AI gets plenty wrong but, says Safira Amazan, the real issue is what the translation industry gets wrong about AI

The AI debate focuses on reducing translation errors through better models. A more important question, however, is why those errors reach the target audience at all. The answer often lies in the workflow, not the technology. After all, translation quality is defined by the processes that transform a first draft into reliable communication. There are several workflow failures that allow AI errors to reach the public, with the greatest impact on lesser-resourced languages.

To understand those failures, let’s consider what translation workflows are designed to achieve. Professional translation relies on multiple layers of review: translation > editing > proofreading > quality assurance (QA) > target audience. ­Each stage serves a distinct purpose. Translators establish meaning, editors refine it, proofreaders improve readability and QA verifies compliance with project requirements. Together, these stages ensure the final translation communicates effectively.

Increasingly, however, AI-assisted workflows are compressed into: AI draft > machine translation post-editing (MTPE) > QA > target audience. The technology is not inherently responsible for this simplification; workflow design is. And every workflow is only as strong as the linguistic infrastructure supporting it.

Weak linguistic infrastructure

The quality of AI-assisted translation is shaped as much by linguistic infrastructure as it is by the AI model itself. AI does not operate in isolation. For dominant languages, a wealth of digital resources and guidance from professional and governmental bodies support both machines and human translators. Although AI produces errors in these languages, reviewers – or the AI itself with effective prompts – can usually identify and resolve them by consulting numerous authoritative references.

Lesser-resourced languages rarely benefit from that same infrastructure. Haitian Creole (one of my working languages) illustrates the challenge. Although its digital presence continues to grow, authoritative linguistic frameworks remain limited. Translators often work with fewer validated glossaries, fewer standardised references, and no recognised authority to support decisions.

These limitations help explain AI errors. They do not explain why those problems survive professional review. A very pertinent question is whether translators, reviewers and clients have the resources needed to identify, justify and correct those errors before they reach the target audience.

Poor terminology management

Translation quality begins long before the first sentence is translated. Translation memories, terminology databases, style guides, client preferences, well-crafted prompts and subject-matter expertise form the linguistic infrastructure that supports both human translators and AI systems.

Nevertheless, language service providers (LSPs) tend to invest heavily in automation while neglecting these essential resources, relying instead on generic prompts, incomplete glossaries, or no terminology guidance at all. The outcome follows a familiar principle: garbage in, garbage out. AI does not compensate for weak linguistic resources; it magnifies them.

Limiting this effect across all languages requires targeted instructions, validated terminology and consistent cross-checking. Glossaries, dictionaries and other references can support that work only when they are properly maintained, integrated into the workflow and made accessible to translators and reviewers. Without broader linguistic guidance, every decision carries greater weight, making human expertise even more important. Ultimately, organisations that invest in AI while neglecting these resources create the conditions in which AI is most likely to fail. 

The productivity myth

Many current pricing and productivity models appear to assume that AI-generated translations require substantially less human effort than human-produced translations. If that were consistently true, extensive review and full accountability would no longer be necessary. Yet LSPs continue to expect translators, editors and reviewers to deliver the same quality standard while holding them fully responsible for errors. The responsibility has not changed, only the time and compensation allocated to achieve that quality.

This contradiction reveals the productivity myth. AI changes how translators work, but it does not reduce the expertise, cognitive effort or critical judgement required to produce a reliable translation. Even so, many productivity models expect linguists to validate thousands of words of AI-generated content per hour while maintaining professional quality standards at low cost. AI may accelerate drafting; it does not proportionally accelerate professional review, nor does it diminish the value of human expertise.

Replacing editing with QA

No professional publisher would eliminate editing simply because the first draft was easier to produce. Still, as AI becomes more widely adopted, many LSPs have shortened, merged or eliminated the editing stage. The result, as illustrated above, is a condensed sequence where editing and proofreading disappear as distinct steps.

This shift assumes that MTPE can absorb the function of editing. Yet MTPE and editing are fundamentally different tasks. In compressed workflows, MTPE aims to make machine-generated text understandable; with pre-allocated time and lower compensation, the translation cannot be fully accurate, natural or appropriate for its intended audience. In fact, editing is what transforms an acceptable translation into an effective one. Therefore, treating MTPE as a substitute for editing – and QA as a substitute for both – removes one of the profession’s most important safeguards. A translation may pass QA because its terminology, formatting and consistency meet specifications, while problems with clarity, tone, style or register remain unresolved.

Each stage of the workflow answers a different question. MTPE asks, ‘Is the translation acceptable?’ Editing asks, ‘Does it communicate effectively?’ Proofreading asks, ‘Does the translation read naturally?’ QA asks, ‘Does it meet the required specifications?’ One stage cannot replace another. Reducing or eliminating editing may improve productivity metrics, but it also increases the risk that formal equivalence is mistaken for effective communication. Multiple reviewers create successive opportunities to identify and correct errors, yet many of these safeguards are being reduced. AI does not make those decisions. It does not remove editors, shorten deadlines, eliminate resources or reduce budgets. LSPs design those workflows.

Lesser-resourced languages

Although such failures affect every language pair, lesser-resourced languages expose them sooner because they have fewer opportunities to recover, and the resulting errors are more likely to persist. Even with carefully designed prompts, AI is less likely to identify and correct its own errors because fewer linguistic resources exist to detect and resolve them.As a result, translators rely more heavily on professional judgement and community-developed resources.

With regards to Haitian Creole, when translators, reviewers or clients favour different English- or French-influenced conventions, it can become especially difficult to reach a consensus because there is often no widely accepted authority to resolve competing linguistic choices. In that vacuum, allowing AI to become that authority is no substitute for linguistic consensus. Consequently, AI is often blamed for errors rooted in limited linguistic resources, unrealistic productivity expectations and weakened human oversight.

Setting AI up to succeed

AI is already improving efficiency, expanding multilingual access and increasing the digital visibility of lesser-resourced languages. These advances should be welcomed. However, better AI does not guarantee better translations. That requires equally sound linguistic infrastructure and professional workflows that support the technology. LSPs must invest not only in improved models, but also in linguistic resources and rigorous review processes that clearly distinguish MTPE, editing and quality assurance.

AI makes mistakes. Translation workflows are designed to catch them. When they fail, the issue is no longer solely technological. It is professional, operational and systemic. And when quality declines, we should ask not only ‘What did the AI get wrong?’, but also ‘What conditions were created for both AI and linguists to succeed?’



Safira Amazan is a cross-cultural communication specialist, university lecturer and general manager of Cultures Communicate.

Safira specialises in translation, localisation, cultural competence training, and medical interpretation across French, English, Spanish and Creole.

This article has been reproduced from the Autumn 2026 issue of The Linguist. Download the full edition here.