the skills required to responsibly interpret AI-generated information.

This study has several limitations that should be acknowledged. First, the sample size (N = 65) is relatively small and restricted to a specific institutional context, which limits the generalisability of the findings. Second, the study relies on self-reported perceptions rather than observed behaviour when interacting with AI systems. Third, as an exploratory pilot study, the analysis focuses primarily on identifying patterns of awareness and concern rather than establishing causal relationships. The findings should therefore be interpreted as indicative of baseline perceptions that may inform future pedagogical research rather than as representative of pre-service teachers more broadly.

Another limitation of the study is that the questionnaire did not include specific items regarding participants’ actual use of generative AI tools, the particular applications they employ, the frequency of such use, or their access to technological infrastructure. Although digital technologies and AI-based applications are increasingly integrated into coursework and academic activities within these programmes, future research will incorporate these variables in order to better contextualise students’ perceptions and experiences regarding bias in generative AI systems.

5. Conclusions

This exploratory pilot study examined pre-service teachers’ awareness of linguistic, cultural and gender-related bias in generative AI systems. The findings reveal a noticeable gap between general concern about the risks associated with AI technologies and the ability to recognise specific manifestations of bias in AI-generated outputs. While most participants acknowledge that AI-generated information may not always be neutral, their understanding of how bias operates in practice remains uneven.

These results highlight the importance of integrating critical AI literacy into teacher education programmes. Future research should expand the sample to include different institutional contexts and examine the effectiveness of specific pedagogical interventions designed to help students critically analyse AI-generated content and identify different forms of bias.