Qualitative data, cross-language research and AI translation: Three icebergs

In qualitative research, language is data. But language is not neat, straightforward data. What is said in an interview is not a transparent, objective record of reality: It is a social construction, shaped by the interviewer's aims and the participant's motivations and perceptions. It is a co-construction, arising from the interaction between researcher and participant. And, as anyone who has read a transcript knows, it is nonlinear, often illogical, inevitably messy. Participants circle back, detour, cut off mid-thought, contradict themselves and mix their metaphors; they send both explicit and implicit messages; they do not always say what they mean or mean exactly what they say. All of this makes interview data inherently complex and subjective, and it necessarily shapes how we make meaning from it.¹

The translation of qualitative interview data further complicates an already complicated situation.

In their focused literature review on cross-language translation in qualitative HPE research, Schumann et al.² found that few papers ‘explicitly discuss the dilemmas and challenges of translation and offer strategies and recommendations’. Beyond this ‘tip of the iceberg’, they report that most papers leave the issue of translation ‘below the waterline’: They avoid translation altogether, translate without explicit acknowledgement or translate with only cursory acknowledgement of the decisions and strategies involved. Drawing on an existing framework,³ the paper offers suggestions for how researchers should address the ‘why’, ‘when’, ‘what’, ‘who’ and ‘how’ of translation to incorporate it into qualitative research in a more methodologically robust way.

Schumann et al. encourage researchers to conceptualize translation as not merely a technical procedure but also one that presents significant methodological and ethical issues. Arguing that translation is ‘an essential means of ensuring the inclusion of diverse perspectives that might otherwise get lost’ (p. 3), they point to challenges beyond mere language accuracy, including cultural context, interpretive complexity and the unavoidable loss or change of meaning. This commentary aims to extend the paper's discussion by reflecting further on two related questions that researchers face in cross-language qualitative work: How much should we translate, and how might artificial intelligence (AI) help or hinder our translation efforts?

The question of how much translation is necessary remains unanswered in Schumann et al.'s review. Yet this question is essential to finding ‘the right balance between methodological rigor and practical feasibility’. Researchers must make individual, context-related judgements about the degree of translation needed to serve their purpose while remaining feasible within their resources. Should entire transcripts be translated? Only the passages that a multilingual team wishes to discuss analytically? Only the excerpts selected for inclusion in the published manuscript? There is no single best answer—only a thoughtful balancing of fidelity and practicality. A precise, phrase-by-phrase translation may preserve nuance, but it could be time-consuming, expensive and impractical, especially in multilingual studies. More selective or interpretive translations, on the other hand, run the danger of removing culturally entrenched meanings that are important to the study. To grapple with the question of how much is enough, researchers might consider these questions:

Recognizing that there is no unilaterally ‘perfect’ translation threshold, how much translation is required should be gauged with reference to the study goals and the demands of both ethical integrity and research credibility.

The question of how to best translate is also critical. It is tempting to think that AI offers an efficient solution. We wonder if it might even lead researchers to ‘over-translate’; after all, if AI translation is fast and free, why not translate everything? This takes us back to the how much question: Is more translation always better? We would argue that it is not, for three main reasons. First, there are the concerns Schumann et al.'s review related to the accuracy and cultural sensitivity of AI translations: They caution that AI systems are technically proficient but less reliable when source texts contain ambiguity or contextual nuance.⁴ Given this, they advise strategic use of both AI and human translation to capitalize on the strengths of both. Second, if we translate everything because AI makes it fast and easy to do so, we will find ourselves in the unwieldy situation of a multilingual research team conducting analysis and interpretation discussions based on reading full transcripts translated from or to multiple languages. Even unilingual research teams find themselves awash in qualitative data and struggling to find the forest for the trees. This challenge (a very real one for the trustworthiness of the analytical process and the robustness of the resulting insights) would be multiplied if teams were working with complete transcript sets in multiple languages.

Third, and perhaps most importantly, the language of transcripts may be exactly the sort of thing that AI is not very good at translating. While the published literature on using AI for translation purposes is limited, a 2024 review of the current state of AI-driven language translation systems suggests that they struggle particularly with idiomatic expressions, emotional subtleties and implicit messages and that they fail to achieve sufficient depth of meaning in translations when dialects, jargon and informal speech are involved.⁵ These features sound suspiciously to us like key characteristics of interview transcripts! Therefore, qualitative researchers who employ AI for translation should take care, knowing that this is not the translation environment that AI was created for and, consequently, a robust and reliable AI translation of their transcripts will be neither fast nor easy. It will undoubtedly require additional steps, such as AI customization on domain-specific glossaries or transcript standardization to remove filler words and correct grammatical errors (‘standardization’ which may violate research principles depending on paradigm and methodology⁶). And it will certainly, as Schumann et al assert, require iteration and hybrid AI–human translation strategies.

So we are dealing with not one but three icebergs when we consider the future of cross-language translation of qualitative interview data. First, what is on the page of a transcript even before translation is only a partial representation. We cannot take transcribed language as transparent, straightforward or complete in its meaning. Second, as Schumann et al. argue, the current state of qualitative cross-language research in HPE is one in which the issue of translation is rarely and incompletely addressed. We must be more reflective and explicit about the role of translation in our research. And third, while AI is a promising translation tool, it struggles with precisely the sort of language features that abound in qualitative transcripts. We must recognize that not only does AI currently have weaknesses related to ambiguity and context, but it may be fundamentally ill-suited to our purposes as qualitative researchers confronted with translating interview data.

Lorelei Lingard: Conceptualization; writing—original draft; writing—review and editing. Jennifer Klasen: Conceptualization; writing—original draft; writing—review and editing.