ChatGPT is Generating Nonsense and No One Knows Why

Every time I start a new post over the past couple of months, it just devolves into a “I told you so” take on some aspect of AI. Which, of course, just gets a little old after trying to tell people that learning analytics, virtual reality, blockchain, MOOCs, etc, etc, etc are all just like any other Ed-Tech trend. It’s never really different this time. I read the recent “news” that claims “no one” can say why ChatGPT is producing unhinged answers (or bizarre nonsense as others called it). Except for, of course, many people (even some working in AI) said this would happen and gave reasons why a while back. So, as usual, they mean “no one that we listen to knows why.” Can’t give the naysayers any credibility for knowing anything. Just look at any AI in education conference panels that never bring in any true skeptics. It’s always the same this time.

Imagine working a job completely dependent on ChatGPT “prompt engineering” and hearing about this, or spending big money to start a degree in AI, or investing in an AI technology for your school, or any other way people are going big with unproven technology like this?  Especially when OpenAI just shrugs and says “Model behavior can be unpredictable.” We found out last week just how many new “AI solutions” are just feeding prompts secretly to ChatGPT in the background.

Buried at the end of that Popular Science article is probably what should be called out more: “While we can’t say exactly what caused ChatGPT’s most recent hiccups, we can say with confidence what it almost certainly wasn’t: AI suddenly exhibiting human-like tendencies.” Anyone that tries to compare AI to human learning or creativity is just using bad science.

To be honest, I haven’t paid much attention to the responses (for or against) my recent blog posts, just because too many people have bought the “AI is inevitable” Kool-Aid. I am the weirdo that believes education can choose it’s own future if we ever just would choose to ignore the thought leaders and big money interests. Recently Ben Williamson outlined 21 issues that show why AI in Education is a Public Problem with the ultimate goal of demonstrating how AI in education “cannot be considered inevitable, beneficial or transformative in any straightforward way.” I suggest reading the whole article if you haven’t already.

Some of the responses to Williamson’s article are saying that “nobody is actually proposing” what he wrote about. This seems to ignore all of the people all over the Internet that are, not to mention that there have been entire conferences dedicated to saying that AI is inevitable, beneficial, and transformative. I know that many people have written responses to Williamson’s 21 issues, and most of it boils down to saying “it happened elsewhere so you can’t blame AI” or “I haven’t heard of it, so it can’t be true.”

Yes, I know – Williamson’s whole point was to show how AI is continuing troubling trends in education. We can (an should) focus on AI or anything else that continues those trends. And he linked to articles that talked about each issue he was highlighting, so claiming no one is saying what he cited them as saying is odd. AI enthusiasts are some of the last holdouts on Xitter, so I can’t blame people that are no longer active there for not knowing what is being spread all over Elon Musk’s billion dollar personal toilet. Williamson is there, and he is paying attention.

I am tempted to go through the various “21 counterpoints / 21 refutations / 21 answers / etc” threads, but I don’t really see the point. Williamson was clear that he takes an extreme position against using AI in schools. Anyone that refutes every point, even with nuance, is just taking an extreme position in a different direction. To do the same would just circle back to Williamson’s points. Williamson is just trying to bring attention to the harms of AI. These harms are rarely addressed. Some conferences will maybe have a session or two (out of dozens and dozens of session) that talk about harms and concerns. Usually “balanced out” with points about benefits and inevitability. Articles might dedicated a paragraph or two. Keynotes might make mention of how “harms need to be addressed.” But how can we ever address those harms if we rarely talk about them on their own (outside of “pros and cons” arguments), or just refute every point anyone makes about their real impact?

Of course, the biggest (but certainly not best) institutional argument against AI in schools comes from OpenAI saying that it would be “impossible to train today’s leading AI models without using copyrighted materials” (materials that they are not compensating the copyright holders for their intellectual property FYI). Using ChatGPT (and any AI solution that followed a similar model) is a direct violation of most school’s academic integrity statements – if anyone actually really meant what they wrote about respecting copyright.

I could also go into “I told you so”s about other things as well. Like how a Google study found that there is little evidence that AI transformer model’s “in-context learning behavior is capable of generalizing beyond their pretraining data” (in other words, AI still doesn’t have the ability to be creative). Or how the racial problems with AI aren’t going away or getting better (Google said that they can’t promise that their AI “won’t occasionally generate embarrassing, inaccurate, or offensive results”). Or how AI is just a fancy form of pattern recognition that is nowhere near equatable to human intelligence. Or how AI takes more time and resources to fix than just doing it yourself first. Or so on and so forth.

(Of course, very little of what I say here is really my original thought – it comes from others that I see as experts. But some people like to make it seem like I making up a bunch of problems out of thin air.)

