Can AI Aid Learning?… Instead of Encouraging Cheating?
Right now, if you sit in on a conversation between two academics, unless the discussion is about university gossip or the papers everyone is writing, chances are it will be about the effect Open AI’s GPT-4 and similar large language learning models are having on teaching and learning. This set of new product examples will explore the relationship between these new AIs and learning.
What GPT is… and What it Isn’t
I want to give a somewhat detailed explanation of large language learning models before I introduce our two new product examples, because I think there is a lot of magical thinking about AI right now, which is informing consumer interest in these new product examples. GPT is an artificial intelligence program that ‘learns’ how to write by using extensive text corpora as training input.
To my understanding, the machine is trained to pick the most probable next word of the sentence it is writing based on the previous words and the prompt, though due to the fact that it uses a neural network, we cannot actually explain the procedure by which it chooses that word. (By ‘most probable’ I mean most probable based on all of the other text it was trained on: so it might follow the words ‘Descartes is a’ with ‘philosopher,’ since that is what most often follows.)
There is also some randomness baked into the operation of the neural net (though the question of how a computer can generate randomness is its own technical rabbit hole), meaning that answers will not be the same every time. One big advantage of AI in general as opposed to a more basic program, coded by hand, is that AI models like GPT can be much more flexible in their input, which is necessary for responding to questions written by someone who is not trained in how to use a program.
I give a relatively detailed technical explanation of how GPT works to try to dispel any sense in the reader that it ‘understands’ things in the way that humans experience ourselves to understand things (though of course hardcore materialists will insist that this self-experience is unreliable, and humans are just large language learning models). So, GPT and similar programs can write a decent essay, more free of grammar errors than many college students can produce.
In Academia, GPT Causes More Concern than Excitement
I’ve been in many conversations about GPT with my fellow grad students and professors, and they usually center around our anxieties:
- We worry that our work will be devalued.
- We worry that students will use these technologies to cheat.
- We wonder about how we will have to redesign college coursework to prevent the misuse of these technologies
However, I have participated in fewer conversations about how GPT and similar tech can be leveraged to assist learners. With these new product examples, I will discuss two apps or technologies that use/will use GPT to help teachers and learners. While these new product examples are aimed at primary and secondary education, they may have implications for higher ed as well. They showcase how AI is not currently very helpful in the education sector, and how magical thinking might be driving consumer interest, insofar as there is consumer interest.
New Product Examples in AI:
Duolingo Max
My first new product example is offered by a well-established company. Duolingo (which recently has been providing me with a way to waste my time while feeling like I am practicing my Latin) is using Open AI’s models to power Duolingo Max. Duolingo has long incorporated some AI into its basic program as a cost-saving measure. In the free version of the app, AI is used to help generate more exercises using the same set of pre-written example sentences. So if I am learning Latin, a human Duolingo employee might write an example sentence: “Rotam fortūnae nōn timent” (they do not fear the wheel of fortune). Then Duolingo will use AI to generate a variety of exercises (for example, maybe the app will ask the user to pick the accusative case ending for the word ‘wheel’).
This new product example, Duolingo Max, uses OpenAI’s models to provide two new features to subscribers: Explain My Answer, and Roleplay. According to Duolingo, with Explain My Answer:
“learners can enter a chat with Duo to get a simple explanation on why their answer was right or wrong, and ask for examples or further clarification.”
The other feature, Roleplay, will allow users to converse with a chatbot in the target language, and:
“after the interaction, learners [will] get AI-powered feedback from Duo on the accuracy and complexity of their responses, as well as tips for future conversations.”
Both of these features seem to have limited but genuine value. It seems that both of these features might be replicated for free by using ChatGPT, for instance. However, it appears that Roleplay in particular might provide a good opportunity for the sort of basic practice that language learners need.
Anthony Lawrence, writing for Medium.com, agrees with me that Explain my Answer isn’t as valuable, noting:
“it seems to me that the AI can’t really analyze mistakes. It knows when an answer is wrong, but it doesn’t always see all the reasons that a student might have answered that way.”
When you think about what GPT actually is, as described above, this shortcoming makes perfect sense: of course the AI can’t actually analyze anything. It doesn’t have an understanding of causation; it can only predict text. So Duolingo Max has some limitations. Insofar as a large amount of practice is needed for language learning, this feature might actually be somewhat attractive, as long as it reliably corrects mistakes (a big if).
New Product Examples in AI: Khanmigo
My next new product example also uses OpenAI’s GPT as an engine. Khanmigo is intended as an extension of Khan academy’s existing offerings. Currently, Khanmigo is only available to those who sign up on a waitlist as beta-testers. While the company is a not-for-profit and it does not seem to be advertising Khanmigo as a product, exactly, it is asking for a $20 donation from those who would like to try out using Khanmigo. The product is intended as a “tutor for learners” and “assistant for teachers,” which can both guide students through challenging material and help teachers to formulate lesson plans.
Salazar Khan, founder of Khan Academy, argues that because GPT-4 is so powerful, it can be helpful to students in ways that earlier AI could not. In an interview with Tech & Learning, Khan explained the advantages of the new system:
“What we’re able to get [the newest version of GPT] to do is something like, ‘Good attempt. It looks like you might have made a mistake distributing that negative two, why don’t you give it another shot?’ Or, ‘Can you help explain your reasoning, because I think you might have made a mistake?’”
Instead of simply giving out an answer, then, the hope is that GPT might be able to guide students to think through a problem on their own, like a teacher or a tutor might.
Khanmigo seems more promising than Duolingo Max, though both of these new product examples have similar limitations. Khanmigo seems aimed at the K-12 space, and in particular for helping students work through problems which can be improved with rote practice. This does seem like a good way to employ AI. Furthermore, many teachers do use lesson plans they find online as a basis for their own lesson plans: insofar as GPT is a collating engine, to have it suggest lesson plans to teachers is not that much different from what is already happening.
Pitfalls of AI Technology Affect Both New Product Examples
However, the fact that GPT is trained on large bodies of text is a major issue. There are many subjects on which most of the information that is written is incorrect or misleading, and like any AI program, GPT is only as good as its data.
Another problem with both of these new product examples is cost. In the same interview in Tech & Learning, Erik Ofgang writes:
“Cost is another factor that the Khan Academy team will be studying. These AI tools require a tremendous amount of computing power, which can be expensive to generate, however, costs have been steadily decreasing and Khan hopes this trend will continue.”
When we compare AI to a human teacher, the AI appears to have no marginal cost of use, whereas the teacher demands wages. However, this intuition is misleading: there is actually marginal cost in the form of computing power.
Ultimately, only time will tell how well both of these new product examples will perform. While consumers might initially be interested in these products based on the highest hopes they might have for AI, the limitations of these new product examples will likely reveal themselves quickly to consumers and be reflected in the market. At the same time, it cannot be denied that educational companies, like all other sectors, will probably stand to gain a lot by automating more of their simplest tasks in order to save costs.
Neither of these new product examples can do everything we might wish for, and neither is completely awful either. It is important to understand how the underlying AI technology works in order to evaluate their usefulness and market potential.
If you are curious to learn more about evaluating new product marketability, the experts at Insight to Action can help. Our resources page features more articles about new products, growth strategy, and more. You can also attend our office hours to speak with an ITA team member, or sign up for our newsletter to hear from us on a more regular basis.
