Leon Furze

Last year I wrote a series of posts on LinkedIn, followed by a longer article, explaining why I think chatbots are a dead end in education.

Amongst the criticisms I received were many variations on a theme of “this is the worst the technology will ever be” and “you don’t understand how the technology works”. In that first article, I also gave a definition of “chatbot” that encompassed the kinds of low-stakes, transactional technologies I was seeing emerging through edtech and big tech. This was an attempt to make clear that I’m not writing off all uses of Generative AI: just the ones that most companies seem to be focused on.

Here’s my earlier definition in full. Chatbots are:

• An extension of earlier (pre-generative AI) technologies that have been used for years in helpdesks and for online support functions
• Extended by generative AI, which has allowed for more “natural” dialogue than the limited earlier technologies
• Largely text-based, with turn-by-turn interactions between the user and the chatbot
• Sometimes “voiced” through text-to-speech software
• Generally, in GAI contexts, a “wrapper” built with an API connected to a foundation model (mostly GPT 3.5 at this stage, but there are many options including open source)
• Tailored to a specific use case, generally by processes like fine-tuning on proprietary data (Khanmigo) or connection to a user’s data through a process like Retrieval Augmented Generation

In my definition, chatbots aren’t:

• Fully multimodal systems, like the upcoming ChatGPT update* which blends image recognition, image generation, advanced data analysis, and internet access
• “System” level rather than application level, like Microsoft’s Copilot is intended to be, or Google’s Duet for Workspace
• Capable of interacting with many other apps and services, like an AI-powered conversational assistant (Siri, Alexa, etc.)

In a nutshell, chatbots by this definition are generative AI-based tools, primarily text based and tailored to specific, limited use cases.

A few updates are probably necessary 10 months on from the original post. This * “upcoming update” was GPT-4, which has since been superseded (and made free) by GPT-4o. That has essentially put multimodal AI, including image recognition and generation, internet search, voice capabilities, and code interpreter into the hands of over 100 million users worldwide, including students.

System level applications continue to be in vogue. Since my article was published, we have seen flashy demonstrations from OpenAI and Google about the integrations of their AI across multiple levels of applications. Microsoft have released their Copilot+ devices, including the controversial ‘Recall’ feature. And Apple have finally joined the party, with their announcement at WWDC of “Apple Intelligence” and a partnership with OpenAI. This includes a much needed update to Siri which, frankly, has never actually worked.

Meanwhile, in education…

And while the world turns its attention to multimodal GenAI that eats entire operating systems, we’re still stuck with this:

Now I’ve criticised Khanmigo before, including the narrow understanding of the writing process implied by the chatbot. I’ve also criticised homework, in general, so a chatbot that mashes the two together is never going to be that appealing to me.

As a nonprofit working in education, I’d love to back Khan Academy, but it’s getting harder and harder. The recent partnerships with OpenAI and Microsoft both point to corporate takeover. Sal Khan’s recent book is a lengthy sales pitch targeted more at parents than educators, and shows clearly enough the interests of the company are no longer just aligned with providing “a free, world-class education to anyone, anywhere.”

While tech companies race to out-do one another with bigger and more powerful models, more integrated features, and multimodal tools that might be used in exciting and interesting ways in the future, I can’t help but feel that education is getting a raw deal.

Why are we stuck with chatbots designed to support outdated practices (like homework), fine tuned on outdated curricula, and predicated on a flawed, outdated understanding of teaching and learning?

If I were a student confronted with a homework helper chatbot, I’d switch of almost immediately, no matter how many emojis you stuck on it.

Here’s a little video of me channelling my 14-year-old self, and thinking about different ways to approach Generative AI.

Transcript via Otter, tidied up by Claude 3 Opus:

Multimodal Generative AI and a Glimpse into My 14-Year-Old Self

Hi, I’m Leon Furze, and this is a quick fly-through of some multimodal generative AI, and a little bit of a glimpse into what I would have been like as a 14-year-old if I was using these technologies.

