AI translation sounds convincing. That is becoming the problem
In January 2021, Spanish-speaking residents looking for information on the Virginia Department of Health website encountered a potentially dangerous machine translation error.
The English-language page said that a COVID-19 vaccine was “not required”. A Google-generated Spanish translation rendered this as “la vacuna no es necesaria”: the vaccine was not necessary.
One statement concerned whether vaccination was compulsory. The other appeared to question whether people needed the vaccine at all.
The wording was subsequently changed, but the incident exposed a problem that has become more pressing as generative AI enters professional translation: organisations can now publish information across multiple languages faster than their governance processes can verify it.
Earlier machine translation frequently advertised its shortcomings through awkward grammar and absurd word choices. Today’s AI translation systems can produce polished, authoritative text while quietly omitting a qualification, weakening an obligation or selecting the wrong specialist meaning.
The risk is no longer confined to embarrassing language. It extends to public safety, financial disclosure, regulatory compliance and legal liability.
How does AI translation work?
Automated translation does not simply consult a bilingual dictionary and apply a fixed set of grammatical rules. Modern systems learn patterns from large collections of text.
Earlier statistical machine translation divided sentences into words or phrases and used probabilities derived from bilingual material to select likely equivalents. The results were often awkward, but their mechanical origins were usually visible.
Neural machine translation changed the process. Text is divided into tokens, which may be words, parts of words or punctuation. The system converts those tokens into numerical representations and processes their relationships before generating a sequence in the target language.
Transformer models, which underpin many current translation systems and large language models, use a mechanism called attention. This allows the model to give different weight to different parts of the input when generating each part of the output. A pronoun near the end of a paragraph, for example, can be interpreted in relation to a noun that appeared earlier.
This produces more natural language and better handling of context. But the system is still generating an output from patterns learned during training and information supplied in the source. It is not retrieving a guaranteed translation from an authoritative record.
A purpose-built machine translation engine is optimised primarily for translation. A general-purpose large language model has also been trained to answer questions, summarise, rewrite and generate text.
Faced with an unusual phrase, missing context or apparent contradiction, a generative model may produce the sentence that appears most plausible rather than the one most faithful to the original. It does not necessarily warn the user that it has resolved an ambiguity in this way.
The same process helps explain why an incorrect AI translation can sound so convincing. The model is highly capable of producing probable, well-formed language. Grammatical fluency is part of the generation process, while precise fidelity remains a separate challenge.
Why does AI translation make mistakes?
Two underlying problems help explain many machine translation errors: bias in the available data and weak grounding in the situation being described.
Data bias
AI translation reflects the data used to build and refine it. Languages with abundant, high-quality parallel texts are generally easier to support than languages, dialects and specialist domains for which reliable training and evaluation material is scarce.
When a dialect, idiom or community-specific use of a word is poorly represented, the model has fewer relevant patterns from which to work. It may normalise the expression into a dominant form of the language, select an apparently similar standard term or generate a plausible meaning that was not intended.
Modern models process tokens and contextual relationships rather than simply looking for the nearest word in a dictionary. The practical result can nevertheless be similar: language used by a smaller community may be interpreted through patterns learned predominantly from larger ones.
Technical language creates a related difficulty. A word may have one meaning in everyday conversation, another in engineering and a third in law. “Consideration”, for example, has a specific meaning in contract law. “Going concern” cannot safely be treated as two ordinary English words.
Without sufficient domain-specific material, approved terminology or surrounding context, a system may select the common meaning rather than the professionally correct one.
Weak real-world grounding
Human translators bring more than knowledge of two languages. They draw on experience of the physical world, cultural conventions, human intentions, institutional practice and historical context.
AI models can encode and reproduce extensive information about the world, and they can infer context surprisingly well. But they do not experience the situation being described or automatically know what the writer intended.
When the source is ambiguous, a human translator may recognise that an instruction would be physically impossible, that a proposed legal interpretation makes little sense or that an expression refers to a culturally specific event. The translator can consult a specialist or ask the client for clarification.
An AI system may instead resolve the ambiguity by generating a statistically plausible output. Because fluency and real-world reliability are not identical capabilities, the result can be grammatically impeccable yet physically, culturally or professionally illogical.
Short text can be especially difficult because it provides little context. A button labelled “charge” could refer to payment, electrical charging or making an accusation. A single omitted “not” can reverse an instruction. Companies nevertheless frequently have to translate isolated spreadsheet cells, software strings or fragments of manuals.
Why minority languages face greater translation risk
AI translation errors are also an inclusion issue. The people most dependent on translated public information may receive the least reliable version of it.
Widely used languages such as English, Spanish, French and German benefit from large quantities of digital material, established terminology and substantial commercial markets for translation. Minority languages, regional varieties and dialects may be represented by smaller or less balanced datasets.
Public authorities face an uncomfortable incentive. Providing professionally translated information in every relevant language is expensive and can take time, particularly when relatively few residents speak each one. Automated translation can expand language coverage without increasing budgets at the same rate.
Yet the languages for which authorities have the fewest qualified reviewers may also be the ones for which automated output most needs checking.
The possible consequences are visible in immigration systems. In 2023, Rest of World reported on a Pashto-speaking asylum applicant in the US whose written statement appeared to contradict an earlier interview. According to the crisis translator who later examined the documents, an automated tool had changed first-person references from “I” to “we”. The discrepancy was reportedly significant enough to contribute to the rejection of the claim.
This is a reported case based on the account of an advocate, rather than a published court finding about the translation system. It nevertheless illustrates how a small linguistic alteration can influence an official assessment of credibility.
