By Hamza Achahoud, PhD candidate in Applied Linguistics at Ibn Zohr University, Morocco

DOI: https://www.doi.org/10.69732/ZMPE3811

Only a few years after the announcement of the first commercialized generative artificial intelligence (GenAI) chatbots, for some of us, they have already found a place in our teaching and learning toolbox. As a doctoral student, I find myself brainstorming and discussing my thoughts and choices on a GenAI chat more often than I do with my peers and advisers. Before I knew it, whenever I got a question or decision that I needed feedback for, whether in my research or my lesson preparation when I am teaching, it always leads to a GenAI platform.

We keep coming back for more once we realize AI has answers; beyond the readily available information on practically any subject, it can also do the thinking and make decisions, saving our minds the cognitive effort. Or so it seems.

Our questions for AI are increasingly bolder and more ambitious. It is reminiscent of the literary work by Isaac Asimov “The Last Question,” in which humans, across millions of years, kept urging the supercomputer, Multivac or AC (an automatic computer) as it was referred to in the story, to answer their ultimate question on how to reverse entropy; that is, how to avoid the universe’s inevitable demise. However, I believe that before we set out to ask questions of that magnitude, we ought to prioritize some more fundamental questions first – what is GenAI currently capable of? and what are its limitations?

Working in a foreign language context, whether in the classroom teaching or at my desk drafting chapters for my doctoral dissertation, I am constantly faced with a decision concerning GenAI: to use or not to use? GenAI seems like it has the potential to do wonders; it can summarize articles that I am reading in a split second, it can provide feedback on my students’ homework (though, of course, we have to be cognizant of privacy laws), and it seems that it can even score exams which would have taken me hours to do manually. Yet, I know that delegating such responsibility ultimately raises the question: to what extent can I truly rely on GenAI to do my job? This question shows up everywhere in my teaching now, and in the following part I am going to work through one example on writing assessment to demonstrate why a fundamental understanding of GenAI mechanisms matters.

The First Question

By “first question”, I mean the first thing we need to think about once we settle on a GenAI tool, and before implementing it in our own language teaching. I believe using technology efficiently depends on how much we understand how it functions. In turn, this understanding also factors into our decision on whether we are going to use it or not. Therefore, I think the first question that matters is: how does a tool actually arrive at its answers? Addressing this will help reveal the tool’s strengths and limitations.

 Researchers have begun to investigate AI in assessment, including in language learning contexts (e.g., review of ChatGPT and L2 assessment, GenAI for essay grading). Yet, before we develop full-fledged research designs, it would not hurt to experiment a little with GenAI, as we usually do on our own devices, only in a more deliberate way; a simple writing assessment test.

Assessing writing is difficult and somewhat subjective, but there are certain indicators that could contribute to our evaluation of a text. For instance, Lexical Diversity (LD) is a metric that can be utilized as a measurement of writing proficiency, it is calculated as the ratio of the number of unique words in a text to its total word count. The higher the LD, the more vocabulary a text uses, which can imply a good command of language through a demonstrated ability to use a diverse range of vocabulary. It is an established metric, but small enough that I can check it by hand; which is exactly why I picked it. If GenAI fails to do basic arithmetic on a paragraph, then surely it cannot be trusted for a holistic score of a whole essay.

For the test, I prompted ChatGPT to measure the LD percentage for a paragraph of my own writing. It missed the count at first and gave a wrong answer. However, when I tried again in a new chat it yielded the correct answer. And so did Claude and Gemini; they all could measure this simple metric without issues.

Nonetheless, the fact that it got the answer wrong once raises a critical concern on whether I can trust it. This is where I came to realize that I need to understand why AI answers were sometimes inconsistent. Looking at the outcome does not tell me enough information.

I recall AI struggling with this exact task a year ago. This small exercise shows how far, or in a more nuanced sense, how fast the AI tech is advancing. It is hard to keep up with this exponential development. On such account, my thinking on what makes up a good first question has shifted; rather than asking how effective AI is in a specific task by looking at outcomes, it may be better to consider how AI in fact makes its decisions and arrives at conclusions.

