Who Checked? The Case for Certifying AI Translation

· Machine Interpreting Translation Future

Who Checked? The Case for Certifying AI Translation

Who Checked? The Case for Certifying AI Translation

AI translation does not need a licence. Like most IT services, it can be built, shipped and used without certification, and the devices that deliver it need no type approval. For the vast majority of uses, that is exactly as it should be. Translation should be available to everyone, in any language, at little or no cost, for private and professional purposes alike.

But "vast majority" is not "all". In acertain settings and use cases, a translation error is not an inconvenience but a harm, and in those settings we should stop treating AI translation as a consumer gadget. My thesis is simple: where translation carry serious consequences, especially in regulated sectors, translation systems should be accompanied by independent, verifiable evidence of their performance, and a form of certification or type approval is the right way to provide it. Three arguments support it:

  1. The stakes are high. In regulated settings, a mistranslation can change a diagnosis or a verdict, and the person affected often cannot detect it.
  2. Deployers need evidence. Institutions that rely on AI translation need a credible basis for their decisions, not vendor claims and polished demos.
  3. It creates economic value. Certification rewards quality in a commoditising market, for the companies that build these systems and for human expertise.

The principle: the greater the harm, the stronger the evidence

We already apply this logic everywhere else. A medical device cannot simply be built, boxed and shipped. Its manufacturer must show that it is safe and performs as intended, manage its risks, and monitor what happens once it is in use. If something goes wrong, there must be a way to trace and correct it.

Compare that with an everyday product such as a phone stand or a kitchen gadget. It still has to be safe, but nobody expects the same level of evidence. If it disappoints, the buyer returns it or chooses another product, and the market does most of the work.

Both approaches are rational, because the cost of failure differs. A broken phone stand is an annoyance; a faulty scanner can seriously injure someone. Regulation is not an ideological stance but a proportionality rule: the stronger the potential harm, and the less the affected person can protect themselves, the stronger the evidence we demand.

1. The stakes are high where translation is regulated

For everyday use, such as translating a menu, a newspaper article, a film or even a conference talk, AI translation is much closer to the phone holder. If the output is poor, the consequences are usually limited, and the user can notice, complain or switch tools. Nothing here calls for regulation.

Hospitals, courts and police interviews are different. These domains are already regulated, for good reasons, and the professionals working in them carry formal responsibilities. A mistranslated sentence can affect a diagnosis, a verdict or someone's future. Crucially, the people using it often cannot spot the error, because they do not speak the language in question. That is precisely the condition under which the market stops doing its work: the buyer is not the person who bears the risk, and the risk is invisible to both.

As automation reaches these sectors, driven largely by the need to cut costs, institutions will increasingly deploy AI translation where they once hired interpreters. When a human interpreter performs badly, the damage is rarely traced back to them, and I am not aware of many cases with significant legal consequences for professionals who delivered poor service. We should not import that blind spot into the automated version. Tecnologisation of translation gives us a chance to do better.

I am not suggesting that every AI translation tool needs regulatory approval. I am suggesting that where society already demands safeguards, it is reasonable to demand some form of evidence about the technology being used as well.

2. Deployers need evidence, not demos

If the stakes are high, the institution deploying a system needs a way to know whether it can be trusted. Today it mostly has vendor claims and demonstrations. The remainder of this section looks at what certification can offer, how modest a first version could be, and who should design it.

What certification can and cannot do

A first objection is that certification cannot guarantee quality. That is true, and it is not the point. Even certified products fail. Certification establishes only that a product was evaluated against defined criteria under specified conditions.

Nor is there a perfect translation to certify against. Quality has many dimensions, depends on context, and is sometimes open to debate. Two excellent interpreters may render the same sentence differently, and both can be right, or both wrong.

What certification can do is give deployers something they currently lack: independent evidence on accuracy, completeness, latency, and robustness to accents and background noise, tested for specific language pairs and conditions. This would give institutions a credible basis for their decisions, instead of vendor claims and polished demos. No more, and no less.

A realistic path: modest, voluntary, independent

A second objection is that we have never regulated human translation and interpreting this way, so why start with machines? It is a fair point. Qualifications and accreditation for human language professionals vary widely, and the field has long relied on training, experience, reputation and institutional trust, with results ranging from excellent to poor. Existing standards such as ISO 17100 are vague by nature: they cover processes, not the quality of individual translations.

