In our earlier article, AI Translation Is Fast—But Is It Fit for Purpose?, we examined how organisations can assess automated translation against the requirements of a particular assignment.
A longer-term question follows: if responsible AI use depends on specialist human oversight, how do we ensure that future generations develop the expertise to provide it?
This question reaches well beyond language services. Across medicine, science, engineering and law, increasingly capable systems are changing how professional work is performed. They also invite scrutiny of how people acquire, maintain and exercise independent judgement.
Expertise develops through sustained study, practical experience, investigation, correction and feedback. These processes enable specialists to recognise an unfamiliar problem, challenge a plausible answer and identify when further evidence is needed.
For multilingual communication, the implications are particularly important. Clients often commission work in languages they cannot independently assess. They depend on professionals whose knowledge extends beyond fluent expression to specialist terminology, cultural context, ethical responsibility and the appropriate handling of information.
The challenge is to harness AI’s capabilities while sustaining the education and professional experience that make informed oversight possible.
What science and medicine are beginning to reveal
Research across several fields suggests that immediate productivity and long-term professional capability deserve separate consideration.
A Nature study examining 41.3 million scientific papers found professional advantages associated with scientists’ use of AI, alongside a narrowing of the collective scientific territory explored. The authors describe an expansion in individual scientific impact accompanied by a contraction in collective reach. [1]
The study examines AI tools across different periods, rather than generative AI alone. It nevertheless raises an important question: how might the technologies that help researchers produce more also influence the questions they investigate?
Medicine offers another example. A 2025 multicentre observational study found that adenoma detection during colonoscopies performed without AI fell from 28.4% before the introduction of AI assistance to 22.4% afterwards. The findings suggest a possible deskilling effect; they do not establish that AI caused the decline. [2]
Research presented at CHI 2025, involving 319 knowledge workers, found that greater confidence in generative AI was associated with less self-reported critical thinking. It also identified a shift in critical-thinking activity towards verification, integration and oversight. [3]
These studies concern different technologies and settings. They do not demonstrate that AI inevitably diminishes human capability, nor do they establish equivalent effects in language services.
They do, however, support examining how professional knowledge is developed and exercised as automation expands.
Human oversight requires more than a person in the process
Human review is frequently proposed as a safeguard for AI-assisted work. Its effectiveness depends on the reviewer’s competence, access to evidence, independence and time to investigate.
Assessing whether an answer sounds plausible is different from examining whether its assumptions, terminology and implications are justified.
In language services, effective review requires comparison with the source, understanding of the intended audience and relevant subject knowledge. Sometimes it requires recognising that the source itself is ambiguous and seeking clarification before proceeding.
NAATI acknowledges that AI can support terminology preparation, drafting and translation workflows. It maintains that human practitioners should remain central to intercultural communication and that AI should not be used in high-risk assignments without expert supervision by a certified practitioner. [4]
That position carries a practical implication: organisations need access to people with sufficient expertise to supervise meaningfully.
Oversight becomes a weak safeguard when those responsible lack the knowledge or opportunity to challenge the output.
Translation expertise must continue to develop
A fluent translation can contain an error that only a knowledgeable reader would recognise.
For example, a contractual permission might be rendered as a requirement. A medical screening result might be expressed as a diagnosis. A familiar word might be assigned its everyday meaning when the document requires a specialist interpretation.
These are illustrative scenarios, but they show why linguistic fluency alone cannot establish accuracy.
Professional translators develop the ability to research terminology, compare authoritative sources, recognise inconsistencies and justify their decisions. Relevant subject expertise helps them identify where further investigation is necessary.
AI can assist this work. Training and professional practice must still provide opportunities to develop those underlying skills.
If developing practitioners mainly accept or lightly edit automated suggestions, will they gain enough experience constructing, testing and defending their own interpretations? How should training evolve so that AI assistance strengthens their judgement?
These are questions for educators, language-service providers and the organisations commissioning specialist communication.
Interpreting requires judgement within live interaction
Interpreting adds the demands of communication as it unfolds.
A speaker may hesitate, change direction, use a regional expression or describe something unfamiliar. The setting may involve distress, disagreement or significant consequences for the participants.
AUSIT’s guidance identifies challenges for machine interpreting involving dialects, unclear or colloquial speech, and cultural or setting-specific meaning. Suitability varies across languages and circumstances. [5]
Consider a hypothetical medical consultation in which a patient uses an unfamiliar expression to describe a symptom. A professional interpreter can seek clarification and make the uncertainty visible to those involved. A confident automated rendering could leave the ambiguity unnoticed.
Developing this judgement involves language proficiency, ethical training, preparation and supervised experience of complex interactions.