A model may also handle the dominant or formal form of a language more reliably than a regional dialect, Indigenous variety or the language used by a displaced community. When officials cannot read the original themselves, they may have no way to recognise that the English translation is misleading.
Automation can make information available in more languages, which is valuable. But if an automatically translated page is treated as equivalent to a professionally verified version, an organisation may provide formal access without providing equal access to meaning.
Fluent AI translations can still be unfaithful
Researchers are developing better ways to measure translation failures. In 2026, an international team introduced HAT, a dataset containing 350,959 professionally annotated samples across 38 language pairs. Its focus is hallucination: output that may read fluently but is not faithful to the source.
Researchers have also identified “harmful factuality hallucination”, in which a large language model corrects what it perceives as a factual error in the source. The resulting statement may be factually more plausible, but it is no longer a faithful representation of the input.
Translation quality consequently has several dimensions. Fluency describes whether the result sounds natural. Fidelity asks whether it preserves the original meaning. Fitness for purpose asks whether it is suitable for a particular audience and use.
Businesses may be evaluating only the easiest of the three.
How a translation error changed Sharp’s financial message
Financial reporting offers another illustration of the consequences of specialist terminology.
In November 2012, Sharp published an English version of a difficult quarterly earnings report in Japanese. It said that “material doubt” existed about the Japanese electronics manufacturer’s ability to continue as a going concern.
“Going concern” has a specific meaning in financial reporting. It relates to whether an organisation can continue operating for the relevant assessment period rather than being forced to liquidate or cease trading.
Four days later, Sharp revised the English text. The corrected version still acknowledged huge losses and negative operating cash flow. However, it said that conditions existed which might create uncertainty, while adding that Sharp believed its countermeasures would resolve them.
The episode nevertheless shows what can happen when a translation changes management’s stated assessment. AI systems can now translate earnings material and policy announcements almost immediately. That speed increases the importance of applying appropriate disclosure controls to every language version on which investors may rely.
Merely labelling a document “AI translated” would not resolve questions about responsibility, accuracy or how the translated information was presented to investors.
Who is liable for an AI translation error?
Consider a mistranslated warning in an industrial maintenance manual. If it contributes to an injury, the manufacturer cannot assume that responsibility automatically passes to its language supplier or AI provider.
The EU Machinery Directive currently requires machinery to be accompanied by instructions in the official language or languages of the member state in which it is placed on the market or put into service. Manufacturers also retain responsibility for ensuring conformity with the applicable safety requirements.
The incoming EU Machinery Regulation, scheduled to apply from 14 January 2027, continues to place conformity responsibilities on manufacturers and includes language requirements for instructions and other information.
A manufacturer might seek compensation from a translation supplier if an error resulted from a breach of contract or negligent professional work. The outcome would depend on the applicable law, what service was purchased, which controls were promised and whether the customer supplied accurate source material and specialist terminology.
Responsibility becomes harder to trace when an employee independently pastes text into a general AI tool. There may be no approved workflow, human reviewer or record of which system and prompt produced the translation.
The provider’s terms may limit warranties or liability for output. A localisation company may never have seen the text. The employee’s organisation can therefore be left carrying a risk it did not formally assess.
Does insurance cover AI translation mistakes?
Professional translators and localisation companies may carry professional indemnity, also known as errors and omissions insurance. Depending on its wording, such a policy can cover defence costs and claims alleging that negligent professional work caused a client financial loss.
It should not be treated as a universal guarantee.
Policies differ considerably. Some may exclude or restrict bodily injury, recalls, contractual penalties, regulatory fines or liabilities that the insured accepted solely through contract. Cover may also depend on the services declared to the insurer and the policyholder following agreed professional procedures.
A supplier insured for conventional translation should therefore confirm with its insurer or broker whether AI-assisted services fall within the scope of its cover.
The customer’s policies raise separate questions. Product-liability insurance may become relevant if defective instructions contribute to an injury. Cyber cover may be relevant if confidential or personal information is exposed through an unauthorised translation platform. Recall expenses and business interruption may require separate insurance.
Whether any policy responds will depend on its wording, jurisdiction and the circumstances of the incident. Procurement teams should therefore look beyond a certificate confirming that insurance exists. They may also need to examine the covered activities, geographical scope, exclusions and claim limits.
How can companies reduce AI translation risk?
Avoiding translation errors does not require every sentence to be translated from scratch by a human.
Routine internal material will generally present less risk than a contract, financial disclosure, medical instruction or machinery warning. Companies can use this difference to determine where automation may be acceptable and where specialist human involvement remains appropriate.
Companies should also avoid assuming that a system performs equally well across languages. Strong results in French or German do not establish comparable reliability in Frisian, Pashto or a regional Arabic variety.
Keeping a record of the approved source, translation system, terminology resources, changes and final approval can help an organisation investigate problems later. An incident process can also make it easier to find and correct the same error across manuals, packaging, websites and software interfaces.
International standards may provide a useful reference point. ISO 17100:2015 sets requirements for translation services, ISO 18587:2017 covers post-editing of machine translation output and ISO 5060:2024 provides general guidance on evaluating translation output. A revised version of ISO 18587 addressing post-editing of non-human translation output is currently under development. Certification cannot guarantee that every sentence is correct, but standards can help establish defined roles and processes.
There is no single safe workflow for every organisation or document, but companies you can ensure that the level of oversight reflects the possible consequences of an error—and that using AI does not make responsibility for the final translation disappear.
AI translation is already good enough to become routine. It now needs governance strong enough to prevent routine use from creating exceptional harm.
Further reading on MoveTheNeedle.news: