Generative AI is currently a hot topic across various industries, including the edtech industry. Lots of companies are keen to demonstrate their AI capabilities, often believing that genAI can solve any problem. There is even debate about whether genAI models possess human-like sentience.
However, there are also fears and concerns about the social and employment implications of this powerful new AI. Prominent scientists and experts have voiced their worries, suggesting that genAI could potentially destabilize our society. Some have even called for a temporary halt to genAI research to fully understand and mitigate its potential risks.
In the education sector, reactions vary. Some companies are embracing genAI models at times perhaps driven by a fear of missing out. Other organizations have rejected AI in education, citing concerns about its impact on the quality of student learning. Some compare the use of chatGPT to the use of calculators in school.
Let’s explore the technology behind the term genAI and its limitations, and take a look at the potential benefits of AI in edtech.
GenAI is a type of AI system capable of generating text, images, or other media in response to prompts. The most common method involves using text as an input prompt. This technique popularized by Large Language Models (LLM) like ChatGPT and GPT-4, generates text output in response to text prompts. These systems have significantly contributed to the popularity of genAI and sparked interest in its applications in society and education.
However, genAI also encompasses algorithms that handle different types of inputs and outputs:
- If we consider a different input modality than text, we can perform image transformations with image as an input (e.g. runway, stability.ai) and voice conversion with voice input(e.g. Sanas, Resemble.ai).
- Considering alternative output modalities, Text-to-Speech (TTS) models can generate speech given an input text (see for example Facebook’s announcement of a model allowing over 1K languages speech output).
- We can also generate images from a text prompt using Dalle-2 or stability.ai or Midjourney, among others.
- Finally, let’s not forget how developers can also generate code using tools like Github Copilot, and even virtual human presenters (or even full videos) using tools like d-id or elai, among others.
So, we can see that the capabilities of genAI extend far beyond simply generating text.
What are the potential problems of AI in edtech?
AI algorithms are trained using human-generated content. However, this content has flaws, just as humans do, and as a result, AI algorithms also have imperfections. GenAI, in particular, has several well-documented flaws:
Factually incorrect content
It is possible that querying an LLM system with a specific question may yield a well-formulated but factually incorrect answer. The model prioritizes word sequence generation (i.e. an answer) rather than information validation, focusing on probable word sequences. Currently, fact-checking is not implemented, leaving users responsible for verifying information, though algorithms may change this soon. This phenomenon is usually called Hallucination. A less common form of Hallucination occurs when the output is nonsensical in English.
Knowledge limited to the training data
When queried about unfamiliar topics, a standard LLM engine without internet access lacks knowledge -producing false or no response.
Disclosure of confidential information
Training data used in LLM models is believed to be “safe”, i.e. it does not include any confidential information like personal contact details, credit card numbers, etc. However, due to the vast amount of information needed to train LLM models, such content filtering needs to be done automatically, and privacy filtering algorithms (like any other AI algorithms) are not perfect. It is possible that some confidential information inadvertently left online by someone falls into the hands of someone else.
Generation of inappropriate language or content
Similarly, even though training data has hopefully been cleansed of inappropriate language before training, it can still generate harmful content (including hate speech). Although conversational LLM models like chatGPT currently use training techniques that include humans in the loop to mitigate these types of content from being generated, it still happens from time to time. So, appropriate content filters need to be developed and applied to avoid the appearance of this type of harmful content.
Leading users to bad behavior for themselves or others
While no single message from the assistant might be flagged as inappropriate, it’s possible that a conversation with genAI could lead the user to harmful behavior. This is especially true for children or other vulnerable groups.
These issues mean that edtech companies need to be very careful of the way they use generative AI in education.
Successful ways to use generative AI in edtech
To maximize the potential of genAI technology in edtech products, identifying its strengths and leveraging them is crucial.
The biggest strength of LLM models is their ability to generate grammatically and syntactically correct language when prompted with existing text. LLM models learn this skill by being trained using correctly written English texts.
When generating the output, each generated word is selected to be one of the most likely words based on the previous history and the knowledge of what the model has learned, therefore strongly leaning on grammar and syntax from the training data.
This ability allowed us to build use cases that leverage this skill:
1. Text summary
We can ask the LLM model to summarize a longer text while keeping its meaning. This is a helpful way of distilling important information from long or complex texts (e.g. poetry and philosophy).
2. Text comparison
LLM models can also be used to compare two texts and identify the differences and similarities between them. For example, we can ask an LLM to compare a written essay to the gold standard (created by the teacher or automatically via LLM).
Another direct use of LLM models is to rewrite texts for clarity and to correct any English mistakes (or any other languages it has been trained with). In addition to just rewriting, we can also ask the LLM to compare both texts and explain why it decided to make the changes, so that we can learn from our mistakes.
4. Definitions/explanations and translations
LLM models can be a very good and complete dictionary, and a reasonable translator. In both cases, the LLM has an advantage over dictionaries. It takes context into account, therefore, it can disambiguate between homographs and different possible translations. Another advantage is that it covers most expressions and idioms in English. Of course, we need to be cautious, as this use-case can suffer from the hallucination and factual incorrectness problems mentioned above.
5. Generation of new exercises
As an extension of the use-cases above, we can also leverage an LLM to generate, based on just a description, language-related exercises for the students. The type of exercises is just limited by the ability of the content operator to create the right prompt to lead the LLM generation. For example, we can easily create content comprehension exercises with multiple-choice answers, grammar exercises with simulated mistakes, vocabulary exercises on how to use the right words in contect, etc (see this post from Duolingo). As mentioned, we should never consider that the output of the LLM model is good as-is and correct if necessary.
6. Language classifier
LLM models can be used as general-purpose language classifiers. This means that they can classify input text into a variety of predefined categories. For example, we can ask it to tell us what language a text is written in, the emotion/sentiment of the text, its writing tone, etc.
7. Content personalization
While the generation of content by an LLM brings in a wealth of information for edtech users, it can sometimes fall beyond their reach when the LLM generates the content using expressions and concepts that users do not understand, as their level of mastery of English is not good enough. This is where using content personalization techniques help bring the generated output closer to the target audience.
This can be done in a couple of ways. When users are well known (e.g. we know their CEFR levels and mother tongue) we can trigger the LLM content generation tasks by adding in such information so that generated text is constrained to be understood by that English level. However, if users are not known but we do know the type of answers that users expect and understand, we have a second option: few-shot prompting. Few shot prompting uses a few input-output examples into the prompts to the LLM to guide them on what types of responses we are expecting from it.
This is not an exhaustive list of the ways to use generative AI in education. You might find that some of these use-cases might not work well with all kinds of data. Learn from successful cases, adapt effective strategies, and apply them to your specific use-case for optimal results.
How genAI can enhance the learning experience
Despite the growing hype around AI in edtech and other industries, it’s important to consider its limitations and potential risks. By leveraging genAI’s strengths, such as language formulation and understanding, the edtech sector can benefit from improved content generation, summarization, translation, explanations and more. However, caution is necessary as genAI can generate incorrect or nonsensical information. By integrating generative AI with expert systems, we can maximize the benefits of this technology in edtech while ensuring the delivery of accurate and reliable knowledge. Hence, enhancing the learning experience for students.