So we should not expect a strict standard for AI translation when none has ever existed for humans. But that is an argument against perfectionism, not against action. Between no oversight at all and medical-device-style regulation lies a wide space.

A sensible first step would be a minimal, voluntary, independent certification stating which system and version was tested, on which languages, under which conditions, and with what results. It would not claim that a product is safe everywhere. It would simply make information comparable. Professional associations, researchers, developers and evaluation bodies could define the basic criteria together, building on existing evaluation methods. Sector-specific assessments, for healthcare or legal settings, could follow later.

Who should steer?

The hardest question is not whether, but who. In the sixth book of the Republic, Plato entrusts his political thesis to an image that would enjoy a long afterlife: a ship whose owner, strong but a little deaf and short-sighted, is beset by sailors squabbling over the helm without ever having learned the art of steering. They jeer at the one man aboard who can actually read the skies and the winds, dismissing him as a babbler lost among the stars. The true pilot, Socrates concludes, is the philosopher, and a well-ordered city should hand him the helm rather than put the course to a vote.

A few decades later, in the third book of the Politics, Aristotle takes up the same ship and turns it on its head with an almost homely observation: in some arts, the practitioner is neither the only judge nor the best one. A house is judged better by the person who lives in it than by the one who built it, a rudder better by the pilot than by the carpenter who carved it, and a banquet better by the guest than by the cook. In Plato, the pilot is the one who knows; in Aristotle, he becomes the one who uses, and that is precisely why he is entitled to say whether the instrument works.

Both views matter for certifying AI translation, and neither suffices alone. Computer scientists are experts in evaluating systems at scale, but they do not always know what counts as quality in a communicative situation. Translation and interpreting scholars understand those situations deeply, but often lack formal methods for evaluating systems. Third-party certifiers have the independence, but may lack the incentive to develop methodologies for a use case that is, commercially, relatively small. These are Plato's experts, each holding part of the knowledge.

Aristotle adds a fourth voice: the people who actually use the instrument and live with its consequences, namely the hospitals, courts, interpreters and, above all, the patients and defendants on the receiving end. A certification scheme designed only by those who build and measure systems would risk testing what is easy to test rather than what matters.

My answer is therefore a joint one. Developers, evaluation specialists, language professionals and institutional users should define the criteria together, with independent bodies performing the testing.

3. Certification creates economic value

There is also a commercial argument, and it points in the same direction. Translation is being commoditised. For general-purpose use, the output of a good machine is already close to free, and there is little reason to expect that trend to reverse. Competing on the translation itself will increasingly be competing on price, and price converges on zero.

What does not get commoditised is accountability. Where a hospital, a court or a public authority relies on a translation, it is not buying words. It is buying the assurance that someone can stand behind them. That kind of value lives in regulated settings, and it lasts only as long as those settings stay regulated. Where there are rules about who may interpret, in what conditions and with what evidence, there is a market for those who can meet them. Where there are none, the market defaults to the cheapest option.

This applies to language professionals as much as to technology vendors. Regulating deployment protects human expertise when it is the right solution to have; it creates roles that did not exist before: validating systems for specific language pairs and domains, defining test sets, supervising use in the field, and signing off on fitness for purpose. These are tasks for people who understand both communication and technology.

For developers, certification turns a weakness into an asset. Today, buyers in sensitive sectors cannot tell a careful vendor from a careless one, so careful vendors cannot charge for their care. Independent evidence changes this: it gives institutions a procurement criterion and lets serious vendors compete on proof rather than on demos.

The caveat is that regulation can also become a barrier that favours incumbents or a shield for protectionism. This is why the scheme must be proportionate, limited to high-stakes uses, and applied to machines and humans on the same evidential footing.

Start now, and keep it modest

Certification will not eliminate errors or settle every debate about quality. But it can bring transparency and accountability that are largely missing today, and it can do so without burdening the countless low-risk uses where AI translation is already doing good. We should begin building in that direction now rather than wait for a perfect framework.

We do not need to certify AI translation because it is AI. We need evidence about its performance whenever people are expected to rely on it, especially in highly regulated settings.

When an institution deploys AI translation where mistakes matter, it should be able to answer one simple question:

Who checked?