The question extends beyond whether a system can interpret a particular exchange today. It includes how future practitioners will acquire the experience needed to recognise when communication is breaking down.
Voice-over requires linguistic and performance expertise
Synthetic voices can offer practical benefits for some projects, including rapid updates and repeatable narration. Their suitability depends on the audience, content and intended effect.
Multilingual voice-over involves decisions about script meaning, pronunciation, emphasis, pacing and cultural appropriateness.
A public-health message may require careful delivery of a warning. An advertisement may depend on humour or emotional intention. A technical recording may contain terminology whose pronunciation cannot be reliably judged by the client commissioning it.
Human voice artists and linguistic directors develop these skills through performance, feedback and collaboration. Synthetic narration also requires appropriate review: a natural-sounding recording does not establish that it communicates correctly.
Professional development therefore needs to preserve both linguistic knowledge and production judgement, including the ability to assess generated speech.
Where an identifiable person’s voice is cloned, responsible commissioning should also establish explicit permission and agreed boundaries for use and reuse. Permission to record a voice should not be assumed to authorise creating a reusable synthetic version.
Subtitling requires attention across the whole production
Automated subtitling can combine speech recognition, translation and timing. Each stage can introduce or amplify errors.
A technical term may be misheard. A speaker may be incorrectly identified. Shortening a sentence to fit the screen may remove an important condition.
W3C guidance states that automatically generated captions need confirmation of full accuracy to meet user needs and accessibility requirements. [6] Translated subtitles introduce the additional task of preserving meaning across languages while keeping text readable and synchronised.
Expert review therefore needs to consider the finished video, including its spoken content, visual context and reading demands.
These skills develop through practical work. Training should enable emerging specialists to understand the reasons behind corrections and make informed decisions themselves.
Professional judgement includes protecting information
Expertise also involves recognising when a proposed workflow is inappropriate for the information involved.
Language-service materials may contain medical histories, legal advice, identity documents, commercial negotiations or unpublished research. Processing them through an external tool raises questions about retention, training use, third-party access and cross-border handling.
The Office of the Australian Information Commissioner recommends, as best practice, avoiding the entry of personal information—particularly sensitive information—into publicly available generative AI tools. Its guidance also addresses product assessment, appropriate use and disclosure, security and transparency. [7]
Data residency concerns where information is stored or processed. Data sovereignty also involves the legal jurisdictions that may apply. Australian hosting alone does not resolve questions about overseas access, provider arrangements or contractual control. Australian Signals Directorate guidance identifies access locations, legal requirements and security controls as relevant considerations for cloud services. [8]
Human-delivered services also require secure information handling. The professional capability worth cultivating includes knowing which questions to ask, when permission or specialist advice is needed, and when a tool should not be used.
Building the next generation of specialists
AI will continue changing professional work. Some tasks will become faster, roles will evolve and new forms of expertise will emerge.
Education and employment practices should support that evolution while preserving substantive knowledge and independent judgement.
For language services, this means combining responsible technology use with terminology research, source evaluation, supervised practice and meaningful feedback. It means giving developing practitioners opportunities to investigate difficult questions and learn from experienced colleagues.
For clients, it means recognising that competent review is professional work requiring appropriate expertise and time.
At Australian Multilingual, the commissioning questions remain practical: who needs to understand the message, what specialist knowledge does it require, what could follow from an error, and how must the information be protected?
Answering those questions depends on people who have developed the capability to assess them.
AI can expand what specialists accomplish. Sustaining that benefit requires continued investment in how specialists are formed.
Across science, medicine and multilingual communication, the capacity to question an answer remains essential to deciding when it deserves our trust.
Contact Australian Multilingual to discuss translation, interpreting, voice-over and subtitling requirements where specialist knowledge, context and confidentiality matter.
References
[1] Hao, Q., Xu, F., Li, Y. and Evans, J. (2026). Artificial intelligence tools expand scientists’ impact but contract science’s focus. Nature.
[2] Budzyń, K. et al. (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. The Lancet Gastroenterology & Hepatology.
[3] Lee, H.-P. et al. (2025). The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. CHI 2025.
[4] NAATI. Position Statement on Use of AI for Translation and Interpreting Purposes.
[5] AUSIT (2025). Position Statement on the Use of Machine Translation, Machine Interpreting and Artificial Intelligence in Translation and Interpreting.
[6] W3C Web Accessibility Initiative. Captions/Subtitles.
[7] Office of the Australian Information Commissioner. Guidance on Privacy and the Use of Commercially Available AI Products.
[8] Australian Signals Directorate. Cloud Computing Security for Executives.
The research examples and professional guidance were checked during preparation. t.co