In this episode, I'm going to be sharing a recent workflow where I used AI tools to help my students perform a rapid research task - I’ll talk about what went well, what could have been better and how you can adopt this same approach in your classroom. The AIcademia podcast is a weekly podcast helping educators like you leverage AI in your everyday practice. I'm your host Andy Fisher. And thanks for joining me.

For those of you who have followed along with this podcast so far, and if you have then thanks for supporting the project, you’ll know that the last three episodes have all been quite information dense. In episode 2 I made some predictions for AI development over the next 12 months, we explored the degree to which teachers jobs may be threatened or augmented by AI in episode 3 and last week we examined the benefits and downstream risks of anthropomorphising our technology. All interesting stuff but I want to strike a balance between these kinds of thought provoking issues and more grounded practical tips for teachers who want to find ways of using AI in their classroom right now. 

With this in mind, what I’m sharing with you today couldn’t be more practical and applicable because I’m going to describe a lesson I delivered to one of my classes this week which was built around the effective use of Large Language Models - and then I’ll offer an after actions report on how I think it went and what I’ve learned from the process.

So, Let me set the scene. I teach a weekly course entitled ‘An Introduction to AI’ to a group of Year 10 students as an alternative to a GCSE course. I am one of a number of teachers in my school piloting this kind of non-examined provision - students have a range of electives to choose from including economics, psychology, Space Science, film studies and Sign Language to name just some of those on offer.

I find this opportunity to shape my own curriculum, adapt the pace and content to the pupils’ needs and upskill my own knowledge so I can facilitate their learning, a really refreshing break from the regime of exam preparation that usually dominates keystage 4 and 5 studies.

I have 80 minutes of contact time each week, in one of the school’s IT suites and as often as possible the lessons are hands on practical sessions, although this is then complemented at times by guest speakers who present on topics like the complexities of AI ethics or ways in which AI is already being deployed in the workplace. 

Since September we’ve touched on the use of LLMs, image generation, no-code platforms like Replit and Bolt, and they ended last term by creating their own illustrated children’s bedtime story by combining some of the skills they had learnt.

This half-term, we're refining our use of large language models like Chat GPT, and I wanted to give them a hands on practical task to dust off the cobwebs of the Christmas break.

I arranged for the students to work in pairs, each pair being provided with a web-enabled device. I posted the lesson assignment online along with a range of links to resources. These included blog posts, articles and long-form videos – all designed to provide a complex and nuanced commentary on my topic of choice. I selected these resources because I knew they would introduce challenging vocabulary that would require a glossary and further investigation. 

The resource materials in this case focused on two core issues which were inspired by some of the breaking AI news that was revealed in the first few days of the new year.

First, I asked the students to interrogate the materials I provided to determine if we have we or have not achieved Artificial General Intelligence, or AGI?  This question was sparked by the recent online debate surrounding OpenAI's announcement of their fronteir 03 reasoning model, with some claiming it has met the conditions necessary for AGI, while others disagree. Students would have to analyze the resources, wrestle with the various definitions of AGI, and then assess the arguments on both sides before arriving at their conclusion.

Then there was an extension challenge inspired by an article about the likely economic impact of AGI as and when it is achieved and some responses to that article. Again, it is a contentious and speculative piece of writing which uses economic terms such as UBI and the ‘transformative leveling of capital’ which were selected to stretch the class.

The pairs would have 60 minutes to research, and prepare a brief report of their findings to me. I intentionally provided a lot of material – roughly two and a half hours of video and around 10,000 words of text which I knew was far too much for them to process in real time using traditional reading and viewing methods.

So this lesson was designed with two objectives in mind - to formatively assess their ability to leverage AI research tools to process, digest and understand a range of previously unseen materials and along the way they would be exposed to some thought-provoking content aligned with the course’s broader ambitions - to prepare them for an AI infused future.

So now that you’ve got a clear idea of how I set up the lesson, lets look at the tools they were encouraged to use and how I faciliated that hour once I hit the stopwatch and let the kids loose on the keyboards.

The first AI research method I encouraged them to use involved large language models. I didn't want them relying on just one; instead, I asked them to use at least two different models – perhaps ChatGPT, Perplexity, Claude, Gemini, or Microsoft Pilot – and then to compare outputs. I reminded them of the importance of good prompt engineering. They would paste a web link to one of the sources and ask for a summary, specifying a certain length, reading age and query focus for the output. I asked them to look for citations and suggested they also used rebuttal prompts, which involves asking the LLM to provide a challenge to the viewpoint being expressed in order to strip away the rhetoric and focus on the merits of the ideas themselves.

For video resources, the students needed to find models that could handle transcripts from the URL provided – many of the current models can manage this, but I recommended a specific custom GPT found on the GPT store called ‘Video Summariser’, which works really well and I'll link to it in the show notes.

