How useful are AI language tutors and will they steal teachers’ jobs?
Over the past year, there has been a flurry of new apps that offer foreign language conversation practice using AI chat. I’ve come across 15 so far (Univerbal, Loora, Membot, Tutor Lily, Verbius, LangBuddy, Makes You Fluent, Praktika, Talkpal, Jumpspeak, Speak, PrettyPolly, Gliglish, Lingolette, Lantalk).
On the surface, these apps offer something very powerful — the opportunity to practise a foreign language in a realistic way with a ‘human-like’ conversation partner who can talk about any topic at any level. And correct your mistakes.
Some of the apps are better than others (I recently did a detailed comparison of 4 of them). Even the better apps still have some weaknesses. Some of these weaknesses are more than minor annoyances and can misinform language learners about what is right and wrong in a language or how to use the language most naturally.
There are also strengths, which I discuss at the end of this article.
Weaknesses
First, here are the weaknesses in order of how negative and impactful I believe they are. I discuss each one in more detail below.
1. They occasionally give incorrect advice about grammar or vocabulary
All the apps I tested had moments where they missed mistakes or said things were incorrect when they weren’t. I uncovered this by using them to practise English (my native language) and by having a Spanish teacher check my chats in my non-native Spanish.
I roughly estimate the apps’ advice was correct 95% of the time, which sounds fairly reliable, but, as a language learner, you don’t know which 5% is wrong, so that casts doubt over everything the app tells you.
If other types of apps are unreliable or incorrect, it’s usually obvious to the user. If you book a taxi through an app and the taxi doesn’t turn up, it’s pretty clear you can’t rely on the app. But if an app confidently tells you something is correct in a language you don’t speak, you have no way of knowing that.
Many language learners still feel that they can only be truly confident about feedback if it’s coming from a human. I run an app called Go Correct, where a human teacher will correct learners’ English and users say they still prefer to have a human correct them.
2. They inadvertently encourage you to speak in a slow and unnatural way
The speech recognition in an AI tutor has problems capturing everything you say, even with correct pronunciation. I found myself speaking really slowly and deliberately so that it could pick everything up, which is unnatural and not a good behaviour to reinforce in language learners.
In Spanish, I’ve made the effort over the years to practise speaking quickly and smoothly in a way that (hopefully) makes me sound more similar to a native speaker. When speaking with these apps, I found myself reverting to a slower, clearer way of speaking in order to have the speech recognition pick up everything correctly. That felt like a step in the wrong direction for me.
3. The voices they use are not like the voices you’ll hear in the real world
The computer-generated voices used in the apps are very different from voices you will encounter in the real world. Some apps have more realistic voices than others but even the most realistic computer-generated voices can’t match the challenging nuances of real humans’ voices. For lower levels, it’s probably helpful to hear a ‘sanitised’ version of a voice, but to really prepare for the real world, a more advanced learner may find these computer voices too easy.
4. The conversations lack the sparkle and tension that you get in a conversation with a human
The ‘people’ you’re chatting to usually lack personality and edge. They’re just so nice and enthusiastic. It reminds me of being on a boring date with someone who’s lovely but who I don’t have any chemistry with. However, some apps (e.g. Univerbal) do let you request a certain type of personality in your bespoke conversations. But even in Univerbal, the AI character mainly paraphrases and reinforces the last thing you said. It doesn’t bring new ideas to the conversation or challenge you. It feels like you’re the one doing the work to keep the conversation going (again, like a bad date).
Considering all this, I think as of now in 2024, a conversation with a real human or teacher will still be the ‘gold standard’ for language practice.
Strengths
I like these apps because it’s rare that a language-learning app is sufficiently challenging for more advanced learners. They’re a demographic that is really under-served. However, these AI chat practice apps can offer sufficiently challenging conversations with a high level of vocabulary, and I have enjoyed some of the challenging conversations I’ve had with Univerbal in Spanish.
The convenience and cost saving is also a major bonus. They provide quick conversations, on demand, at short notice. They are cheaper than paying a human to talk to you but, with a roughly £10/month ($13) subscription fee, you have to use them a lot to make the cost worth it.
Who am I and why am I doing this research?
I have a BA in Linguistics & Phonetics and an MSc in Digital Product Development. I have been working in language learning technology since 2008 and, before that, wrote my master's thesis on it.
I run a product called Go Correct for language learners who want to improve through daily practice, with human feedback. I like to stay aware of the latest trends and think about how I might incorporate them into my own product. You can connect with me on LinkedIn.
