The first program I developed locally, for my own use, and which could be considered a prototype of Octopus T2L, was based solely on dictionary lookups. We’re talking about the years 2018–2020, when artificial intelligence (AI) models weren’t yet very effective. In just a few years, the technical landscape of digital learning tools has completely changed thanks to the rise of AI.
A dictionary search involves linking an unknown word to a definition. The search returns a list of words that may match the query. For example, for a learner transitioning from French to English, “atonishment” yields “atonish” → étonnement, ébahissement, foudroiement. “Atonish” is also a verb. It’s up to the reader to do the mental work of determining the most appropriate meaning based on the context of the text. Cognitively, this effort can be quite significant, especially when faced with polysemy. This effort is not a bad thing in itself, since it leads to deeper encoding of the word in memory. However, it can also cause a form of fatigue in the reading process whenever the text’s difficulty level is far removed from the learner’s proficiency level.
Word learning never happens in isolation; it always takes place within a context. This is also a fundamental principle of learning on Octopus T2L: a word is never learned on its own, as in a vocabulary list. Whether in the spaced repetition, dictation, or pronunciation modules, words are consistently presented within sentences.
As someone who regularly tests AI models, I have to admit that they’ve made tremendous progress in less than a year. Across all my tests, I haven’t encountered any inaccuracies in the definitions returned by the AI. The big advantage is that it helps clear up ambiguity and often provides another example of how the word is used in a context similar to the one encountered while reading. It now does a very good job of identifying all semantic false friends, which is a huge advantage over dictionary lookups: “Be careful—this word cannot mean this in this context.”
The downside is that the feedback can sometimes be very wordy, interrupting the learner’s main reading activity. But the biggest problem is the tendency to provide a complete translation—which isn’t always very natural—of the sentence that provides the context. In a previous article, I explained that learners should avoid translating entire chunks of text so that the mental effort required to assimilate the language’s structures can take place.
In conclusion, I felt less fatigue when using AI than when using a dictionary. It’s possible that the encoding is less deep. But less fatigue also means more reading, and thus a broader linguistic range, as well as a lower likelihood of giving up on learning. In a society where we aren’t trained to focus on a single task for more than ten minutes, I believe the “Context AI” feature should be a priority for learners. That’s why I’ve enabled it by default on Octopus T2L.
01 octobre 2026