This deductive way of thinking about the issue ensures a greater sense of reliability; understanding the foundational rules of AI will help us know what to expect of it. If we can assert that AI is built in a way that allows word counting in a text, then as a consequence we know that all applications including a count of words are likely to be reliable. Along the same lines, if we aim to use AI for writing assessment, we may look at whether its inner or built-in functioning allows it. And the most direct way to do this, I believe, is checking the tool’s own official documentation; there you can find all the explanations as well as the new updates (more on how I learn about AI shortly).

Because when we talk of AI in terms of safety, trust, and ethics, we need to acknowledge that the conversation includes questions about how it was built and programmed, and how it treats this data we feed into it; what happens to the students’ personal work if I upload it to an AI platform? That is a question of student consent and privacy, as well as about how AI works.

Our first-hand experience using AI, the growing body of research, and everything in between, mostly only focus on the AI outcome instead of the process. For instance, some research (e.g., check this synthesis on ChatGPT in L2 writing feedback and assessment) suggests a positive influence for AI as a feedback mechanism on writing, though how this was concluded is usually by comparing to human feedback, or by exploring the impact on student performance after introducing it. This reasoning is not completely flawless; positive outcomes in this kind of experiment do not answer in fact whether AI can assess writing, it just delivers feedback that has a positive effect, or a score that correlates to that offered by human raters. The logical oversight in this example means that our trust in AI is not necessarily strengthened even with encouraging results, and that is because we inherently take on the role of an overseer who constantly needs to look over AI’s shoulder to ensure it is working correctly.

What I suggest is to learn the foundational rules upon which AI is built. From our standpoint as teachers and education practitioners, we are usually focused on the applications of AI at the cost of any attention to the technical building blocks of the technology itself. It is time that we engage in the AI conversation right next to engineers. The article “Should I (You?) Have a Seat at the AI Table?” stresses that teachers too should take part in AI discussions and create their own policies. The thoughts that I am sharing in this piece are in effect a response to this call, from my position as practitioner in an EFL environment. However, I also believe that what precedes the policy is a profound understanding of the technology itself.

For this reason, the very first question we need to ask needs to center around AI itself, and how does it function?

Sufficient Data for a Potential Correct Answer

Generative AI, running on large language models (LLMs), is constantly in progress with new features implemented frequently to enhance the user experience. Said features range from technical additions that are simple to understand for the average customer, like the “search the web” function that triggers a wide search of the internet to confirm an answer, making use of the Retrieval-Augmented Generation (RAG) method. Or the “extended thinking” option that reduces hallucinations by basically producing a plan, breaking down the user prompt to steps and testing multiple assumptions before proceeding with an outcome.

Other features sit deep within the underlying architecture of the conversational chatbots we use and love. Such is the Reinforcement Learning from Human Feedback (RLHF) technique; used to align AI models with our values and preferences. After a model is pre-trained on a massive amount of data, human reviewers rate different responses to the same prompt, effectively training it to produce answers in ways that humans prefer; for example, a polite response is favored over the same response told in an aggressive tone (described in InstructGPT Paper).

Another example that significantly improved the efficacy of chatbots is developing the ability to utilize tools; now LLMs can independently generate code to calculate or look up some specific information . It is the ability of a model to recognize that a task can be done more efficiently by writing and running a piece of code, or calling an external function (a technique known as ToolFormer). And that is probably what happened previously in our little exercise, according to the available documentation on this feature that basically relies on writing code to conduct precise data analysis. The chatbot I prompted to compute the LD of a paragraph realized it would need a separate tool to make the calculations, identified the suitable tool or function and then called it, and finally interpreted results in plain English. This is only my best explanation, as I cannot see the model’s internal process directly.

