Richard Good
The thought of artificial intelligence intruding on our lives is unsettling.
And yet. Let’s suspend any reservations we have and ask the question whether AI can help us learn a language.
Trial and error is a big part of the learning process. We express a thought, it comes out wrong, our parents or teachers or peers correct us. We progress.
So what if a chatbot were to take the role of the parent, the teacher or the peer? We could study whenever we want. We wouldn’t bother our teacher with stupid mistakes. We could take our time weighing up what we want to say without getting waylaid by social pressures.
Let’s give it a go.
The test
I chose ChatGPT andBard. I’d tell them about a film I’d just watched and ask them to correct any mistakes in my French and make any improvement suggestions they had. And to control the experiment, I’d check the results with both Deepl and a human French teacher.
The experience
On a failli être amies starring Karin Viard and Emmanuelle Devos is not a must-see-before-you-die film. But the pitch – a vocational trainer falls in love with the husband of one of her trainees – was appropriately straightforward for the experiment.
Neither ChatGPT nor Bard respond directly to the spoken word. You need to use the voice-to-text functionality of your device to dictate your speech and then press enter to send your message to the bots.
This in itself is no bad thing: it gives you the opportunity to correct any transcription errors before submitting your work. But it turns out to be a much more comfortable experience on an Apple device than an Android. iOS continues listening until you hit a big blue stop button – so if you pause for thought, the recording keeps going. Android on the other hand annoyingly stops listening if you pause for more than a couple of seconds – meaning you have to reactivate the voice-to-text functionality to resume talking.
Once I overcame the sense of ridicule that comes from talking to a computer, I found the experience relatively enjoyable. Both ChatGPT and Bard immediately understood what I wanted from them and played the game – offering improvement suggestions to my utterances.
They also encouraged me on the way ( « Je suis tout ouïe ! » « Continuez, j’écoute attentivement ») and offered were very polite when slapping me down (« Votre phrase est presque parfaite ! Voici une version légèrement modifiée pour la fluidité! »).
ChatGPT vs. Bard – Conversational style
The user experience with the two chatbots is very similar. But there were noticeable differences in style.
Bard is noticeably more verbose – asking for feedback on its own contributions and surrounding its corrections with lengthy explanations.
This was excessive for my tastes and I preferred the more compact responses from ChatGPT.
At one point in the conversation, I halted my narrative and asked the bots how they thought Emmanuelle’s character might react to the discovery that her vocational trainer was trying to seduce her husband.
ChatGPT gave a very circumspect reply, saying that as a machine it was unable to predict human emotions, though it did outline a number of possible reactions. Bard showed no such inhibitions, replying:
Je pense que Emmanuelle a été surprise et déçue par la révélation de Karin. Elle se sentait probablement trahie par sa conseillère, qui était censée l’aider à retrouver un emploi, mais qui a finalement cherché à manipuler la situation pour avoir une relation avec son mari.
I was stopped in my tracks by Bard’s ability to assimilate different elements of a conversation and come up with such a detailed response.
ChatGPT vs. Bard – The Corrections
Both chatbots were able to pick up on grammatical errors and correct them. They also offered some helpful vocabulary improvement suggestions: proposing elle assiste à un stage as a more elegant than elle suit un stage, for example, or car instead of parce que and so on. In a text of 200 words, I counted ten improvement points to take away. Not a bad result.
However, even the bot’s corrected versions proved clunky when placed side by side French Classes founder Françoise Le Roux’s model text. Chatbots adjust your output – but they will not provide the radical changes of syntax that can be needed to develop a truly fluent turn of phrase. And in terms of vocabulary improvements, it was Deepl that proposed the closest text to Françoise’s model.
Conclusion
Activities that you can do on your own are gold dust for language learners, and I was pleasantly surprised by this experience.
Chatbots are not about to replace the value or pleasure you’ll get from conversing with human beings. But if you use them for training, as a complement to authentic experience, they’re well worth considering as an option.
The conversational style of exchanges mirrors the demands of real life. And because you get to choose the topics for discussion, you work on language that will be useful for you. The caveat is that the corrections only offer incremental improvements to your own imperfect speech, not models to adopt.
