Most companies already use AI to translate something. A supplier's email, a product page, a slide deck for a prospect, the occasional contract. It makes sense: it's fast, it's cheap, and the output usually reads beautifully.
That last point is exactly where the trouble starts. A translation that reads well is not necessarily a translation that is right. This guide looks at the real AI translation risks, the errors that slip through most often, which content is safe to hand to a machine, how to keep quality under control, and a question few people ask until something goes wrong: who is responsible when an AI translation is wrong?
To be clear from the outset, this is not an argument against AI. It's an argument for using it with a safety net.
Why companies now translate with AI
Neural machine translation engines and assistants built on large language models (LLMs) have improved dramatically in a short time. The benefits deserve an honest hearing:
- Speed. Turnaround drops from days to seconds.
- Cost. Translating large volumes no longer needs its own budget line.
- Self-service. Anyone on the team can understand a document in another language without waiting for help.
- Natural output. The text flows, with grammar and phrasing that would have been unthinkable a few years ago.
- Reach. Small teams can run a multilingual website or support customers in several languages.
For plenty of tasks, AI is simply the right tool. The real question isn't whether to use it, but for what and with what checks.
Fluent output hides errors
Older machine translation gave itself away. Garbled sentences and broken grammar told everyone the text needed fixing. Today the opposite is true. AI produces sentences that look polished, and polish breeds trust that isn't always earned.
A changed figure, a vanished negative or a sentence that quietly drops out does nothing to spoil the flow. Unless someone compares the translation with the source, line by line, nobody notices.
The risk with AI translation isn't that it gets things wrong. It's that it gets things wrong in a way that looks right.
So, is AI translation reliable? There's no yes-or-no answer. It's reliable for getting the gist, drafting and much internal content. It stops being reliable when the text is going to be signed, published or submitted to an authority without anyone checking it.
Six common AI translation errors
These are the error types that come up again and again in machine translation and LLM output. Every example below is invented for illustration and has nothing to do with any client's content.
1. Modal verbs that change legal force
Illustrative example.
- Source (EN): The Licensee shall pay all fees within 30 days of the invoice date.
- AI output (ES): El Licenciatario podrá abonar todas las tarifas en un plazo de 30 días desde la fecha de la factura.
- Correct: El Licenciatario deberá abonar todas las tarifas en un plazo de 30 días desde la fecha de la factura.
"Shall" creates an obligation; "podrá" means "may". The Spanish reads perfectly, yet a duty to pay has become an option to pay.
2. Dates and number formats
Illustrative example.
- Source (EN, UK document): Completion date: 03/04/2027.
- AI output (ES): Fecha de finalización: 4 de marzo de 2027.
- Correct: Fecha de finalización: 3 de abril de 2027.
The tool read a British date as if it were American and shifted the deadline by a month. Decimal points and thousands separators carry the same risk: €1,500.00 in English is 1.500,00 € in Spanish, and swapping conventions can change an amount entirely.
3. Negatives that disappear
Illustrative example.
- Source (EN): Do not apply to broken or irritated skin.
- AI output (ES): Aplicar sobre piel dañada o irritada.
- Correct: No aplicar sobre piel dañada o irritada.
One missing word and a warning becomes an instruction. It's not the most frequent error, but it's among the most serious, and it's almost invisible if you only read the translation.
4. Omissions: the clause nobody misses
Illustrative example.
- Source (EN): Neither party shall be liable for indirect losses, except in cases of gross negligence or wilful misconduct.
- AI output (ES): Ninguna de las partes será responsable de los daños indirectos.
- Correct: Ninguna de las partes será responsable de los daños indirectos, salvo en caso de negligencia grave o dolo.
Everything the translation says is accurate. The problem is what it leaves out: an exception that changes the whole allocation of risk. Because the sentence still makes sense, omissions are among the hardest errors to catch.
5. Added content (hallucinations)
Illustrative example.
- Source (EN): Returns are accepted within 14 days of delivery.
- AI output (ES): Se aceptan devoluciones en un plazo de 14 días desde la entrega, sin necesidad de justificar el motivo y con reembolso íntegro.
- Correct: Se aceptan devoluciones en un plazo de 14 días desde la entrega.
Language models sometimes "complete" a text with information that isn't in the source. Here, the result is a commercial promise the company never made.
6. Broken placeholders and ignored terminology
Illustrative example.
- Source (EN): Hi {first_name}, your order {order_id} has been dispatched.
- AI output (ES): Hola {nombre}, tu pedido {id_pedido} ha sido enviado.
- Correct: Hola, {first_name}: tu pedido {order_id} ya está en camino.
The variables were "helpfully" translated, so the software can no longer fill them in. The same thing happens with glossary terms: a product name, a job title or a defined term in a contract rendered three different ways across one document.
Other recurring problems include the wrong register (formal versus informal address), a tone that doesn't match the brand, untranslated fragments and text that no longer fits its table cell or slide. If contracts are your main concern, our upcoming article covers the most common machine translation errors in contracts in more detail.
Low, medium and high-risk content
Not every text carries the same risk. A useful test: if this translation contained an error, who would read it, and what would happen next?
| Risk level | Typical content | Suggested control |
|---|---|---|
| Low | Understanding incoming emails or documents, internal notes, drafts, background research | AI on its own, or a quick read by someone who knows the subject |
| Medium | Website copy, product descriptions, newsletters, sales decks, customer support replies, non-critical technical documentation | Glossary and style guide, plus full post-editing or professional revision |
| High | Contracts, terms and conditions, regulatory submissions, medical and pharmaceutical information, safety instructions, investor information, documents for public authorities | Specialist translator plus revision by a second linguist; sworn translation where the procedure requires it |
The rule of thumb is simple. The further a text travels from your own team, and the closer it gets to a customer, a court, a patient or a regulator, the more control it needs.
