If AI handles the first draft, how will the next generation develop the judgment needed to check it?
For more than two decades, our founder has worked in translation and localization, first as a translator and later managing multilingual teams. That experience raises a question that matters far beyond the translation industry: if AI takes over the work that used to help junior professionals learn, who will develop the expertise needed to supervise AI tomorrow?
Before AI could produce fluent text in seconds, a new translator learned by doing the difficult work. They read unfamiliar source material, researched terminology, decided how to express the meaning, made mistakes and received feedback. Over time, the decisions became more informed. They learned that a grammatical sentence isn't necessarily a correct translation—and that a correct translation isn't necessarily a good one.
Today, that learning process is changing.
Can someone learn translation by correcting AI?
Machine translation post-editing is not a simple task. Fluent output can hide a wrong meaning, a missing condition or inappropriate terminology. Experienced linguists know how to question text that appears perfectly natural.
The concern is what happens when a newcomer works almost exclusively on generated drafts. Someone who has translated independently for years may immediately recognise a subtle change in meaning because they have made—and corrected—similar decisions many times. But how does a junior linguist acquire that instinct if the first decision is always made by a model?
This is not an argument against post-editing. Reviewing multiple translations, comparing alternatives and analysing AI errors can be excellent training. But reviewing other people's—or machines'—work should not completely replace learning to create a good translation independently.
Language changes. Does that mean quality is declining?
The discussion becomes more complicated when we consider how language itself changes. Every generation has its own communication habits. Short messages, unconventional punctuation and new expressions are not automatically signs of poor writing. English, Hungarian and every other living language evolve.
We no longer write like Shakespeare, but that does not mean that language has simply deteriorated. Some expressions considered awkward today may become ordinary in the future.
There is, however, a difference between the natural evolution of language and losing the ability to read, interpret or express complex information accurately. This distinction matters when the text concerns vehicle instructions, medical devices, software controls or other technical information. The style may change; the meaning still has to be right.
Reading skills are part of the challenge
The OECD’s PISA 2025 findings add context to the discussion. Across 35 OECD countries with comparable data, average reading performance among 15-year-olds fell by 25 points between 2018 and 2025. Across 74 countries and economies with comparable results, 56 recorded statistically significant declines.
These figures do not show that AI caused the decline, nor do they predict the competence of tomorrow's translators. They do remind us that reading and understanding complex material cannot be taken for granted. For professional localization, careful reading, interpretation and sound decisions remain essential.
Source: OECD, PISA 2025 Results, reading trends.
What happens when AI-generated writing becomes the reference?
There is another question worth exploring. AI systems are trained on large amounts of human language, while people are now consuming increasing amounts of AI-generated text. Could this eventually influence what readers regard as natural or professional writing?
That influence is not necessarily negative. AI can help people express themselves clearly, make information accessible and improve consistency. But anyone regularly reviewing generated content will recognise patterns: familiar phrasing, predictable structures and text that is acceptable yet sometimes strangely generic.
The long-term effects are uncertain. Still, it is worth asking whether people who mainly read and edit machine-generated material will remain as sensitive to the difference between text that merely sounds correct and text that communicates with genuine precision and purpose.
Developing the next generation of language specialists
The response should not be to stop using AI. It should be to design training and production workflows that allow professionals to develop independent judgment alongside new technical skills.
Some practical approaches are straightforward:
- Let juniors translate independently. Give them source passages to translate before showing an AI draft, so they can form and defend their own interpretation.
- Compare decisions, not just sentences. Review why alternatives differ in meaning, terminology, register or audience suitability.
- Provide structured feedback. Experienced reviewers should explain significant errors and the reasoning behind corrections, rather than only delivering a changed file.
- Teach AI evaluation explicitly. Linguists need to understand how apparently fluent output can omit a condition, invent a detail or resolve ambiguity incorrectly.
- Preserve mentoring opportunities. Teams should make space for junior specialists to work with experienced professionals on difficult content, instead of assigning them only repetitive corrections.
These are starting points, not a complete training system. But they acknowledge that expertise takes time to build, and that efficiency today shouldn't eliminate the learning opportunities needed tomorrow.
Keeping expertise in the loop
In an earlier article, Are We Training AI with the Wrong Translators?, Miklós Tóth questioned whether the industry involves enough experienced linguists in training and evaluating AI systems.
The other side of that problem may be just as important. If new translators rarely have the chance to develop deep linguistic expertise, who will evaluate the next generation of models?
At MITO Solutions, we see human expertise, careful linguistic quality assurance and responsible AI workflows as complementary, not competing, priorities. Technology can make production faster. It does not eliminate the need to develop professionals who know when the output should be trusted—and when it shouldn't.
The greatest risk may not be that AI replaces translators, but that we gradually lose the experiences that make good translators possible. Rebuilding that expertise later could be far harder than preserving it now.
For discussions about AI evaluation, linguistic quality and localization workflows, visit MITO Solutions or contact contact@mitosolutions.com.
Original article by Miklós Tóth, published on LinkedIn on 28 October 2026. Adapted for MITO Solutions.