‘Real time, real change’: the role of artificial intelligence in providing real‐time language translation services in regional medicine

Abstract

Artificial intelligence (AI) is rapidly evolving worldwide, enabling greater flexibility and applicability to the field of language translation within healthcare. Australia is currently one of the most culturally and linguistically diverse countries in the world, creating a growing pressure on translators to ensure there is equitable and culturally safe access to healthcare services. Emerging research supports the use of AI as an adjunct to the current translator framework in being able to support low‐risk communication tasks, such as appointment coordination or simple instructions, but it should be avoided in high‐stakes contexts including informed consent or complex management discussions. Importantly, human oversight is needed to ensure clinical accuracy and patient safety during all translations.

Over the past decade, there has been rapid growth in the capabilities of artificial intelligence (AI), expanding their applications within the healthcare industry. This growth has seen large language models (LLMs), as a form of AI, expand in their processing times and accuracy of answers, enabling users to utilise real‐time language translation. This form of language translation is currently being used in platforms such as Google Translate with the recent adoption of Gemini models, allowing users to translate between languages anywhere in the world with internet access. ¹ This advancement represents a unique opportunity for regional medicine to offer LLM‐generated, real‐time language translations. These translations have the potential to improve cultural healthcare equity by directly overcoming language barriers. After all, ‘real‐time’ language translations represent a ‘real change’ towards improving broader, regional healthcare outcomes.

Australia is currently one of the most culturally and linguistically diverse countries worldwide, with more than 23% of the population speaking a language other than English at home and 3.4% of the population not speaking English well or not at all, representing a major barrier to accessing the Australian healthcare system and a potential ongoing issue relating to cultural safety. ² This barrier is more pronounced in regional healthcare locations due to the difficulty in accessing translated healthcare information and language interpreters. ³ Further, research published in the Internal Medicine Journal (IMJ) by Ekladious et al. highlighted lower access to language services among immigrant patients, raising significant concerns of cultural safety during consultations. ⁴ Medical doctors are also impacted by these barriers through increased costs in organising interpreters, as they often require longer consultations. ⁵ In response to these concerns, the Australian Government funds the National Translating and Interpreting Service (NTIS), a free, 24/7 service available to medical practitioners. However, in a review of the Primary Health Network Commissioned Mental Health Services, this system has been found to be unable to meet more than 20% of requests for a translator, mainly due to lack of translator availability. ⁶

It is important to distinguish between translation and interpreting in this context. Translation refers to the conversion of words between languages, whereas interpreting involves the conveyance of meaning, tone, cultural nuance and non‐verbal communication within a clinical interaction. ⁷ , ⁸ While LLMs may approximate real‐time translation, they do not replicate the relational and contextual functions of trained interpreters. As such, even high‐performing AI systems should be considered adjuncts rather than substitutes for professional interpreting services. This distinction is particularly important in high‐stake clinical interactions, where misunderstanding of nuance may directly impact patient safety.

LLM translators represent a unique opportunity to ‘bridge’ the current unmet demand and provide real‐time translation for patients. LLMs have been shown to produce equivalent quality translations to that of human translators for patient information materials in Spanish; however, they are less accurate when translating other languages including Arabic, Chinese and Vietnamese. ⁹ , ¹⁰ This introduces a potential equity paradox; populations with the least access to human interpreters, particularly speakers of less common languages, may also be those for whom LLM translation is least accurate, reflecting limited representation in training data. This mirrors the same populations for whom NTIS access is most limited.

Some further limitations of LLM translators include the potential for inaccurate translations of medical terminology and medication names. ¹⁰ LLMs require high‐quality microphone and audio devices to collect audio, which may not be readily accessible in regional settings. ¹¹ LLMs also have the potential to misinterpret reported clinical findings from patients. These limitations are likely due to the lack of medical data produced in the native language being spoken, increasing the risk of LLM hallucination. ¹⁰ Hallucination refers to inaccurate or untrue results that are generated by an AI model due to a lack of data or incorrect assumptions from when the model was constructed. ¹² Additionally, the medico‐legal implications of relying on AI‐generated translations remain unclear, particularly where errors may contribute to miscommunication or adverse clinical outcomes. Consideration must also be given to data privacy and confidentiality when using cloud‐based translation tools in clinical settings. Importantly, LLM performance is not fixed; task‐specific optimisation strategies, including domain‐specific fine‐tuning, retrieval‐augmented generation and hybrid human–AI workflows, have demonstrated potential to improve accuracy in specialised medical contexts.

Despite the widespread availability of LLM translators, there is limited formal uptake in regional medicine due to a lack of data supporting their use, along with systemic infrastructure constraints. These constraints have been identified by Hains et al. and include lack of internet connectivity, network hardware and training and development. ¹³ Further analysis into these specific constraints was presented by Leslie et al., with solutions including Low Earth Orbit network providers such as Starlink, functioning electronic medical records and introducing education and training into medical school curricula. Importantly, the article emphasised that upgrades in metropolitan centres will have flow‐on benefits to regional centres, allowing decision‐makers to take time to plan deliberate and carefully architected infrastructure rollouts. ¹⁴ However, investment in digital infrastructure should be considered alongside alternative strategies, including expansion of the trained interpreter workforce. While expanding interpreter services requires workforce training and retention, LLM deployment involves upfront infrastructure, governance and ongoing validation costs, the relative cost‐effectiveness of these approaches in regional settings remains uncertain, particularly in geographically dispersed populations.

Furthermore, alongside the plan for infrastructure improvements there needs to be a focus on clinician education. This is supported by Shinners et al., who identified the need to develop comprehensive education to promote understanding and hence adoption of AI technology. ¹⁵ To facilitate comprehensive education on LLM translators, it is important to gain support from government organisations such as the National Accreditation Authority for Translators and Interpreters (NAATI). This organisation is the main governing body providing nationally recognised accreditation to translators and interpreters within Australia and could play an important role in guiding the adoption of LLM translators. As part of this research, it would be useful to evaluate LLMs against current NAATI examination standards, allowing clinicians to understand where current LLMs perform at levels similar to NAATI‐certified translators. The extent to which clinicians will engage with additional training remains uncertain, particularly in already resource‐constrained regional settings, and may represent a barrier to implementation.

Therefore, LLM translators represent a potential tool to improve healthcare access and equity for patients who are non‐fluent in English in regional locations across Australia, strengthening cultural safety during consultations. In practice, clinicians are already informally using unvalidated, LLM‐based translation tools. At present, their use should be limited to low‐risk communication tasks, such as appointment coordination or simple instructions, and should be avoided in high‐stake contexts including informed consent, diagnostic disclosure or complex management discussions. We propose that LLM translators be introduced as adjuncts to existing services within a structured evaluation framework, including benchmarking against current standards for medical translation such as those provided by NAATI, alongside ongoing real‐world evaluation, safety monitoring and clear clinical governance to guide adoption by national bodies. Clinicians should therefore approach LLM translation as a supplementary tool, using clinical judgement to determine appropriate contexts while prioritising patient safety and access to professional interpreting services where available. We therefore advocate for cautious, adjunctive use of LLM translation in low‐risk clinical settings at present, alongside parallel investment in validation, governance frameworks and interpreter services before any broader clinical adoption.

Data availability statement

Data sharing not applicable to this article as no datasets were generated or analysed during the current study.