What AI Tools for Education Don't Tell You and Why It Matters
More teachers are embracing AI tools than ever before. Yet, a crucial element is often ignored: the Large Language Models (LLMs) powering the tools they use.
I've reviewed hundreds of AI for education tools, and they harbor one almost universal secret: their LLM. The Large Language Model is the single most important component of any AI-based tool. It's essentially the engine that drives the tool, impacting the AI's ability to understand and learn context from user input and provide high-quality, tailored content. It also affects the tool's creativity, among other considerations.
Choosing an AI tool without knowing its underlying LLM is like buying a car without the knowing its engine specifications. Imagine purchasing a vehicle based primarily on its dashboard features and comfortable seats, and without any information about its engine's power, efficiency, or reliability. Yet, that’s exactly what we do.
Naturally, users would prefer a high-quality LLM driving an AI tool. A more powerful engine takes you further. However, purchasing access to a powerful and advanced LLM is expensive for developers. So, often they don't.
The result? Many teachers use AI tools driven by GPT-3.5. Or, they use the free version of an AI tool, which GPT-3.5 likely drives. If you pay up, you might get GPT-4. But many teachers only sign up for the free version of AI tools.
(Note: GPT-4 level variants include GPT-4 Turbo, GPT-4o (Omni), and now GPT-4o mini — more on that later.)
This issue becomes more prominent as teachers spend more time in AI teaching assistants such as MagicSchool, SchoolAI, Brisk Teaching, and Eduaide. These assistants offer great dashboard features (lots of tools) and comfortable seats (minimal prompting), but they often pale in power and flexibility compared to ChatGPT’s GPT-4o (Omni), or an equivalent. It’s almost impossible to find any information about their LLMs. This issue is compounded as students are now entering some of these AI teaching assistants to work with AI chatbots.
To be fair, AI developers often introduce features specifically designed for educational settings, beyond what ChatGPT offers. Examples include:
• Teacher controls for implementing AI in classrooms.
• Transparency in student engagement with AI models.
• Specialized features like AI-assisted assignment grading.
These enhancements aim to provide greater utility and save teachers time. And they cost time and money. While GPT-4o, for instance, offers valuable tools, it lacks several capabilities essential for schools.
That said, GPT-4 and GPT-3.5 are worlds apart in their capabilities and complexity. GPT-4's vastly larger parameter count - estimated to be in the trillions compared to GPT-3.5's 175 billion - allows it to grasp and generate information with far greater nuance and accuracy.
My new book AI Tools & Uses: A Practical Guide For Teachers is now available at Amazon! It is the first of its kind to be organized by K-12 subject areas and teaching responsibilities, and it offers unprecedented depth and specificity for teachers. The book provides detailed analyses of AI tools, including both subject-specific tools and versatile platforms that can be leveraged across disciplines.
The advantages of GPT-4 extend beyond just size. Its architecture is more sophisticated, enabling it to understand context more deeply and maintain coherence over longer interactions. It's like comparing a basic calculator to a graphing calculator - both can do arithmetic, but one is capable of much more complex operations.
For K12 classrooms, these differences become significant. In math tasks, GPT-3.5 can handle basic arithmetic and straightforward algebraic problems but struggles with more complex mathematical concepts, especially those involving multiple steps or abstract reasoning. Errors are more frequent, particularly in problems requiring precision and logical progression. Meanwhile, GPT-4 excels at handling complex math tasks, thanks to improved pattern recognition, logical reasoning, and computational accuracy. It can more reliably solve multi-step problems, understand advanced mathematical concepts, and offer accurate explanations. GPT-4 is far from flawless, and can flub math tasks, but AI’s math progression is impressive.
Language learning with GPT-4 is also leaps better. GPT-3.5’s language abilities are relatively strong, but it has limitations in understanding nuanced contexts, idiomatic expressions, and less-common language constructs. It may also struggle with maintaining coherence in an extended conversation or delivering contextually appropriate responses. GPT-4 exhibits a more advanced understanding of language. It can handle nuanced language use, including idioms, metaphors, and cultural references, with greater accuracy. It also maintains coherence in extended conversations and provides more contextually relevant responses. So, why converse with GPT-3.5 to learn a language?