The more holsitic research option was to use Google Labs’ NotebookLM. This is a tool that allows you to upload multiple sources at the same time and then work with them. The free version allows you to create up to 100 notebooks, and each of those can contain as many as 50 sources including web links, PDFs, word documents and YouTube videos. Each of these sources in turn can be up to 500,000 words long which is another indication that we are well past the point of worrying about context windows. 

If it would be helpful, I’d happily create a video tutorial on how to use NotebookLM although there are already several excellent ones on YouTube already - those of my collegaues who I have shared the tool with are usually very excited by the possible use cases. 

In a nutshell, you load your sources by clicking and dragging them into the left hand panel, and you can then type into the chat window and ask questions about those materials which are conveniently now all in one place. You can choose which sources it draws on when answering or choose to draw on all of them at the same time. You can ask for a summary, a study guide, a glossary of terms and it will even generate a podcast conversation between two AI hosts who convincingly chat away about the materials you’ve provided. Here is an except from an example podcast - this one is a response to Sam Altman’s new year blogpost, in which he claims that superintelligence is now achievable and nearer than we might think...

It’s pretty wild isn’t it and in a recent update, NotebookLM now even allows you to interrupt the hosts by raising your 'hand' and you can then ask questions, and receive responses as if this were a live call-in show. This interactive feature is really quite amazing and I can see a future where the response latency will become so negigable that we will be able to create subject specialist tutors with whom we can chat in real time - if this makes you a little twitchy and you’ve started wringing your hands about the future role of teachers, I’d invite you to listen to last week’s episode which I hope will offer some reassurance.

There is also a paid version of this platform called Notebook LM Plus, with five times the output and workspace, but the free version was more than adequate for the lesson I had planned. Though if your school uses Google Workspace for Education you may already have access to this enhanced version so it might be worth exploring.

So, some pupils chose to use NotebookLM, some worked with LLMs and divided the research load between them and I floated from station to station, overseeing their progress and encouraging the use of best practise prompting while also managing the inevitable login issues which plague most of our sessions.

I can imagine this same work flow being suited to history students researching the causes of the First World War,  a geography lesson which involves getting a quick overview of a case study or a Biology lesson at the start of a unit to lay the groundwork for the class’s exploration of genetics.

Combined with tried and tested teaching activities were already familiar with, there are ways we could build out this AI augmented research approach. For example, the class could be divided in half with two different research tasks being investigated and then recombine the pairs with an ‘each one teach one’ feedback process where each student teaches their partner what they discovered. This ‘learner as teacher’ model has the added advantage of allowing the students to consolidate their understanding and identify any gaps.

So, how did my session go? Well, most students found it to be a positive experience. They valued how quickly they could get an overview of the complex material, but they also recognised the challenge of balancing efficiency without losing the nuances of the original source material. I don’t think this approach is a replacement for careful primary reading of sources - it’s really more a strategy for getting an overview of a topic or concept.

Those who used the LLM approach expressed frustration at hitting prompt limits on the free accounts for models like Claude, which highlights once again the issue of equitable access to AI tools. Unless we can provide our learners with unlimited access to A.I. while on campus which has obvious cost implications, their ability to leverage A.I. will always be bound to the budgets we have and the resources that are currently free to users. While they could, in most cases, switch between platforms, this access barrier is worth noting. 

Those who used NotebookLM did not have to deal with these kinds of constraints but for some reason three pupils were unable to access the platform and none of us were able to figure out why - such are the mysteries of working with technology as a non-tech specialist!

Interestingly, one student was off-task and required robust redirection to actually start the activity and then engage fully. This is quite unusual given it’s an elective course but it reinforces the point  that these AI tools still need a motivated learner at the helm to prove effective. You can lead a horse to digital water but you cannot make him drink or think!

Which brings me neatly back again to my discussion in episode three about whether AI will replace teachers altogether, where I emphasised the importance of teaching standards seven and eight –  we work hard to create a safe and well-managed learning environment and to hold pupils accountable to their highest potential – which is something AI will not replace any time soon. I did get that pupil on track, and by the end, they were producing some useful work but if I hadn’t intervened I suspect they could have filled sixty minutes of fiddling with the mouse, poking the classmate on the next station and typing irrelevant prompts into ChatGPT!

For the rest of the class who were actively engaged throughout, they found that 60 minutes was just a little tight for the amount of content they had been asked to summarise and digest, so the next time I run this approach, I will reduce the number of sources they are working with.

On balance though, I see great potential for this kind of AI-augmented research in schools. I also think It’s a key lifeskill in a world where keeping up-to-date with research and development in any field will be essential for career progression. To remain relevant we will all have to be lifelong learners who can absorb firehoses of information as efficiently as possible. As Alvin Toffler, writer and futurist reminds us, ‘the illiterate of the 21st century will not be those who cannot read and write, but those who cannot learn, unlearn, and relearn.’

Thanks for listening - I hope you’ve found some useful takeaways from the conversation. Please do spread the word if you think others would like the show, and do check out the AIcademia Youtube channel where you’ll find practical tutorials that complement the topics covered on this podcast. Have a great week and I look forward to catching up again soon.