Official procedures are a special case. In Spain, a sworn translation (traducción jurada) must be signed by a sworn translator appointed by the Spanish Ministry of Foreign Affairs. An AI translation is not a sworn translation and won't be accepted as one. We look at this in more depth in Is an AI-translated document legally valid?
How to keep AI translation quality under control
Translation quality control doesn't need to be invented from scratch. The industry already has well-established methods and international standards.
Give the AI context: glossaries and style guides
Many errors can be prevented before translation starts. A glossary of key terms (including those that must stay untranslated), a style guide that sets register and tone, and a line or two about the audience all improve the output. They don't replace checking, but they make it faster.
Light versus full post-editing
Post-editing means a human correcting machine translation output. There are two broad levels:
- Light post-editing: fixes errors that block understanding or change the meaning. The result is usable but not polished, which is fine for internal use.
- Full post-editing: aims for a text comparable to human translation: accurate in meaning, terminology, style and formatting. This is the right level for anything you publish.
Revision versus review
These two words are often used interchangeably, but they describe different jobs. Revision is a bilingual check: a second professional compares the translation with the source, segment by segment. It's the only reliable way to catch omissions, altered figures and lost negatives. Review, in the sense used by ISO 17100, is a monolingual read of the target text to check it's fit for purpose. Useful, but it can't spot what isn't there.
Measuring quality: MQM and ISO 5060
To move from "it looks fine to me" to something you can measure, the industry uses error typologies. The best known is MQM (Multidimensional Quality Metrics), which classifies each error by category (accuracy, fluency, terminology, style, locale conventions and so on) and by severity (minor, major or critical). ISO 5060:2024 provides guidance on evaluating translation output, whether human, machine or post-edited, using a similar approach: error types and penalty points weighted by severity. With either, you can compare tools, suppliers or workflows on objective grounds.
ISO 18587 and ISO 17100 in brief
- ISO 18587:2017 sets out requirements for the full human post-editing of machine translation output and the competences of post-editors. A second edition, widening its scope to AI and LLM output, is in preparation.
- ISO 17100:2015 covers professional translation services: translator qualifications, project management and, crucially, mandatory revision by a second linguist. It explicitly excludes raw machine translation plus post-editing from its scope, which is where ISO 18587 comes in.
In practice, the sensible approach combines the two: AI and post-editing for speed, and revision by a second person whenever the content calls for it.
Who is responsible when an AI translation is wrong?
This is the uncomfortable question. If a mistranslated clause sparks a dispute, a product page promises something it shouldn't, or a set of instructions loses a safety warning, who answers for it?
The short answer: the company that publishes, signs or sends the text. The tool doesn't sign the contract or deal with the complaint. As far as your customer, partner or regulator is concerned, the document is yours, whether a person, a machine or both translated it.
Beyond that, the full answer gets complicated. It depends on your contracts with any language service provider, your sector, the jurisdiction and the type of document. The regulatory picture is also shifting. The EU AI Act (Regulation (EU) 2024/1689) emphasises transparency, human oversight and the accountability of organisations that deploy AI systems. Since 2 February 2025, for instance, Article 4 has required organisations using AI systems to take measures to ensure a sufficient level of AI literacy among their staff. None of this is translation-specific, but the direction of travel is clear: using AI doesn't remove the need to oversee what it produces.
There's no single answer to close this debate, but there is a practical rule: for high-risk content, have a professional check the translation before it leaves your organisation. Then, if anyone asks "who checked this?", you'll have an answer.
Before you publish: what to check
Before you sign off an AI translation, compare it with the source and check at least these: figures and dates, negatives, obligations (words like "shall" and "may"), missing sentences, terminology, formatting and how risky the content is.
TRIDIOM has been translating from Madrid for more than 25 years. We're certified to ISO 17100, ISO 9001 and ISO 27001, and our sworn translators are appointed by the Spanish Ministry of Foreign Affairs. If your team already translates with AI and wants a safety net, we can help with professional post-editing and specialist revision for legal, medical and sworn translation.
Translating with AI already? Tell us how you check the output at info@tridiom.com.
AI translation: frequently asked questions
Is AI translation reliable?
It's reliable for understanding a text, producing drafts and translating internal content. For anything that will be published, signed or submitted to third parties, it isn't enough on its own, because it can make meaning errors that don't show up on a quick read. That content needs post-editing or professional revision.
What are the main AI translation risks?
The most common are altered numbers, dates and amounts; modal verbs that change obligations; missing negatives; omitted sentences or clauses; added content that isn't in the source; and inconsistent terminology. How much they matter depends on the type of content.
Can I use an AI translation for an official procedure?
Not where a sworn translation is required. In Spain, a sworn translation must be signed by a sworn translator appointed by the Ministry of Foreign Affairs. An AI translation can serve as a working draft, but it has no validity as a sworn translation.
What's the difference between post-editing and revision?
Post-editing is human correction of machine-translated text, either light or full. Revision is a bilingual comparison of a translation against its source by a second professional, as required by ISO 17100. For high-risk content, the two can be combined.
Who is responsible if an AI translation contains errors?
In dealings with customers, partners or authorities, it's the company that publishes or uses the text, not the tool. The exact split of liability depends on contracts, sector and applicable law, which is why high-risk content should be checked by a professional.
