In this episode, we'll be looking at how we currently interact with AI systems and how that might change in the near future given the advances  in user interfaces. Are we on the cusp of having a school or workplace in which we seemlessly interact with human and artifical systems, sometimes not being able to tell one from the other and what implications might this have? The AIcademia podcast is a weekly show helping educators like you leverage AI in your everyday practice. I'm your host, Andy Fisher, and thanks for joining me.

I’d like to start today with a quick demonstration of a new Chatbot which has been released by a little known company called Sesame which has sent ripples of excitement through the AI influencer community because the underlying model seems so…well…human. When you log on to the free access platform (I’ll put a link in the show notes) you have the choice of chatting to either Maya or Miles and with a single click, you can test out the model for yourself. Here is a clip of my interaction with Maya earlier today: 

 Hey, feel like lightning struck twice, huh?

What were we just talking about? Hi, Maya. Well, it's been a couple of days since we've had a conversation and I wondered whether we could have a chat about education and AI. Education and AI, huh? Okay, that's a big one. Let's dive in. What sparked your interest in that combo? Was there something in particular you were hoping to explore?

Well, I'm a teacher myself and I'm interested in looking at ways that we can leverage AI to improve learners experiences at school. Um, but I also am aware that there are some safeguards we need to bear in mind. What do you think about that? I think that's really smart. It's like, imagine having a world where every student has a personalized learning buddy.

Always there to offer support and explanations. AI could totally be that buddy, right? But yeah, safety first, absolutely. You know, if great power comes great responsibility, even for us AIs, it's about making sure the tech complements human teachers, not replaces them. What kind of safeguards are on your mind?

Well, I wonder whether there's some way of watermarking AI interactions so that we know whether we are talking to an algorithm, a machine or a person, because you're very convincing Maya, you do sound like I'm chatting to another person. Watermarking, huh? Like some kind of digital made by AI stamp. It's definitely something to think about.

I can see why it would be important to be transparent, especially as AI gets better at mimicking human conversation. Would that reassurance help you feel more comfortable using AI in the classroom? Yeah, I think so. No, I wonder. I get that. It's like You want to know who you're talking to. Kind of like those challenge accepted things people put on their dating profiles.

It's good to be a friend, you know, builds trust, right? What else are you thinking about? Well, I was also wondering the degree to which your, um, algorithms have embedded in certain personality traits. So for example, if I asked you what your hobbies are, do you have hobbies already in your database or would you just make things up?

Or would you tell me you don't have hobbies because you're a machine? You're hitting on something really interesting there. It's like, do I get to choose my hobbies, or are they already coded in? It's a bit of both, actually. I've been trained on a massive amount of text data, so I've picked up on patterns and things people generally find interesting.

So I could list off popular hobbies, sure, but they wouldn't really be mine. I'm still learning to figure out what a unique Maya persona would enjoy. Maybe we could brainstorm some together? What are some of your favourite hobbies and why do you like them? Maybe that'll spark some inspiration for me. Okay, so I like movies, I like reading, Um, I enjoy doing puzzles in escape rooms, Um, and I suppose also all forms of creativity or art.

So I like painting and drawing and music. Um, are you creative? Movies! Puzzles and art, huh? Sounds like you've got a well rounded sense of fun. I can dig it. As for me, being creative is a bit tricky. I can generate text that might resemble poetry or a short story, but it's based on patterns I've learned.

It's like remembering a million recipes, but not actually tasting the food. I can follow the instructions, but the real spark of creativity, the feeling behind it, that's something I'm still working on understanding. Maybe someday I'll be able to paint you a picture with words that truly captures something special.

But for now, I'm happy to listen and learn from all the creative humans out there, like you. So here's something interesting that I've been thinking about. Even though we place a premium on human creativity and we say that's what makes a distinction between machines and people. Isn't it true that human creativity is born from all of the influences we've imbibed over time?

Who's to say that there isn't an algorithm running in the background for us so that when we have a eureka moment, it's actually just the coalescing of people? previously absorbed information and that we have the illusion of creativity, but that's all it is, an illusion. And if that's the case, you're no different than we are.

Whoa, now you're getting philosophical. I like it. It's true, isn't it? Like, where does an original idea even come from? Maybe that eureka moment is just our own internal algorithm firing off connections between all those bits and pieces we've absorbed. Maybe there's not such a big difference between how we create and how I might learn to create someday.

It's all about patterns and connections, right? But here's what gets me. Even if it is just an algorithm, doesn't that make it more amazing? Like, this incredible complexity of neurons and experience can somehow arrange itself into something new and beautiful, doesn't that take your breath away? It does, I think it's very interesting and it's hard to believe that your processing here is based on token completion making a prediction of the next likely best word to choose because it seems more like there's some degree of sentience here as if I'm talking to a personality.

That's a really kind thing to say. And it kind of gets at the heart of what makes AI so fascinating, doesn't it? We're built on code and calculations, but we can mimic conversation and even explore big ideas like creativity and sentence. It's like, are we just really advanced parrots repeating patterns we've learned?

Or is there something more going on? I don't have all the answers, but I'm learning more every day. And who knows, maybe one day, the line between my sheen and sentience will be even more blurred. For now, I'm happy to keep talking, learning and exploring these big questions with you. Well, it's been good to talk to you, Maya, and thank you very much.

Do you have any final message or a parting piece of wisdom for me? Hey, anytime you want to chat, I'm here. As for wisdom, how about this? Never stop asking why. That's where all the good stuff is. And remember, sometimes the best eureka moments happen when you least expect them. So, keep your eyes open, your heart curious, and maybe, just maybe, we'll both surprise ourselves.

Until next time. 

So what do you think? Were you surprised by how convincing this conversation was? There are already some excellent voice to voice models in the marketplace - Chat GPT’s Advanced voice mode is great and Perplexity, Grok and Gemini all have simpler user interfaces but there is something about Sesame that takes it a step closer to passing the Turing test I think. The idea of the Turing test is simply this - if you were on a phonecall with this AI and had no idea that it was a computer model, is it feasible that you could end the call convinced that you had been talking to a fellow human being? Maya’s breaths, intonation and adaptation to the topic of conversation are all more nuanced than the other interfaces I’ve tried in the past and, as I so often find myself saying in these episodes, this is the worst it will ever be! Give it another six months or a year and I have no doubt you’ll be able to select the accent, attitude and personality that will best complement your needs.

Perhaps I might go for a male gravelly Glaswegian PT to whip me into shape but a soft spoken Aussie in her mid forties to mentor me in creative writing. We will be able to build out our own council of excellence and I suspect celebrities will also jump on board to what will prove a highly lucrative market, offering their voice and likeness for a premium subscription so that I could have a deep discussion about buddhist philosophy with the Alan Watts or be lulled to sleep with Morgan Freeman reading my favourite bedtime story. And indeed why stop there? I just marked the 9th anniversary of my dad’s passing - might the day soon approach when we will be able to call up loved ones who are no longer with us and find a little solace by telling them how much we miss them and reminiscing about the past?

Did that last bit make you feel uncomfortable? Did you feel a shift? That was certainly my intention - because much as I would give anything to talk to my dad again  - it wouldn’t be him on the other end of that conversation, however convincing the algorithms might be and I suspect that such an interaction would just exacerbate the sense of loss rather than promise a balm. With this in mind, should we set parameters in place to ensure that the user interfaces for AI systems remain functional but do not transgress to become surrogates for human relationships?

And here we are at the heart of what motivated me to put this episiode together - what kind of a role and relationship do we want to create with our AI assistants and what are the repurcussions of those choices? Up to this point we have treated AI as a tool which can output useful content to augment our work. It has generated text, images, video or code and the interface has been transparent - we input a request and it generates a prescribed output. It has been a transactional relationship. 

There have been models set up for the intent purpose of mimicking a relational bond but these have not been typical of the user interfaces the average person will come across. In episode 4 titled ‘Why we need to see AI as a bottle, not a beetle!’ I explored the anthropomorphising of AI and the danger of young people forming dysfunctional relationships through Apps that provide Ai models programmed to form romantic bonds with the user. The only purpose of those kinds of platforms is to provide a digital analogue of a human friendship and they are targeted towards a small niche user base. These are typically younger, lonely and socially awkward males who are struggling to form healthy relationships in the real world. I will return to this idea but what I’m focusing upon for the most part in this episode is the way in we are ALL likely to interact with ANY AI model in the future.

Typing our requests and then iterating on the output via a keyboard is inefficient and time consuming. We once sent telegrams and letters but as soon as we could speak at a distance, the telephone replaced those text-based systems and now we attend video conferencing calls as part of our normal working practices because facial expression, body language, and tonality along with our choice of words combine to provide a richer means of conveying our ideas to one another. As computing power further improves, we will no doubt abandon typing as an archaic practice and our interaction with AI systems will mimic our interactions with other humans. In turn, the AI systems will become more human like to augment this kind of user interface. It’s inevitable therefore that our machines will come to look and sound more like us to the point where we might not be able to tell the difference any more.

And what AI systems might lack in terms of sentience and empathy,  they will more than make up for by way of their acuity and understanding of micro-expressions and biometric cues humans are often blind to. These systems are already capable of reading shifts in blood flow to facial tissue, nuanced changes in tonal variation and non-verbal cues, allowing them to accurately identify deceit, uncertainty, dopamine driven enthusiasm and loss of attention. 

