Three years after its arrival in language learning, conversational AI has moved past the novelty effect. It offers continuous oral practice and error diagnosis, but its value lies primarily in how it is articulated with human instruction.
A wave of enthusiasm, then a real market
For just over three years, conversational AI has occupied a place in the language learning landscape. The first months were marked by massive enthusiasm: almost all app publishers added an AI conversation feature, presented as the revolution capable of solving the most stubborn problem in learning: oral practice.
Three years later, the initial enthusiasm has subsided, as is often the case with a technological novelty. The question is no longer whether the tool impresses on first try, but what it offers once the allure of novelty has faded, and how it integrates into a broader learning system.
Two contributions, and one pitfall
Beyond the demonstration effect, two contributions stand out. The first is the availability of oral practice. Unlike a human trainer, a conversational AI is accessible at any time, without perceived judgment, which removes part of the psychological barrier that many adult learners associate with speaking in a foreign language.
The second contribution, less visible, is the fine-grained analysis of errors. A conversation generates a quantity of data that the human ear cannot process with the same systematicity: recurring grammar errors, structures avoided by the learner, hesitations on certain types of sentences. Well exploited, this data allows for the identification of precise areas for improvement, rather than relying on a general impression.
The downside lies in the phenomenon of the novelty effect, well-documented in educational technologies. According to several studies, a new tool, especially if it relies on game-like mechanics, causes a spike in engagement that then declines sharply, often within a few weeks. This reflex is partly explained by neurobiology: according to the work of Bunzeck and Düzel, the brain reacts to novelty for its own sake, at the level of the midbrain. However, this reaction is measured in seconds or minutes, whereas the use of a tool plays out over weeks.
The interface also matters. A study involving over 20,000 students, which compared different interfaces for the same pedagogical engine using A/B tests, showed that targeted work on presentation could improve engagement by over 25%, without changing the content. This work focuses mainly on intelligent tutoring systems in mathematics and science. Its transposition to languages remains an extrapolation, but it reminds us that a pedagogically sound tool can fail if its interface discourages regular use.
A marked need in France
The issue of oral practice takes on particular significance in the French market. School language teaching there remains historically focused on grammar and writing. It produces learners capable of reading and analyzing a language, but often blocked as soon as it comes to speaking it spontaneously. This blockage, regularly highlighted in international level comparisons, partly explains the resonance of a feature focused on oral skills.
The debate now centers on the tool’s place in a complete learning path. Used alone, conversational AI offers repeated practice and error diagnosis, but it does not replace the contextualized correction of a trainer, nor adaptation to a specific professional objective, nor long-term follow-up. The emerging approach relies on a form of flipped pedagogy: the AI handles repetitive training and the identification of weak points, which frees up trainer time to address these points rather than a generic program.
Three years after its arrival, conversational AI is therefore judged less on its ability to impress in a demonstration than on its use after several weeks, the quality of the diagnosis it produces, and how this diagnosis is exploited in the subsequent learning journey.
