I am the Director of an online Master’s program in Translation at a time when neural networks, AI technologies, and Large Language Models (LLMs) are transforming the industry. These rapid changes have required us to adapt the way we teach, since the Master’s program I am honored to lead started its activities nearly 20 years ago, in a very different context. Today, however, genuine innovations are often surrounded by so much hype that it can be difficult to distinguish reality from marketing. How should we steer our teaching?
Hype vs. History
As books such as AI Snake Oil (by Arvind Narayanan and Sayash Kapoor) or The AI Con (by Emily Bender and Alex Hanna) remind us, much of what is claimed about current AI systems is exaggerated, if not entirely misleading. Even the label “Artificial Intelligence” can be deceptive.
It is possible that one day technology will produce an intelligence comparable to—or even surpassing—human intelligence. But, as researchers like Gary Marcus have often pointed out, there is no clear path from today’s technologies to that hypothetical future. The past few years have revealed both the impressive capabilities and the very real limits of LLMs. Despite the billions of dollars invested in reducing hallucinations (or confabulations), this problem has not been solved—merely reduced. As a result, these tools remain unreliable for many real-world applications.
In short, these systems are powerful, but if we must choose between two extremes, they are still better described as “autocomplete on steroids” than as true intelligences. And there is little evidence that this will change rapidly.
A degree of skepticism about the pace of progress is also justified by history. Technologies such as chatbots and machine translation showed impressive possibilities at the start but then stagnated for decades, relying on simple tricks whose possibilities were quickly exploited to the full. For example, the capacities of the very first modern chatbot, ELIZA—developed in 1966 by Joseph Weizenbaum—remained essentially unmatched until the recent boom in neural text generators. The same may happen with LLMs: incremental improvement will surely continue, but transformative breakthroughs are far from guaranteed.
Where We Stand
Today, machine translation systems can already produce surprisingly high-quality output. However, experts know that unsupervised automatic translation is fully reliable only in specific contexts. Even the most advanced LLMs regularly produce grammatical errors, awkward phrasing, or stylistically poor passages. Human oversight remains indispensable to achieve professional-quality results.
This is not a defensive claim by professionals worried about their jobs, but rather the conclusion of numerous state-of-the-art evaluations. Since 2013, I have worked with other researchers in systematically assessing the performance of linguistic technologies for Italian—from speech-to-text to, more recently, neural machine translation and LLMs. The most recent results consistently show that even for basic grammatical correctness, the best systems still commit non-trivial errors—often about one problematic sentence per page. For specialized tasks, error rates rise even further, with no clear path to eliminating these weaknesses.
For now—and for the foreseeable future—the world of professional translation is therefore bound to a reality of “almost -but-not-quite-accurate” machine translation. What does this mean for professionals?
Translators and Editors
Ethan Mollick has popularized the metaphor of the “centaur”: a human using AI to complete difficult tasks. Cory Doctorow has, however, remarked the limits of that idea, pointing out how demanding this role can be, since humans must remain vigilant to catch mistakes that AI presents with undue confidence. In translation, though, this challenge is not new. Human editors have always revised the work of translators, often with remarkable expertise and skill. The use for this kind of competences is already growing and it has many chances to continue its growth.
In today’s context, therefore, it seems probable that many translators will feel the need to expand their skill sets to include the revision of AI-generated translations. This of course does not mean abandoning traditional expertise; rather, it means extending it. Increasingly, good translators will probably need to be also good editors—or, more specifically, skilled post-editors capable of refining AI-assisted translations that fall short of professional standards.
Training for the Future
Educational programs must reflect this reality, and many already do. The online Master in Specialized Translation from English to Italian (one year), jointly offered by the Universities of Pisa and Genoa, is designed precisely with this in mind:
👉 Master in Specialized Translation
https://master.italicon.it/it/traduzione-specialistica
Within its curriculum, AI-related instruction is integrated with traditional translation training, ensuring that participants are well prepared for the realities of the translation market in the near future.
For professional translators interested in specializing in post-editing, the program also offers a dedicated 3-months course:
👉 Post-Editing Course
https://master.italicon.it/it/traduzione-specialistica/il-post-editing-la-traduzione-specialistica
Most importantly, the faculty and instructors of both the Master’s program and the post-editing course are convinced that the abilities of human translators remain central—and, indeed, more important than ever in an AI-augmented world. This conviction, rooted in both market knowledge and technological awareness, guides our teaching and defines our work.
📅 Enrollment for the post-editing course is open until September 1, 2025. Enrollment for the 18th edition of the Master’s program is open until November 22, 2025, with classes starting in January. The programs are taught in Italian.