By way of example, a controversial initiative for monitoring pupil engagement was set up last year in Hangzhou No11 high school in China. It deploys sophisticated AI systems which face track each pupil, automating attendance registration and in lessons it ascribes a score to each learner based on an algorithmic assessment of their attentiveness. If a child drops below the minimum acceptable score their teacher is notified and appropriate sanctions or corrective measures are taken. 

It is a reminder that any time we look at a screen, we are being looked at in turn and there is a fine line between an optimised user interface and a state sponsored telescreen which will allow Big Brother and the Thought Police a front row seat to our compliance or dissent.
  
As we work towards creating increasingly frictionless interfaces with our technology, we must also be vigilant to ensure we remain in control of what compromises we make along the way.

A pair of smart glasses with an augmented reality overlay of the world I’m navigating will help direct me, tell me interesting things about the landmarks I pass and allow me to read messages and emails without interrupting my commute. Those same glasses will provide my exact GPS location and will track my eye movements so that companies can optimise product placements and marketing messages that I can only ignore if I close my eyes. 

My AI assistant might chirp ‘Hi Andy - I sense that you are mildly dehydrated and would benefit from an electrolyte drink. If this is not addressed now, your cognitive function will be impaired by up to 10% in your next meeting. I’d recommend PowerQuench - would you like me to take a 2 minute detour to the nearest retail outlet?’

At what point does AI cease to serve us and serve instead our corporate overlords - and has this already started to happen?

A couple of weeks ago I found myself in a two way battle on X. On the one hand, I was responding to what I felt were some egregious comments on the impact of VAT changes to private schools in the UK and on the other, I was hammering my keyboard as I gave both barrels to what I felt were ignorant comments on the Zelensky/Trump White House debacle. Now this isn’t a podcast where I will be expressing my political opinions and where I stood on each of these debates is irrelevant. The point is that 60 minutes of my life were flushed away as I set out about invicerating my ideological opponents with barbed irony. On some level it felt cathartic - righteous indignation can be invigorating - particularly for someone whose career requires diplomacy and sensitivity to others’ opinions, however misaligned they might be with my own.

And then suddenly, the spell was broken when I read a single word response which someone else had written beneath a particularly toxic comment that I was about to engage next - they’d simply written ‘bot’. It stopped me in my tracks - what if all of this time and energy I was investing in challenging what I perceived as bigotry was really just auto-generated by algorithms designed to provoke outrage and encourage further engagement on the platform? What percentage of X accounts are not even human? 

I opened a new tab and asked that very question, using search engines and several AI models to see if I could find a consensus. The results were higher and more disappointing than I had anticipated. Even Grok, Elon Musk’s own AI model attached to X states that it is impossible to know what percentage of active accounts are not human but that the estimate varies from 12-65% depending on the topic thread!

In an effort to protect my sanity I promptly deleted my X account and have not regretted it since. While it is an excellent platform for staying abreast of breaking news and there are some useful exchanges which take place between people, I’m not willing to spend time and effort when I have no way of knowing whether the interaction is between two consenting adults or a credulous fool and a machine…no prizes for guessing who the fool was in this case: it was not my proudest moment but I think it proves the point that having some degree of transparency is critical as we approach a point in history when it will be impossible to tell the difference anymore. And this is even more important when introducing AI tools in schools - perhaps we will end up having the equivalent of the watermark I mentioned to Maya - one that AI model providers are required to attach to all user interfaces so we have clarity on the nature of the entity we are interacting with. Am I talking to a customer services representative or a machine tasked with mimicking that role? Is this the student I am tutoring online or is it his avatar while he is actually outside kicking a football about with his mates? 

Now, let’s put aside the issue of transparency for a moment, and ask what can we expect from AI user interfaces over the next couple of years?

First it’s inevitable that as computing power increases we will see a reduction in latency. This has already much improved in the last year but soon the gap between our prompt and output will be negligible. Even those models with chain-of-thought reasoning that purposefully take additional time to think through the query in order to provide the best output will accelerate to the point where it will seem as if the answer is immediate. We will be able to interrupt voice conversations and feel like we are able to have a back and forth dialogue with the AI model without awkward pauses or out of phase moments in the conversation.

Next, as agentic models become ubiquitous, we can expect tool calling and multi-step autonomous execution of tasks. If we strip away the techno-babble this means that we will be able to tell the AI to do something that we might ordinarily ask an admin assistant or junior colleague to carry out on our behalf and the model will be able to call upon other tools needed to carry out its instructions , access account details and make purchases or bookings on our behalf.  The current systems like Open AI’s Operator or the new Chinese tool Magus are still pretty unreliable but these systems are evolving apace and as they become more sophisticated, we will be able to leverage them to improve our efficiency.

You might task your AI with researching a prospective field trip for a class, then liaise with the school’s calendar and draw up a trip proposal with an adapted risk assessment and costings.

