Educators beware: AI can harm learning

After having experienced the use of Large Language Models (LLMs) in schools and universities since the introduction of ChatGPT in 2022, there is a palpable sense of unease among educators. They suspect that excessive reliance on these tools for academic work by students could lead to ‘cognitive atrophy’ since LLMs like ChatGPT provide ready-made answers that short-circuit the whole process of learning by undermining the application and development of thinking and reasoning skills.

Two recent studies, one undertaken by the Massachusetts Institute of Technology (MIT) Media Lab and the other by the University of Pennsylvania (UPenn), aim at answering this question by focusing on the impact of artificial intelligence (AI) on learning outcomes.

The MIT study involved 54 participants divided into three equal groups of eighteen students. One group was tasked with writing essays using ChatGPT and was not allowed to use anything else. The second group used the search engine, Google, and was not permitted to use any LLM. The third had no digital assistance (neither LLMs nor search engines) for their essay writing and had to rely solely on their brains.

Each of the three groups completed three sessions under the same conditions, with electroencephalography (EEG) used to monitor and record their brain activity during the writing sessions.

In a subsequent fourth session the group tasks were flipped. Those using ChatGPT for the first three sessions were asked to write using only their brains whereas the group that had previously been writing without any digital tools was instructed to write their essays using ChatGPT.

The results were assessed by both human and AI judges. The outcome provides a sobering realization that AI may not be the silver bullet to educational uplift that its champions imagine it to be.

Thus, the MIT researchers found that while essays written with AI were more grammatically polished and properly structured, they lacked the spark of originality and creativity that was more evident in the essays written by the group that had no digital help.

Shockingly, upon being interviewed only a few minutes after task completion essay writers using LLMs could barely recall anything they had written. This might be termed, to borrow a term from Marxists, ‘cognitive alienation’. But what accounts for this blanking out of memory? The EEG data explains: their brains simply weren’t encoding the information effectively because they weren’t processing it to learn the material. ChatGPT users showed weaker activity in parts of their brains linked to attention and critical thinking.

In contrast, those reliant on their own brains showed ownership of their work. They had built the mental scaffolding required to write an essay that involves taking on a much bigger cognitive load than simply resorting to copy-paste procedures that most participants in the ChatGPT batch were wont to engage in.

(The test group restricted to using Google showed a moderate degree of brain activation that was less than the group relying only on their brains, but more than those using ChatGPT. The MIT finding suggests that traditional search engines like Google also require users to engage in substantial cognitive work, such as formulating queries, evaluating sources, synthesising information and constructing their arguments.

The MIT study’s limitations should also be noted. The sample size was small, the task focused only on essay writing, and the study lasted for only four months. Neither has it been peer reviewed. Nonetheless, the study’s findings point to a significant learning gap between those who use AI and those who rely on their memories and past learning.

The UPenn research study employed a randomised controlled trial (RCT) conducted at a high school in Turkey. It involved nearly 1,000 high school students who were studying math. Over four class sessions, students were randomly assigned to one of three groups: those who relied only on textbooks (the control); another used a version that mimics a standard ChatGPT interface that the researchers call GPT Base; the third used a version with learning-focused prompts and teacher-designed safeguards that guided students to answers but never quite completed them, called GPT Tutor.

The research objective was to compare the performance and learning retention across these groups, evaluating both immediate task performance and knowledge retention when AI was removed.

The results were significant. In practice problems, students using GPT Base performed 48 per cent better than those using textbooks. Those using GPT Tutor scored 127 per cent higher than those using textbooks only.

However, when a test was undertaken without AI help, the outcome was reversed. The GPT Base group did 17 per cent worse than the textbook group (the control). The GPT Tutor group did about as well as the control.

The reason for the result reversal was that students using GPT Base often asked the AI for complete answers. They copied them without really trying to solve the problems. During tests, when the AI was not there, they couldn’t solve the problems on their own.

There was also an illusion of learning. The ChatGPT Base group thought they were improving. In post-test surveys, many said they felt confident about their performance. But the data on their sub-par performance suggested they were mistaken.

What connects these two studies is a basic idea: when people let AI do the heavy lifting involved in learning, they incur what the MIT researchers term ‘cognitive debt’. This refers to the build-up of thinking deficits that accumulate over time when students offload their thinking and critical reasoning skills to AI.

It’s like when students with experienced human tutors in, say, math get all their homework done by their tutors and receive full marks for their home assignments throughout the school term. However, when taking tests in school, they struggle to answer questions because they have not seriously engaged with and/or internalised the subject matter.

Learning something new involves discomfort and setbacks, but helps erase cognitive debt. Metacognition, which essentially involves self-reflection about how one learns anything, only develops when students immerse themselves in the learning process by, for instance, taking and summarising notes during lectures, asking questions like ‘do I understand this?’, etc. AI tools can disrupt this process by providing quick answers (sometimes wrong) without requiring inner reflection.

There is a famous quote attributed to Euclid who, when asked by King Ptolemy I if there was a simple way of learning geometry, was reported to have replied: “There is no royal road to geometry".

Obviously, we can’t ban AI – and even if we could, that would be like throwing the baby out with the bathwater. The real challenge is finding the sweet spot. We want to utilise AI’s ability to personalise learning for each student without letting it do the thinking for them. A hint as to how to approach this is provided by the experience with UPenn’s GPT Tutor interface.

One useful learning model of the GPT Tutor type is Khan Academy’s AI tool, Khanmigo, which offers personalised, self-paced instruction. This helps students explore the subject, ask questions of their AI tutor and in turn be quizzed by AI. According to Khan Academy’s website description, Khanmigo focuses on the process not just the answer. It therefore ‘encourages students to explain their reasoning, explore different approaches and arrive at solutions themselves.’

But incorporating tools like Khanmigo in Pakistani schools may not be cost-effective, so alternative remedies should be examined. These could be discussed in a workshop with participation from relevant stakeholders. The need is to encourage responsible AI use by both students and teachers.

The writer is a group director at the Jang Group. He can be reached at: iqbal.hussain@janggroup.com.pk