Why AI alone is not enough for medical translations

Automated translation has come a long way. But it’s simply too risky for critical medical translations. Raw AI output won’t stand up to an audit, a quality review, a regulatory submission, or the standard of care that patients deserve. Let’s take a look at the risks, and what a safer workflow looks like. 

When is an AI translation the right solution? 

AI translations have gained traction for a few reasons, namely speed and cost. First, let’s look at some scenarios where you might want to use raw AI output that hasn’t been reviewed by a human. 

• If you just want to get the gist of a non-critical document; the translation will never be published or distributed, and a minor mistake in the translation isn’t critically important. Keep in mind that the quality of AI translation varies a great deal by language, because the large language model (LLM) has a lot more data to work with in Spanish than in Amharic. 
• If you’re doing first-pass research or triage, to identify documents that need to be sent for human translation or review. Everything doesn’t need to be your highest priority, and an AI tool can help you categorize your documents. 
• If you’re translating informal, non-medical communications. Maybe you’re e-mailing or messaging a colleague in another country about a low-stakes topic, and the subject is nothing with legal or clinical significance. An AI tool can be a big help. 

Now, let’s look at some scenarios where AI-only translation falls short. 

AI’s shortcomings: the big picture 

ChatGPT stormed onto the scene in late 2022, promising unprecedented speed and scale gains. And when it comes to translation, these gains are needed. KFF reports that 31% of immigrants with Limited English Proficiency (LEP) have faced barriers to accessing health services, and are twice as likely (28% versus 14%) as fluent English speakers to report that they are in fair or poor health. 

And yet, faster and cheaper translations aren’t the solution. The following table shows some of AI’s disadvantages when it comes to medical translation. 

It’s not hard to see the potential impacts of these issues:

• Fines and legal repercussions due to HIPAA violations 
• Patients are provided with incorrect information and don’t understand how to take their medications, resulting in a repeat visit or even a medical emergency
• Sensitive topics such as a life-altering diagnosis are handled clumsily, resulting in emotional stress and a loss of trust in the healthcare system 

Medical translation isn’t just about the words

AI translation might be a good option if it were simply a question of word replacement: English word=Spanish word. But that’s not the case. Additionally, AI translations often sound very fluent when they’re incomplete, ambiguous, or even completely incorrect. 

Let’s take the example of a translated prescription label, which some people might think of as the simplest form of medical translation. But even in English, as reported in a Science Daily article, prescription labels serve pharmacies more effectively than they serve patients. Many patients literally cannot read their prescription labels, even in their own language. In a study of 85 prescription labels, the vast majority (84%) gave prominence to the pharmacy’s name or logo, versus the dosing instructions, which often appeared in nine-point type, despite the U.S. Pharmacopeial Convention’s recommendation to use 12-point type. The USP also reports that many patients misinterpret dosing instructions such as “take two pills twice daily,” taking a total of two pills instead of four. 

Now, add an AI-generated translation into the mix, potentially provided to the patient as a separate printout rather than on the prescription bottle itself. 

Medical translation needs to convey a message, not just individual words. Even the concept of “pain” can be a good example. Many languages and cultures have different words for physical and emotional pain, and terms commonly used in U.S. English (sharp pain, stabbing pain, throbbing pain) may not translate directly. An AI tool doesn’t know this; AI-generated translations also have a tendency to confuse terms like “dreaming” (as in, while you’re asleep) and “dreaming” (as in, your hopes and dreams). 

Regulatory and compliance risks

Regulatory authorities, including the U.S. Food and Drug Administration, and the European Medicines Agency, mandate that translations must be accurate, clear, and culturally appropriate. This means that automated translations without human oversight are generally prohibited or rejected. 

The U.S. Department of Health and Human Services states that “…machine translation…for critical documents…must be reviewed by a qualified human translator to ensure accuracy, when accuracy is essential, when the source documents or materials contain complex, nonliteral, or technical language, or when the underlying text is critical to the rights, benefits, or meaningful access” of LEP individuals. Free, public AI engines are also non-compliant with HIPAA and other medical privacy laws. In short, unreviewed AI translations, particularly if generated by a free, online AI tool, are simply too risky. 

Another issue is the lack of context that LLMs are working with, when dealing with complex texts. In its guidance on safe and responsible use of LLMs, the European Medicines Agency notes that, “LLMs may not have been exposed to information that answers a scientific or regulatory question, due to the novelty of many tasks performed by regulators (e.g. those concerning new active substances).” 

On the Language Scientific blog, we recently reported on the use of AI for specific regulatory purposes, including the European Union’s new In Vitro Diagnostic Medical Devices Regulation (IVDR). AI tools can help, they can hurt, and subject-matter experts are still key to a safe and compliant finished product. 

FAQs on AI for medical translation, and what a safer workflow looks like 

Q: Can AI be used for medical translation?
A: Yes, AI can be used for some types of medical translation. But AI alone is not enough. You need an expert assessment of whether you need AI-leveraged, AI-assisted, or human-only translation. This depends on the languages you are translating from and into, the purpose of the translation, the complexity of the document, and the risk level of the document. 

Q: Is AI translation accurate for medical documents?
A: AI translation has come a long way, but medical documents often include highly technical vocabulary, and even novel terms for new drugs and devices. Most large language models (LLMs) are trained on general material, not specialized medical content. 

Q: What are the risks of using AI for medical translation?
A: The main risk of using AI for medical translation is an error slipping through in a high-risk document. For low-risk documents, a minor translation error may not be a big deal. But in clinical trial protocols, patient intake questionnaires, medical instrument manuals, and patient-reported outcome documents, an unreviewed AI translation is too risky. 

Q: Do AI-translated medical documents still need human review?
A: Most  AI-translated medical documents do need to be reviewed by an expert human. AI-translated documents can contain errors, and regulatory authorities like the U.S. Food and Drug Administration and the European Medicines Agency caution against, or even prohibit the use of unreviewed AI translations in regulatory submissions. 

Q: How do medical translation companies use AI safely?
A: Safe use of AI in medical translations means keeping human experts at the controls. We know that clients want to benefit from AI’s speed, scale, and scope advantages, while minimizing risk and keeping patients safe. This is possible, but it demands an AI risk evaluation from a highly qualified translation provider. 

It’s an exciting time in the world of clinical trials, drug development, and medical device innovation. A recent article by the American Hospital Association states, “Patient recruitment cycles that used to span months are shrinking to days; study builds that took days now take minutes.” At the same time, it’s unwise to ignore the risks. As discussed on the Language Scientific blog, public AI engines don’t comply with confidentiality, security, and accountability risks in the medical world. By following the strategies described above, you’ll be on your way to leveraging AI’s speed and scale gains, while keeping patient safety and regulatory compliance at the forefront.