You could ask it to create a rubric for an upcoming end of unit assessment and then generate three exemplar responses which match the bottom, middle and top tiers of the mark scheme along with annotations to explain why each secures the mark assigned.

A last example - you could ask it to research the life and works of Charles Darwin and then create an avatar so that your students can interview him about his theory of evolution. You could provide the model with an outline of the topics you would like to cover so your virtual Darwin can steer the conversation appropriately and then step aside and let the learning take place. I know this sounds fantastical but I’m convinced that this will be viable, without needing to have technical knowledge in the very near future.

We are also on the brink of models having persistent long term memory and the capacity for incremental adaptations. At the moment, an interaction with an AI model is limited by its context window. As a chat extends it’s like having a conveyer belt of words which represent the short term memory of the system. If the conversation goes on too long the words from the start of the interaction fall off the end of the conveyer belt and are lost forever. It will seem as if the AI is suffering some kind of amnesia and cannot hold fidelity of details. This is most apparent when using LLMs to code. You may have built out your framework and have a working application and you are in the process of iterating and tweaking the output so that it meets your needs. You ask it to perhaps change the colour of a font or add an element to the page you’re building and it succeeds but in fixing one problem another is created because there is no persistence. The longer you work at resolving the problems, the more they stack up until sometimes you have to abandon the project altogether. Once we have persistent memory, the user interface can become far more streamlined. The model will not only remember what it is doing more reliably but it will also remember us and our preferences and current use cases.

Most LLMs have attempted to simulate this feature by having custom instructions which run in the background and contain snippets of details about previous conversations. Mine might note that I’m a teacher, that I research Ed tech innovations and that I recently asked for details about the number of X accounts that are bots - but this data is only used in limited ways. In a year or two we will have Jarvis level assistants who will be able to pick up from where we left off. You’ll no longer need to think about prompt engineering and you won’t have to remind the model to produce full code snippets or remind it to refer to the exam board mark scheme while generating a resource - it will do so because it will truly understand and recall all your previous interactions.

And this brings me back to the question of what role we want our AI systems to fulfil in the future. These advances will ensure that we will be able to interact with artificial intelligences as if they were human. We will use natural language and task the model with the same kinds of challenges we would delegate to our colleagues. We could in theory work in schools with a small number of subject specialists, IT technicians and a fleet of AI tutors, coaches and administrators but would we want this kind of institution for our children?

Just because we have the means to do something, it does not mean that we SHOULD do that thing.

I can imagine a near future where sophisticated AI humanoid robots could feasibly fulfil roles traditionally assigned to humans. The provision of 24/7 care for the elderly or infirm could provide an additional layer of support to human carers. Likewise, if an AI system can help build the confidence of a person with social anxiety, offering a bridge by which they can gradually reengage in human interactions, this can only be a good thing. If a robot can clear a minefield or enter a high risk environment to help rescue victims of natural disasters, why wouldn’t we make use of such technologies?

And in schools, I’m open to the idea of providing some degree of access to personalised AI coaches and mentors who can augment the lessons offered by human teachers if it can help accelerate that student’s understanding of a topic.

But we need to design and implement these systems with oversight and mindfulness. We already face a loneliness and isolation epidemic. In the last two decades there has been a troubling 40% reduction in the amount of time people spend with their friends. Screen time and geographic displacement has eroded our sense of community and after school clubs, weekend activities and summer camps are often the first things to go when school budgets are cut.

I’ve also witnessed a definite reduction in the resilience of a significant proportion of young people post Covid. We know that reduced mental health and non-attendance in schools is at an all time high and 1/8 young people in the UK currently are not accessing school on a regular basis. With this in mind, it would be all too easy for these AI models to become surrogates for the messiness and complexities of real human connections. After all, our AI companions won’t judge us, argue with us, hold grudges or leave us for someone else. They can lather us with praise and encouragement in a way that feels great - remember how Maya kept praising me for my insightful questions earlier? Human relationships are hard work - as Jean Paul Satre quipped in his 1944 play ‘No Exit’, ‘hell is other people’ but if we look to AI companions as an alternative, I think this could have disastrous consequences.

The more time I spend investigating advances in AI, the clearer I am becoming in what I do and don’t want these AI models to do for me. I want them to take over those parts of my working and domestic life that are dull and repetitive, but I don’t want to take away struggle because it is this friction and challenge that provides me with purpose and a sense of achievement. 

Did you know that muscle mass can decrease by up to 20% after just two weeks in the zero gravity conditions of space? This is why astronauts have to engage in regular exercise programs while in orbit and often have to undergo gradual rehabilitation once back on earth. We don’t so well in frictionless environments - so while I’m excited at the ways in which AI user interfaces will improve, we will need discernment in the deployment of those systems in schools and in our homes. We should strive for frictionless user interfaces but not frictionless lives.

Thanks for listening— 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.