For those of us that actually have to respond to AI and use AI tools in actual classrooms, AI (especially ChatGPT) has been mostly a headache. It increases administration time due to dealing with all the bad output it generates that needs to be fixed. Promises of “personalized learning for all” are almost meh at best (on a good day). The ever present existence of the uncanny valley (that no can really seem to fix) makes application to real world scenarios pointless.

Many are saying that it is time to rethink the role of the humans in education. The role of humans has always been to learn, but there never really was one defined way to do that role. In fact, the practive of learning has always been evolving and changing since the dawn of time. Only people stuck in the “classrooms haven’t changed in 100 years” myth would think we need to rethink the role of humans – and I know the people saying this don’t believe in that myth. I wish we had more bold leaders that would take the opposite stance against AI, so that we can avoid an educational future that “could also lead to a further devaluation of the intrinsic value of studying and learning” as Williamson puts it.

Speaking of leadership, there are many that say that universities have a “lack of strong institutional leadership.” That is kind of a weird thing to say, as very few people make it to the top of institutions without a strong ability to lead. They often don’t lead in the way people want, but that doesn’t mean they aren’t strong. In talking with some of these leaders, they just don’t see a good case that AI has value now or even in the future. So they are strongly leading institutions in a direction they do see value. I wish it would be towards a future that sees the intrinsic value of studying and learning. But I doubt that will be the case, either.

What if We Could Connect Interactive Content Like H5P to Artificial Intelligence?

You might have noticed some chatter recently about H5P, which can create interactive content (videos, questions, games, content, etc) that works in a browser through html5. The concept seems to be fairly similar to the E-Learning Framework (ELF) from APUS and other projects started a few years ago based on html5 and/or jquery – but those seem to mostly be gone or kept a secret. The fact that H5P is easily shareable and open is a good start.

Some of our newer work on self-mapped learning pathways is starting to focus on how to build tools that can help learners map their own learning pathway through multiple options. Something like H5P will be a great tool for that. I am hoping that the future of H5P will include ways to harness AI in ways that can mix and match content in ways beyond what most groups currently do with html5.

To explain this, let me take a step back a bit and look at where our current work with AI and Chatbots currently sits and point to where this could all go. Our goal right now is to build branching tree interfaces and AI-driven chatbots to help students get answers to FAQs about various courses. This is not incredibly ground-breaking at this point, but we hope to take this in some interesting directions.

So, the basic idea with our current chatbots is that you create answers first and them come up with a set of questions that serve as different ways to get to that answer. The AI uses Natural Language Processing and other algorithms to take what is entered into a chatbot interface and match the entered text with a set of questions:

Diagram 1: Basic AI structure of connecting various question options to one answer. I guess the resemblance to a snowflake shows I am starting to get into the Christmas spirit?

You put a lot of answers together into a chatbot, and the oversimplified way of explaining it is that the bot tries to match each question from the chatbot interface with the most likely answer/response:

Diagram 2: Basic chatbot structure of determining which question the text entered into the bot interface most closely matches, and then sending that response back to the interface.

This is our current work – putting together a chatbot fueled FAQ for the upcoming Learning Analytics MOOCs.

Now, we tend to think of these things in terms of “chatting” and/or answering questions, but what if we could turn that on its head? What if we started with some questions or activities, and the responses from those built content/activities/options in a more dynamic fashion using something like H5P or conversational user interfaces (except without the part that tries to fool you that a real person is chatting with you)? In other words, what if we replaced the answers with content and the questions with responses from learners in Diagram 1 above:

Diagram 3: Basic AI structure of connecting learners responses to content/activities/learning pathway options.

And then we replaced the chatbot with a more dynamic interactive interface in Diagram 2 above:

Diagram 4: Basic example of connecting learners with content/activity groupings based on learner responses to prompts embedded in content, activities, or videos.

The goal here would be to not just send back a response to a chat question, but to build content based on what learner responses – using tools like H5P to make interactive videos, branching text options, etc. on the fly. Of course, most people see this and think how it could be used to create different ways of looking at content in a specific topic. Creating greater diversity within a topic is a great place to start, but there could also be bigger ways of looking at this idea.

For example, you could look at taking a cross-disciplinary approach to a course and use a system to come up with ways to bring in different areas of study. For example, instead of using the example in Diagram 4 above to add greater depth to a History course, what if it could be used to tap into a specific learner’s curiosities to, say, bring in some other related cross disciplinary topics:

Diagram 5: Content/activity groupings based on matching learner responses with content and activities that connect with cross disciplinary resources.

