Daejeon, July 27–31
This short paper addresses the conference theme of “Engagement” and the subtheme “Translating | Translinguality in the Era of AI” by examining prompting as a site where human linguistic competence, cognitive self-management, and AI systems intersect. It investigates how selected strategies from Neuro-Linguistic Programming (NLP); specifically, reframing, anchoring, “chunking” (up and down), and meta-model questioning can be critically repurposed to structure human interaction with large language models (LLMs) used for translation and multilingual text generation.
While NLP has been widely critiqued and remains controversial as a therapeutic modality (Tosey and Mathison 2009), its emphasis on language patterns, framing, and goal-oriented communication offers an intriguing heuristic framework for studying engagement in AI-supported translation. The project asks three core research questions: How can NLP-inspired techniques help translators formulate clearer, more controllable prompts for AI-driven translation systems? How might these techniques support systematic “debugging” of unsatisfactory or biased AI outputs? Can they contribute to maintaining professional agency and ethical awareness in increasingly automated translation workflows?
These questions respond to current concerns in translation studies and machine translation research about overreliance on automation, opacity of model behaviour, and the erosion of human responsibility in post-editing settings (Chesterman 2016; Castilho et al. 2017; Kenny 2022). They also resonate with debates in AI ethics on transparency, contestability, and responsible governance of generative systems, particularly concerning linguistic diversity and the dominance of English in AI training data (Floridi and Chiriatti 2020; Bender et al. 2021; Tanwar et al. 2025).
Empirically, the study draws on a small corpus of translator-AI sessions involving twelve participants (Polish L1: advanced students and professional translators) who use GPT-4 and Claude 3.5 Sonnet as support tools in EN→L1 translation tasks. The choice of these two LLMs allows for comparative analysis of how different model architectures and training approaches affect translation quality, prompt responsiveness, and bias patterns. Participants are recruited from MA-level translation programmes at European universities and have 1–3 years of professional translation experience.
Participants receive a 90-minute, carefully de-mystified introduction to selected NLP strategies, explicitly separated from their therapeutic origins and presented instead as optional linguistic heuristics. To clarify the distinctions and relationships between these strategies:
Overlaps occur when strategies are combined iteratively: for instance, a translator might chunk down to a specific sentence, reframe it for a particular audience, anchor it to genre conventions, and then use meta-model questioning to challenge an inadequate AI response. Divergences lie in their primary functions: chunking and reframing are generative (shaping the prompt), while anchoring and meta-model questioning are evaluative (constraining and critiquing outputs).
Participants are then asked to apply these heuristics while completing translation tasks in a controlled university lab setting. The resulting data include complete prompt histories and AI responses (captured via API logs), screen recordings and keystroke logs (using Inputlog software) that capture on-the-fly revisions, and 30-minute semi-structured retrospective interviews in which participants comment on their strategies, frustrations, and ethical concerns. These multimodal data are analysed using a combination of qualitative coding (following Charmaz 2014) and lightweight corpus techniques.
First, prompts are categorised according to their predominant NLP-inspired strategy (e.g., reframing prompt, chunking prompt, meta-model challenge) and their function within the translation workflow (initial drafting, clarification, style adjustment, error repair). Second, corresponding AI outputs are examined for changes in adequacy, fluency, register, and cultural nuance, with particular attention to how prompts steer the handling of idioms, politeness, and references to local culture. Outputs from GPT-4 and Claude are compared to identify model-specific patterns in responsiveness to structured prompts. Third, interview data are coded for perceptions of control, trust, responsibility, and cognitive load, with attention to participants’ reflections on language hierarchies and the models’ differential competence in English versus their L1s.
By mapping NLP concepts onto concrete prompting patterns, the paper develops a provisional “NLP-for-prompts” toolkit for translator education and multilingual DH practice. This toolkit is not presented as a prescriptive method, but as a vocabulary for making explicit the otherwise tacit strategies that experienced translators use when “talking to” AI systems. Early analysis suggests that NLP-inspired prompts encourage participants to externalise evaluation criteria (e.g., “Act as a critical editor for a Polish academic audience interested in digital humanities…”), challenge the system’s first answer rather than accept it (e.g., “What is another way to render this idiom that avoids a literal calque?”), and articulate ethical and contextual constraints (e.g., “Do not erase gender-neutral language; preserve non-binary forms where possible.”).
Such practices align with current debates on responsible AI and the need for transparent, contestable AI outputs in the humanities (Floridi and Chiriatti 2020; Hagendorff 2020). They also resonate with Chesterman’s notion of translator ethics as a set of emergent “memes of translation” that circulate and stabilise professional behaviour (Chesterman 2016). Crucially, the study reveals how linguistic hierarchies embedded in LLM training data shape translation quality: participants consistently report that the models produce more nuanced, contextually appropriate outputs when translating into English than when generating Polish, German, or Czech text. This asymmetry raises urgent questions about AI’s role in reinforcing or challenging global language inequalities (Bender et al. 2021). The NLP-inspired prompting strategies, particularly meta-model questioning and cultural reframing, offer translators tools to explicitly name and resist these biases, but they cannot fully compensate for structural imbalances in training data and model design.
At the same time, the paper critically reflects on the risks and limitations of importing NLP into DH and translation contexts. It discusses the problematic history of NLP as a quasi-scientific therapeutic movement and emphasises the importance of re-situating its techniques within empirically grounded research and clear ethical guidelines (Tosey and Mathison 2009; Sturt et al. 2012). The analysis considers whether “optimised” prompting might inadvertently increase cognitive and temporal pressure on translators, normalising high levels of automation and shifting even more responsibility onto individual workers for detecting and correcting errors produced by opaque models. Interview data reveal mixed responses: some participants report increased agency and control, while others express frustration at the additional metacognitive labour required to “train” the AI through iterative prompting. The paper argues that prompting strategies should be understood not as individual solutions, but as part of a broader institutional and infrastructural response to AI-mediated translation that includes transparent model documentation, multilingual training data governance, and fair labour practices (Moorkens 2020; Porsdam Mann et al. 2023).
The contribution of this short paper is threefold. Conceptually, it reframes prompting as a form of meta-translation and translingual negotiation, rather than a mere technical interface with AI. Methodologically, it demonstrates how small-scale, process-based studies of human-AI interaction can illuminate the micro-practices through which translators negotiate agency, quality, and ethics in AI-mediated environments. Practically, it offers a critically informed, adaptable toolkit that can be used in translator education, DH classrooms, and multilingual project teams to foster more reflective, responsible engagement with AI translation technologies.
In doing so, the paper speaks directly to the conference’s focus on “Engagement” by foregrounding the linguistic and cognitive work required to make AI systems meaningful, accountable partners in translingual practice, and by exploring how European and global translation communities can actively shape, rather than merely adapt to, the evolving landscape of AI-driven multilingual infrastructures. By explicitly addressing linguistic diversity, model-specific performance differences, and the ethical implications of language hierarchies in AI systems, the study contributes to ongoing efforts within DH to build more equitable, transparent, and culturally responsive digital research infrastructures.
digital humanities, translation studies, generative AI, Neuro-Linguistic Programming (NLP), prompting, machine translation, translinguality, ethical AI, linguistic diversity