will AI replace translators?

The tasks where you're genuinely irreplaceable aren't marginal edge cases. They're 13 of the 17 tasks O*NET identifies for this role. That's a clear majority. And they're not interchangeable with each other.

Think about what it takes to identify and resolve a conflict between two languages at the level of concept or cultural practice. A Spanish legal term like "amparo" doesn't map cleanly onto English law. Choosing how to render it requires understanding two legal systems, the context of the document, and the expectations of the reader. No current AI does this reliably. The same applies when you're conferring with an author to make sure a translation keeps the feeling of the original, not just the literal meaning. That's a judgment call about tone, intent, and reader experience. It's also a relationship. The author needs to trust you.

The supervisory and educational tasks matter too. Training junior translators, educating school staff about the role of educational interpreters, guiding tourists who speak another language — these require presence, authority, and the ability to read a room. According to the Anthropic Economic Index, tasks involving direct human coordination and cultural mediation score among the lowest for AI penetration across all occupations. That's exactly what this work involves. If you're building expertise in legal, medical, or technical translation, you're in the safest part of the profession. If you're doing high-volume, low-context document translation for commodity clients, that's the part feeling the most pressure right now.

The core problem with AI translation isn't speed or vocabulary. It's accountability. When a medical interpreter mistranslates a dosage instruction or a legal translator renders a contract clause incorrectly, someone is liable. AI tools produce plausible-sounding output with no mechanism for catching their own errors in context. They can't tell when a term is technically correct but pragmatically wrong for the specific reader or situation.

Hallucination is a real problem in this field. AI translation tools occasionally produce fluent, confident output that's factually or legally wrong. In a scientific document or a court filing, that's not a minor nuisance. It's a professional and legal risk. Models trained on general web text also struggle with domain-specific terminology that hasn't been well-documented online in both languages simultaneously. Low-resource language pairs, regional dialects, and specialised fields like patent law or pharmaceutical labelling are all weak spots.

There's also the cultural layer that AI consistently misses. Adapting software to another language and culture, for instance, means knowing what a user in that culture expects from an interface, not just what the words say. AI can translate the strings in an app. It can't tell you that the phrasing will read as rude or confusing to a native speaker of that region. That gap closes slowly because it isn't a data problem. It's a judgment problem.

The two tasks AI handles most confidently are reading source documents and rewriting them in a target language, and adapting educational translations to specific grade levels. These are real capabilities today, not theoretical ones. Tools like DeepL can produce first-draft translations of legal documents, news articles, and business content at a quality level that would have been unusable five years ago. For high-resource language pairs like English-Spanish, English-French, or English-German, the output is often good enough to publish with light editing.

For simultaneous and consecutive interpretation, AI is less dominant but is genuinely speeding up the workflow. Tools like Kudo and KUDO AI are being used in conference settings to provide AI-assisted interpretation support. They don't replace skilled human interpreters in high-stakes environments, but they lower the barrier to providing multilingual access in lower-stakes settings like corporate meetings or webinars. Proofreading and editing translated materials is also faster now. Tools like Lilt and memoQ use translation memory combined with machine suggestions to cut revision time on large documents.

For document translation at volume, the workflow has shifted from translate-from-scratch to post-edit-machine-output. That's a real change in how the work gets done. Some clients now pay for post-editing at a lower rate than full translation, which affects earnings for translators who work in that segment. Knowing how to work efficiently inside tools like SDL Trados or Phrase (formerly Memsource) is now a baseline expectation for many professional translation roles, not an optional skill.

The rhythm of your day has shifted most noticeably at the document translation end. If you work with a CAT tool like SDL Trados or memoQ, you're spending less time generating first drafts and more time making decisions about whether the machine's output is actually right for this document, this client, and this audience. That's a different cognitive task than translating from scratch, and it's faster in volume terms but it can be more fatiguing because you're constantly evaluating, not creating.

What hasn't changed at all is the front end and the back end of complex projects. The initial brief, the conversation with the client about tone and audience, the decision about how to handle an untranslatable concept, the final read-through where you're asking whether this feels right in the target language — none of that has been touched. Interpreters working in healthcare, courts, or live conference settings report almost no change to the actual experience of doing the job. The human is still in the room, still making real-time calls.

What you're spending more time on, if you're adapting, is quality control and specialisation. The translation work that pays well now is the work where someone needs a human expert, not just fluent output. That means clients are more likely to come to you with harder problems, not routine ones.

The BLS projects 1.7% growth for translators and interpreters through 2034. That's not a collapsing market, but it's well below the 4% average across all occupations. With 75,300 people currently employed and about 6,900 annual openings, there are still jobs. But the growth number needs context: demand is being sustained by interpreting and specialised translation work, not by the document translation segment where AI has had the most impact.

The segments holding up best are healthcare interpreting, legal interpreting, and immigration services. These are areas where a mistake has consequences, where accountability matters, and where in-person or real-time human presence is often legally required. Sign language interpreters are in a different position from document translators — demand there is growing faster, and remote video interpreting services like Sorenson and ZVRS have created new channels for work. According to BLS occupational data, states with large immigrant and refugee populations are seeing steady demand for community interpreters regardless of AI trends.

The segment under real pressure is general document translation, especially in common language pairs. If your work is primarily translating English marketing copy into French or Spanish at volume, the economics are worse than five years ago. Clients who used to pay for full translation are increasingly paying for post-editing rates or using AI without human review at all for internal documents. That's not speculation. It's already happening in the translation services market, and the shift is reflected in flat freelance rates for generalist document work across major platforms.

The exposure score for this role sits at 57%, and it's more likely to rise than fall over the next five years. The main driver will be continued improvement in machine translation quality for high-resource language pairs and an expansion into more language pairs as training data improves. If multimodal AI gets better at real-time audio translation, the lower end of conference interpreting will face more pressure by 2030. Live, high-stakes interpreting in courts and hospitals is more protected, but not permanently immune.

The scenario where AI genuinely threatens the whole profession requires two things that haven't happened yet: reliable domain-specific accuracy in legal and medical terminology across many language pairs, and some form of accountable AI that clients and institutions can trust with liability-sensitive documents. Neither is close. Legal and medical clients aren't going to remove the human layer until liability frameworks change, and that's a regulatory and institutional problem, not just a technical one. If you're building depth in a specialised field and a specific language pair, the realistic timeline for serious threat to your work is beyond 10 years. If you're doing commodity document translation, the pressure is already here.

The clearest move is to double down on the tasks with 0% AI penetration and build credentials in a domain where errors have consequences. Legal, medical, and technical translation command higher rates and are the last areas clients will hand to unreviewed AI. If you don't have domain specialisation yet, picking one and building toward a certification is a better use of time than trying to compete on volume with machine translation output.

The supervisory and training tasks in this role are also worth taking seriously as a career direction. As organisations adopt translation technology, someone needs to manage translation workflows, evaluate output quality, and train junior staff. That person needs to understand both the linguistic and the technical side. Positioning yourself as a translation project manager or quality assurance specialist adds a layer of value that doesn't compete with AI at all. Courses through the American Translators Association or the Chartered Institute of Linguists can build formal credentials in this direction.

For interpreters, the live presence tasks are your strongest ground. In-person legal interpreting, medical interpreting with certification from the National Board of Certification for Medical Interpreters, and educational interpreting all have institutional protections that document translation doesn't. Remote video interpreting is also growing, and getting certified for platforms that serve healthcare or court systems is a concrete way to access that demand. The translators facing the most career risk right now are generalists doing commodity document work in common language pairs without any domain specialisation. If that's you, the shift toward specialisation isn't optional anymore.