Of course, there could be many different ways to approach this. What if you could also utilize a sociocultural lens with this concept? What if you have learners from several different countries in a course and want to bring in content from their contexts? Or you teach in a field that is very U.S.-centric that needs to look at a more global perspective?

Diagram 6: Content/activity groupings based on matching learner responses with content and activities that focus on different countries.

Or you could also look at dynamic content creation from an epistemological angle. What if you had a way to rate how instructivist or connectivist a learner is (something I started working on a bit in my dissertation work)? Or maybe even use something like a Self-Regulated Learning Index? What if you could connect learners with lessons and activities closer to what they prefer or need based on how self-regulated, connectivist, experienced, etc they are?

Diagram 7: The content/activity groupings above are based on a scale I created in my dissertation that puts “mostly instructivist” at 1.0 and “mostly connectivist” at 2.0.

You could also even look at connecting an assignment bank to something like this to help learners get out of the box ideas for how to prove what they have been learning:

Diagram 8: Content/activity groupings based on matching learner responses with specific activities they might want to try from an assignment bank.

Even beyond all of this, it would be great to build a system that allows for mixes of responses to each prompt rather than just one (or even systems that allow you to build on one response with the next one in specific ways). The red lines in the diagrams above represent what the AI sees as the “best match,” but what if it was indicating the percentage of what content should come from which content pool? The cross-disciplinary image above (Diagram 5) could move from just picking “Art” as the best match to making a lesson that is 10% Health, 20% History, 50% Art, and so on. Or the first response would be some related content on “Art,” then another prompt would pull in a bit from “Health.”

Then the even bigger question is: can these various methods be stacked on top of each other, so that you are not forced to choose sociocultural or epistemological separately, but the AI could look at both at once? Probably so, but would a tool to create such a lesson be too complex for most people to practically utilize?

Of course, something like this is ripe for bias, so that would have to be kept in mind at all stages. I am not sure exactly how to counteract that, but hopefully if we try to build more options into the system for the learner to choose from, this will start dealing with and exposing that. We would also have to be careful to not turn this into some kind of surveillance system to watch learners’ every move. Many AI tools are very unclear about what they do with your data. If students have to worry about data being misused by the very tool that is supposed to help them, that will cause stress that is detrimental to learning. So in order for something like this to work, openness and transparency about what is happening – as well as giving learners control over their data – will be key to building a system that is actually usable (trusted) by learners.

Social Presence, Immediacy, Virtual Reality, and Artificial Intelligence

While doing some research for my current work on AI and Chatbots, I was struck by how much some people are trying to use bots to fool people into thinking they are really humans. This seems to be a problematic road to go down, as we know that people are not necessarily against interacting with non-human agents (like those of us that prefer to get basic information like bank account balances over the phone from a machine rather than bother a human). At the core, I think these efforts are really aimed at humanizing those tools, which is not a bad aim. I just don’t think we should ever get away from openness about who or what we are having learners interact with.

I was reminded about Second Life (remember that?) and how we used to question how some people would build traditional structures like rooms and stairs in spaces where your avatars could fly. At the time it was the “cool, hip” way to mock the people that you didn’t think “understood” Second Life. However, I am wondering if maybe there was something to this approach that we missed?

Concepts like social presence and immediacy have fallen out of the limelight in education, but they still have immense value (and many people still promote them thankfully). We need something in our educational efforts, whether in classrooms or at a distance online, that connects us to other learners in ways that we can feel, sense, connect with, etc. What if one way of doing that is by creating human-based structures in our virtual/digital interactions?

I’m not saying to ditch anything experimental and just recreate traditional classroom simulations in virtual reality, or re-enact standard educational interactions with chat bots. But what if incorporating some of those elements could help bring about more of a human element?

To be honest, I am not sure where the right “balance” of these two concepts would be. If I enter a virtual reality space that is just like a building in real life, I will probably miss out on the affordances of exploration that virtual reality could bring to the table. But if I walk into some wild trippy learning space that looks like a foreign planet to me, I will have to spend more time figuring out the way things work than actually learning about the topic I am interested in. I would also feel a bit out of contact with humanity of there is little to tie me back to what I am used to in real life.

The same could be said about the interactions we are designing for AI and chatbots. On one hand, we don’t need to mimic the status quo in the physical world just because it is what we have always done. But we also don’t need to do things that are way out there just because we can, either. Somewhere there is probably a happy medium of humanizing these technologies enough for us to connect with them (without trying to trick people into thinking they are humans) while still not replicating everything we already know just because that is what we know. I know some Social Presence Theory people would balk at the idea of those ideas being applied to technology, but I am thinking more of how we can use those concepts to inform our designs – just in a more meta fashion. Something to mull over for now.