Why LLM Chatbots Won’t Be Replacing Language Teachers Any Time Soon

When was the last time you used a chatbot? It’s probably more recent than you think. A 2022 survey from Tidio found that 88% of participants had a minimum of one conversation with a chatbot in the last 12 months, and throughout the world, there are roughly 1.5 billion people using chatbots.

However, the time of the preprogrammed chatbots—capable of providing five answers— is over: Generative AI and large language models (LLMs) have arrived. And while LLM chatbots are revolutionizing customer service, providing multilingual support, and analyzing data-driven insights—saving businesses millions of dollars—its uses have now diversified and expanded to education, in particular, language learning.

Theoretically, LLM chatbots could be the most innovative tool language learning has ever experienced, as they’re interactive, accessible, and affordable. Bill Gates has even claimed they could teach children to read in 18 months rather than years. After all, using an LLM is a more interactive experience as they’re not deterministic, and their answers cannot be controlled. So it’s their capability of absorbing, processing, and creating new data—specializing in the human language—that makes them especially applicable to education, particularly language learning.

But will their erratic hallucinations and inaccuracies, caused by the data sets they’ve been trained on and their inability to understand the prompt information, make them unsuitable for teaching?

Let’s dive into what makes an LLM chabot tick, its educational use cases, and why it won’t be vying for language teacher jobs yet.

The science behind the chatbots

The way that chatbots work varies depending on the chatbot; however, in this article, LLMs—a subset of generative AI, like Google’s Bard and OpenAI’s ChatGPT—will be the main focus.

However, it’s right to acknowledge where chatbots in language learning have come from, as other famous language apps, such as Duolingo, also use AI chatbots, particularly machine learning (ML) and natural language processing (NLP) models. Although, the limitations here are that this type of framework is back-coded, meaning it’s programmed to check and only validate correct answers; it’s deterministic. While this type of question-answer language testing is beneficial when acquiring a new language, LLMs can delve much deeper.

LLMs specialize in NLP and can generate and mimic human language, making them ideal for language learning. Real-world uses can include machine translation, text summarization, question-answering systems, chatbots, and content generation—they excel at understanding the context and semantics of text data. And due to their expansive uses, LLMs are characterized by their enormous size and computational requirements as they contain billions of parameters.

On the other hand, traditional ML generally focuses on understanding data, making predictions, solving specific tasks from pattern recognition, and labeled and unlabelled data. And within language learning, the applications of ML can vary from speech recognition to fraud detection (finding students who like to cheat) and predictive analytics. It’s best used for tasks where data is structured and represented numerically. And unlike LLMs, traditional ML models are smaller and computationally less demanding, depending on the complexity of the task and the chosen algorithm.

LLMs: The new super tool for language classes?

In a typical language lesson, a teacher could have anywhere from 8 to 20 students with varying levels (and interests) in the language. So, in a lesson where a teacher may need to answer questions from 50% of the class, this spreads the tutor’s time and energy a little thin.

Therefore, LLMs are ideal for teaching the basics, like grasping simple grammar rules and practicing vocabulary. With LLM chatbots, users can use school-sanctioned laptops and ask for grammar rules to be explained as if they were five, 10, or 15 years old, catering to every level of student. Or they can ask for a grammar rule to be explained using football terminology, making it personal for individual students and hence, something they’re more likely to pay attention to. This increased engagement means users are likely to return and practice again.

Additionally, LLM chatbots are excellent alternatives for conversation practice as they can mimic a native speaker. Plus, their interactive and entertaining nature is another key reason which keeps people interested. And all teachers know that engagement can be one of the biggest hurdles in language learning.

One of the significant benefits of LLM chatbots is their widespread accessibility. Anyone can create an account and access applications such as Chat GPT or Google’s Bard, as they both have competent free versions. So, as long as you have a stable internet connection, any student from around the world can leverage the practical capabilities of an LLM chatbot. However, nothing in life, and AI, is perfect.

Unraveling the pitfalls of LLMs in language learning

While LLMs are valuable tools for language learning, one of the foundational pillars of grasping a new language is fluency, and this skill cannot be achieved through chatting online with an LLM. As chatbots are mainly based on written communication, they could hinder pronunciation and speaking skills if they’re the only tool used in language learning. Only a human teacher can help students perfect their abilities, provide demonstrations, and, more importantly, corrections.

One of the most significant issues surrounding the development of generative AI and LLMs is the AI hallucination state. This occurs when chatbots don’t recognize or understand the question or information they’ve been given. Therefore, they go ‘AWOL’ and start fabricating responses meaning they can provide inaccurate information and may reinforce errors or misconceptions.

LLMs absorb data from the internet and then generate new data based on these sources—and since anyone can post on the internet—incorrect and prejudicial information can infiltrate LLMs which struggle to differentiate between fact and fiction.

These types of hallucinations can undermine trust, and the BBC reported that ChatGPT has been known to ask students for sexy pictures. So, when used in an educational setting with children, the non-deterministic aspect of LLMs creates a safeguarding issue. Human teachers must go through rigorous checks to work with children: If similar moderation isn’t applied to LLM chatbots, this could place children in a potentially harmful situation.

Moreover, learning a language doesn’t only consist of mastering conjugations, tenses, and grammar rules. Language learning encompasses much more, such as cultural understanding, idiomatic expressions, and contextual usage. Although an LLM chatbot can provide some idioms or slang on request, it cannot try to rival a knowledgeable teacher, who, for example, can provide the essential nuances between Spanish-speaking countries. Further to this, generally speaking, students must take the initiative when using a chatbot. Users need to know what prompts to ask, and beginners, for example, are unlikely to possess this knowledge.

Additionally, the potential of LLM chatbots to aid students in cheating during exams—such as providing answers or speech recognition capabilities—has prompted some Seattle and California school districts to block platforms like ChatGPT.

There’s no denying that generative AI, particularly LLM chatbots, is an incredible tool that is revolutionizing many aspects of business and is on a trajectory to be implemented in schools soon. However, while the model’s accessibility and affordability can overcome two major hurdles when learning a new language, its unpredictable inaccuracies and hallucinations mean that the current technology cannot rival and replace a good language teacher.

Generative AI has a significant obstacle to overcome regarding its hallucination issue, but; in the right circumstances, LLMs can be helpful tools for supervised homework and grammar practice. Overall, teachers should be safe in the knowledge that, at present, generative AI won’t be taking over the education system.