GPT-4's multimodal capabilities - its ability to process and understand text, images, audio, and video - open up even more possibilities for rich, interactive learning experiences across K12 subjects, well beyond GPT-3.5 This versatility makes GPT-4 a powerful tool for creating engaging, multifaceted lessons that cater to different learning styles.
Moreover, GPT-4's advanced training techniques have made it significantly safer and less biased than its predecessor. This increased reliability is crucial in educational settings where accuracy and appropriateness of content are paramount. One estimate puts GPT-4 as 82% less likely to generate disallowed content than GPT-3.5. Why would you choose an LLM that is much more likely to produce biased content?
These differences have real-world implications for student learning and teacher effectiveness. Students using tools powered by GPT-3.5 might receive less accurate information, oversimplified explanations, or struggle with more complex topics. Teachers relying on GPT-3.5-based tools for lesson planning or content creation might find the generated material lacking in depth or sophistication.
In many instances, teachers would be better off using GPT-4o in ChatGPT. However, this requires better prompting skills. Developing these skills takes time but can lead to more effective use of AI in education, allowing teachers to harness the full potential of advanced LLMs like GPT-4o.
Another barrier is student privacy. ChatGPT doesn’t protect it, so many school districts shy away from it.
In a move that could reshape the AI tools in education landscape, OpenAI launched GPT-4o mini in July 2024. This cost-efficient, smaller model is destined to replace GPT-3.5, bringing better performance at a more affordable cost. It's particularly appealing for developers looking to create AI applications without the high compute costs associated with larger models.
However, I don’t expect the adoption of GPT-4o mini to be universal. In my conversations with several AI for education developers, I've found a hesitancy to purchase GPT-4o mini. Despite its promised improvements over GPT-3.5, some developers are adopting a wait-and-see approach, seemingly due to integration issues or budget constraints.
Adding another layer of complexity to the decision-making process is the anticipated release of GPT-5. Expected any week now, GPT-5 is likely to widen the capability gap even further. This imminent release may be another factor causing some developers to hesitate in upgrading from GPT-3.5, as they weigh the benefits of immediate improvements against potential near-future advancements.
For educators and institutions, this rapidly changing landscape underscores the importance of staying informed about the AI tools they're using. While GPT-4o is now available to free ChatGPT users, offering features like multimodality and voice chat previously reserved for paying customers, the full rollout is still in progress.
As educators explore AI tools this summer, it's crucial to ask whether GPT-4o could do a better job than the specialized tools they're considering. While these tools offer convenience and a bevy of features, relying solely on them can be misguided.
I recommend the following:
1. Leverage GPT-4o as much as possible. Its superior capabilities can enhance various aspects of teaching and learning.
2. Invest time in improving your prompting skills. This will allow you to use more powerful LLMs effectively.
3. Resist the urge to sign up for a slew of free versions of AI education products. Many of these might be using inferior LLMs.
4. When exploring new AI tools, always question whether GPT-4o could perform the same task more effectively.
In conclusion, the power of AI tools in education lies largely in their LLMs. As we explore new AI tools, let's prioritize the quality of the AI "engine" over flashy features, ensuring we're providing our students with the best possible learning experiences.
AI Tools & News
Google School Time - a new setting for Android phones and other devices that empowers parents to limited their functionality during school hours.
How Schools Are Holding Edtech Products to a Higher Standard - EdSurge
Teaching Writing in the Age of AI - An Experiential Approach - Phillip Alcock
Zoom Is Going After Google and Microsoft With AI-Driven Docs - Wired
Dezn - transform diagrams into interactive experiences
Leveraging AI to help special education teachers - eSchoolNews
Creating a Student Interactive with Claude: It’s Easier Than You Think! - TCEA
AI Revolution: How one Utah teacher and a student say AI brings learning to life - KSLTV
CAPE AI Prompt Maker: Applying Chain-of-Thought to Instruction - TCEA
Why I’m Banning Student AI Use This Year - Edutopia
Imagine Me - new Meta AI to create AI version of yourself based on a prompt
Should Educators Put Disclosures on Teaching Materials When They Use AI? - EdSurge