Using ChatGPT for English Homework

The main way that I’m seeing these technologies used in education is obviously as a little tutor chatbot thing. So here’s me using ChatGPT to do my English homework, something pretty standard: grammar exercises, subject-verb agreement, pronouns, tenses, punctuation. And this is really my opinion about chatbots like this, in general. This is pretty boring. I would get distracted almost immediately. I’d be looking for other things to do.

Creating a Lava Lamp Visualisation

So let’s see if ChatGPT can help me make a lava lamp style visualisation in HTML. Of course it can. Pretty straightforward, it’s going to create all of the HTML, CSS, JavaScript, and we can grab all of that, then grab the downloadable files, and bring those in and run our little lava lamp. So I’m already distracted from English homework. The English homework helper chatbot isn’t holding my attention. We’ve made this pretty cheap and cheerful lava lamp. It doesn’t look exactly like the lava lamps that I remember from my youth, but you know, it’ll do. It’s a good effort. I’m going to run this in the background.

Generating Music and Visualisations

And whilst I’m doing that, I’m going to jump into a platform like Suno or Udio and generate some music. Now, let’s just point out at this stage that I’m not actually using Suno and Udio for anything, because I’ve got some serious questions about the copyright and intellectual property. But this is 14-year-old me. I probably don’t care. So let’s create a music visualiser that goes along with those audio tracks. Let’s take the file input as a web browser. And because this is 14-year-old me, let’s make it look like Winamp from the 2000s. It was and still is the best music player ever. We’ve made some little equaliser bars. And now we can get some of this music playing in the background.

All right, great. So I’ve got my very basic Winamp style music visualiser. I’m going to jump over to MetaAI on my phone just in the Facebook Messenger app, and create some album artwork for a Lo-Fi chillhop album. We’ll make it illustrated in an anime style with bold colours and a cat, because it’s not Lo-Fi if it doesn’t have a cat. And just there on my phone, I obviously have access to Meta’s image generation model. So I’m switching around between apps now, like a typical 14-year-old. I’m not really maintaining a great deal of focus on my chatbot. Let’s see if I can bring the artwork into my Winamp app.

Of course, that’s as simple as just updating that initial bit of code that we got from ChatGPT. And there we go. So we’ll manually load the tracks in. We’ve got the equaliser, we’ve got the album artwork, and we’re ready to go.

I think the next step, instead of doing my English homework with my homework helper chatbot, should probably be to make a few more tracks, generate a few more tracks in Sooner, and let’s have some playback controls. So we’re gonna get some things that can cycle through the tracks in that folder. And again, because this is Winamp circa the 2000s, we’re going to load all of those tracks manually as WAV files. So again, nice and easy. Let’s have a listen to a few of those tracks.

So it sounds pretty much like every Lo-Fi album I’ve ever listened to on Apple Music. But now I’m generating all of these things myself and, importantly, I’m doing it whilst I should be doing my English homework, which is much less interesting to me.

Turning the Project into a Product

Alright, so 14-year-old me was a little bit entrepreneurial and always looking to make a quick pound or dollar. So I’m gonna go over to Google’s Gemini Ultra Pro now and get it to analyse the screen recording that I’ve just made of that whole process and ask me how I can actually turn this into a product. So it’s gone away. It’s analysed the video, a very capable multimodal model. It can take a lot of data. It’s taken the 20-minute screen recording and it’s given me a step-by-step process for staging this for development.

Now I’ve got my music player, I’ve got my lava lamp, I’ve got the visualiser, my Winamp, I’ve got some songs. I should probably get back to doing my homework.

I think we can do better than homework helper chatbots, and our students will certainly think of more interesting ways to use Generative AI. I talk more about the (near) future of Generative AI in my online course, Practical AI Strategies.

I regularly work with schools, universities, and faculty teams on developing guidelines and approaches for Generative AI. If you’re interested in talking about consulting and PD, get in touch via the form below: