Welcome! Let's start with a question
Here's something to think about before we begin. When you ask an AI to help with your homework, what are you actually after?
Is it a finished piece of work? Or is it a brain that can do that work on its own next time?
Most of the time we tell ourselves it's the second one. In practice, though, it can slide into the first without us noticing. That's what this whole course is about, and in this first level we'll look at what the research says.
Doing a task vs learning from a task
Let's get one idea clear first, because everything else builds on it.
Every assignment you're given has two outputs. There's the product: the essay, the worked answers, the lab report. Then there's the learning: the skill or understanding that the task was meant to build in your head.
Your teacher doesn't need another essay on the causes of the First World War. They've read hundreds. The essay is just a way of getting you to practise building an argument.
Think of a gym. If you get a forklift to lift the weights, the bar still goes up. The product is there. Your muscles haven't changed one bit, though.
So using AI to do a task means the AI does the thinking and you collect the product. Using AI to learn from a task means you still do the thinking, and the AI coaches, explains, questions or checks you along the way.
One quick test you can use: who did the hard thinking here, me or the machine?
Finding 1: the crutch effect (PNAS, 2025)
So does any of this actually show up in results? Yes, and quite clearly.
In 2025, Hamsa Bastani and colleagues published a study in the journal PNAS (Proceedings of the National Academy of Sciences). They ran a field experiment with nearly 1,000 high school maths students in Turkey.
Some students practised with GPT Base, which worked like an ordinary chatbot. Others used GPT Tutor, which was designed with teacher input to give hints rather than hand over answers. A third group had no AI at all.
What happened during practice? Both AI groups did better. GPT Base students scored 48% higher than the no-AI group, and GPT Tutor students scored 127% higher.
Then came an exam with no AI allowed. The GPT Base students did 17% worse than students who'd never had AI. The GPT Tutor students scored about the same as the no-AI group, so the harm had largely gone.
Do you see what happened there? Doing well in practice felt like learning, but it wasn't. The researchers concluded that students had used the AI as a crutch, and once it was taken away they struggled.
Finding 2: a well-designed AI tutor (Scientific Reports, 2025)
Now for the hopeful side.
In June 2025, Gregory Kestin, Kelly Miller and colleagues published a randomised controlled trial in Scientific Reports. It involved 194 students on an introductory physics course at Harvard, and it ran in autumn 2023.
They compared two things. One was a well-run in-class lesson using active learning. The other was a custom AI tutor built on the same teaching principles. It took students through each problem step by step and kept them actively working instead of just giving them solutions.
The result? Students using the AI tutor achieved more than double the learning gains, in less time. They also said they felt more engaged and motivated.
Put those two studies side by side and a pattern jumps out. It isn't really "AI good" or "AI bad". How you use it matters most. When AI does the thinking for you, learning drops. When it makes you do the thinking, learning can improve a lot.
Finding 3: an early warning sign (MIT Media Lab, 2025)
One more, with a health warning attached.
In June 2025, Nataliya Kosmyna's team at the MIT Media Lab posted a study called Your Brain on ChatGPT. They split 54 adults into three groups to write essays: one used ChatGPT, one used a search engine, and one used no tools at all. The researchers measured their brain activity with EEG.
The ChatGPT group showed the weakest brain connectivity, and many of them struggled to quote from essays they'd only just written.
There's an important catch, though. The study was small and hadn't been peer reviewed when it came out, and the authors themselves said the findings were preliminary. They also asked journalists to stop claiming it showed that AI makes you "dumb". So treat it as an early signal and not as proof.
That's a useful habit for the rest of this course. When you hear a claim about AI, ask yourself: when was it published, where, and how big was the study?
Your turn: sort your own tasks
Let's make this personal. Pick three study tasks from the last week or two. For each one, go through three steps in order.
First, name the skill the task was meant to build. Second, ask who should do that thinking, and the answer is almost always you. Third, decide what job AI gets: coach, quizzer or checker if it helps your learning, or nothing at all if any use would take over the thinking.
Here's an example. You're solving chemistry equilibrium questions, and the skill is applying Le Chatelier's principle. If AI just gives you the answers, that harms your learning. If it gives you a hint when you're stuck, or checks your reasoning after you've tried, that helps.
Write your three tasks down. We'll keep coming back to them.
Key takeaways
- Every task has a product and a learning outcome. Your goal is the learning.
- Bastani et al. (PNAS, 2025): open AI access raised practice scores but led to 17% lower exam scores. A hint-based tutor largely removed that harm.
- Kestin et al. (Scientific Reports, 2025): a carefully designed AI tutor more than doubled learning gains compared with an active-learning class.
- The MIT Media Lab study (June 2025) was small, not peer reviewed and preliminary. Treat it as an early signal.
- Ask yourself: who did the hard thinking?