GenAI at its very foundation remains a predicting machine; the current consensus views it as unable to understand natural languages the way humans do, but since the data it was trained on is so enormous, it learned how to imitate and generate a response that sounds human; a guess based on statistical prediction rather than actual comprehension. Many researchers argue that these systems understand patterns and speak in probabilities and mathematical vectors, rather than concepts or emotions. AI today is not yet at the same level as AC in Asimov’s short story; the ever-evolving supercomputer that can answer humanity’s biggest questions; however, it has learned sufficiently to provide answers that are potentially correct.

In the language classroom, the question persists: To what extent can we rely on this statistical imitation for writing assessment or other language teaching tasks? That remains up for debate, and depends on the way this specific task is carried out within the AI system. In other words, teachers need to question how the AI currently handles processing of language.

Today, we need more than ever to have a front row seat at the AI table. It is urgent and critical that teachers and educators in general cross the bridge into the engineering side of GenAI. Raising awareness about the need to understand the magic of GenAI could, in turn, provide some vital assistance to address our pedagogical worries. When we quit thinking of AI only as a tool, and start taking interest in learning how it functions, only then can we be well-equipped to make better choices. Next, we could make informed decisions regarding our own AI policy, and eventually, we will be able to contribute to the conversation and have our thoughts and experiences influence future AI development, especially its technical aspect, ensuring that our perspective and knowledge are considered when building the tools.

So, here is what I have started doing, and perhaps you could try too. The next time you find yourself hesitant to use AI in your teaching, remember to ask yourself two simple questions: What is it that I expect the AI to do precisely about this task? You want to assess a student’s writing; are you expecting a holistic score for the outcome, or a specific metric you need to compute? Then ask: Can the underlying design of current GenAI, or the exact LLM-based tool you are using, actually support what you are asking of it? If I need AI to enhance my work, to provide feedback that is often time consuming, is GenAI built for that? and Do the tool’s methods follow safety and privacy rules? Those are the questions you need to think over before you start using a chatbot.

It is obvious that the second question can be a bit complex, as it requires the extra effort to learn how GenAI works and keep up with any technical developments. However, thinking of these details would then push you to consider what GenAI can and cannot do for you. Practically, to address this, I always check the official documentation. It may sound boring, but it really is not if you are interested in learning about the tool you are using. The docs for ChatGPT – its capabilities page, or the all things ChatGPT collection,- Gemini, and Claude are all user-friendly and presented neatly for a regular user with no tech background.

For instance, once I learned that GenAI can run code that it writes specifically to respond to a prompt, I started searching for this online, and then I was directed to the code-execution page, where you can see that explicitly using the keywords “run”, “compute”, or “execute” triggers this feature. This explains the inconsistency that I had noticed in GenAI answers to my writing test. Using these keywords would ensure that the outcome is not based on prediction, but on a precise code that can count easily the number of unique words in a text.

I think that going through this process of learning about how AI functions will make you aware of AI’s limits, and more informed to make the most of it in your practice. Moreover, if you have other types of questions, such as privacy; you can look up how the tool deals with your data. Often, GenAI tools let you disable settings that allow it to use your data for training; however, it is not unreasonable to still not trust your students’ work to an AI platform. In which case you can learn about local LLMs, these are self-hosted models that can live only on your computer, safeguarding any information you have on it. When I tried this on my laptop, I discovered that my personal machine can only run a few of the available tiny models. This quest showed the solution to my issue with privacy, yet it revealed a need that I was not aware of: high-end computers that can run the models I need. This newfound knowledge is something I can take to my administration, and that is because I learned how this technology works. And if I learned how to manage a local LLM, I can even set up one computer that all teachers in my school can access to use its safe, local GenAI.

It is questions like these, about the nature of the technology, of the AI and how it functions, that can help teachers use AI to its fullest potential, and participate in the AI conversation based not only on our use, but of our understanding of it as well. Keep up with the updates on official blogs and websites, take introductory courses on GenAI, especially those tailored for teachers like those at Claude’s Anthropic Academy, or Google’s AI educator series among others. As we learn about this, we eventually answer those questions we have carried since the beginning.

Because in the end, it may not be our questions that need to get audacious, but our understanding.

AI disclosure: Minimal use of AI: Generative AI was used only for spelling and grammar suggestions.