AI Hallucinations in Translation: When Cutting Costs Comes at A Price

AI translation can save businesses time and money. From product descriptions to legal contracts and technical manuals, companies are increasingly using AI to translate content that once required extensive manual work.

But there's a catch: a translation can sound perfectly natural and still be dangerously inaccurate.

One risk is AI hallucinations, when AI models generate information not supported by the original text.

For businesses dealing with specialized content, these inaccuracies can have serious consequences, from compliance issues and financial losses to potential safety risks.

What Are AI Hallucinations in Translation?

AI hallucinations occur when an artificial intelligence model generates information that isn't supported by the data it is supposed to rely on.

In machine translation, this can happen when the translated text contains information that doesn't appear in the source. The result may be grammatically correct, well-written, and entirely natural in the target language, yet significantly different in meaning from the original.

However, it's key to distinguish AI hallucinations from other types of translation errors.

An omission, an incorrectly translated technical term, or a misinterpreted sentence isn't necessarily a hallucination. Nevertheless, these errors can be just as damaging to translation accuracy and reliability.

Translation quality isn't simply about how naturally a text reads. It's about how faithfully it conveys the original meaning.

How can AI Hallucinations Affect a Business?

Legal translation: When AI invents a legal provision

A Greek company uses AI to translate a supplier agreement into English.

In a clause concerning contract termination, the model inserts a reference to a nonexistent legal provision, presenting it as the legal basis for ending the agreement without prior notice.

The fabricated reference blends seamlessly into the English translation, making it difficult to spot without checking the original contract.

If the company relies on the translation, it may misunderstand its contractual rights and obligations or make decisions based on a non-existent legal provision.

The critical point: Checking whether the legal terminology is correct isn't enough. Each clause must be compared against the source text, and legislative references must be verified to catch even those errors that appear entirely convincing.

Technical translation: When a fabricated instruction appears legitimate

Imagine a manufacturing company using AI to translate an equipment maintenance manual.

The model adds an instruction that doesn't appear in the original and presents it as part of the prescribed maintenance procedure.

An employee relying solely on the translation might assume the instruction is valid and follow it.

In technical documentation, an invented instruction isn't just a translation error. It can have real-world consequences, potentially affecting equipment operation, compliance with established procedures, or employee safety.

These scenarios are hypothetical, but the risks they illustrate reflect documented limitations of machine translation.

What Does Scientific Research Say about AI Hallucinations?

Hallucinations in machine translation have been the subject of scientific research.

In their 2023 study, Hallucinations in Large Multilingual Translation Models, Guerreiro et al. investigated this phenomenon across different AI translation models and more than 100 language pairs.

The researchers found that the models they examined were particularly vulnerable to hallucinations when translating into languages with limited available linguistic data, as well as when translating from English into other languages.

The study also identified instances where models generated content absent from the source text, including toxic content.

These findings don't mean that all AI models behave in the same way or that every AI-generated translation contains hallucinations.

They do, however, highlight a critical challenge for businesses: fluent AI-generated translations can contain errors that aren't immediately obvious without checking the source text.

For companies working with legal, technical, or regulatory documentation, this makes translation accuracy a matter of more than linguistic quality.

The True Cost of An Inaccurate Translation

One of the main advantages of AI translation is its potential to reduce upfront costs.

However, assessing a translation's financial value shouldn't stop at the production cost.

Depending on the content and its intended use, a translation error can lead to:

• Regulatory compliance issues: Incorrectly translated legal or technical requirements can create confusion about a company's obligations.
• Safety risks: Inaccurate operating instructions, maintenance procedures, or safety warnings may put employees or equipment at risk.
• Additional expenses: Correcting errors may require document revisions, reissues, or reviews of existing procedures.
• Damage to business credibility: Inaccurate information reaching customers, business partners, or regulatory authorities can undermine trust.

The cost of these consequences can far exceed the initial savings.

The question, therefore, isn't simply how cheaply a translation can be produced, but whether it's reliable enough for its intended purpose.

AI & Human Translation: Competition or Collaboration?

The possibility of translation errors doesn't mean businesses should exclude artificial intelligence from their workflows.

AI tools can help process large volumes of content more quickly and support various stages of the translation process.

What matters is how you use these tools and the level of review each project requires.

A short internal communication, for example, may not require the same level of scrutiny as a legal contract, technical manual, or regulatory document.

For specialized content, professional translation review may include:

• Source-text comparison: Identifying additions, omissions, and changes in meaning.
• Terminology verification: Ensuring specialized terms are translated accurately and consistently.
• Verification of valuable information: Checking legal references, technical specifications, and safety instructions.
• Linguistic and functional review: Ensuring the final text is clear, accurate, and appropriate for its intended audience.

This process goes well beyond correcting spelling or grammar. It requires a thorough understanding of the content and how it will be used.

Artificial intelligence can be a valuable tool. However, specialized human expertise remains paramount when accuracy matters most.

Choose Words Collective for Technical Translation Quality

At Words Collective, technical translation starts with understanding the source material, its terminology, and how the translated content will be used.

Whether we're translating technical documentation or reviewing machine-translated content, our focus is on preserving meaning, checking terminology, and identifying inaccuracies that could otherwise go unnoticed.

We pay close attention to the details that matter in specialized texts, from consistent technical terminology to accurate procedures, specifications, and safety-related information. Our goal is to help businesses communicate accurately and consistently across languages, without sacrificing credibility for speed.

After all, creativity is part of what we do at Words Collective. Inventing technical specifications isn't. We'll leave that to science fiction.

Need professional technical translation or a review of AI-generated content? Contact Words Collective to discuss your project's requirements.