Categories
Oak updates
27 November 2023Our new AI tools for teachers are just the start
John Roberts
Product and Engineering Director
There is no doubt: interest is soaring in the potential of artificial intelligence (AI) to enhance education. Many conversations focus on pupils and how AI could personalise studying or provide digital one-to-one tutoring.
Less has been said, though, about how it might help teachers develop resources and save time. This is the area where Oak National Academy already works, with thousands of teachers using our resources every week to plan lessons, so it was the natural place for us to start experimenting. This is just the start of the process for Oak, and there are many things we can do to improve our tools.
We faced the same problem as everyone else when working with generative AI tools: the results are not always accurate enough and the content they generate is not always safe enough for classroom use. It’s critical that education AI products are optimised for quality and safety.
Generative AI draws on the general internet. Its training is vast and comprehensive but it’s riddled with inaccuracies, there are in-built biases and parts of it are unsafe for children.
So Oak has been exploring how we can improve this. Not just for our own AI experiments but for everyone.
What are AI tools for teachers?
Accurate and safe resources
We can provide a high-quality and safe set of education resources for AI to draw from and train on. Last month we started to roll out our new curriculum and teaching resources, which will eventually cover the entire national curriculum for England.
They are developed by teachers and have gone through many layers of quality assurance.
All of these resources will be on an Open Government Licence, which means not just Oak but any school, publisher or AI developer can use them as a launchpad for their AI products, safe in the knowledge the content is accurate and safe.
We are also making sure teachers are continually feeding back on how to improve these resources, and their suggestions are acted on.
We've built an AI quiz generator and lesson planner
We’ve been experimenting to see if we can develop AI tools to help teachers create and develop their lessons.
The first tool we have built is a quiz generator. It produces questions and high-quality “distractors” (wrong answers) for multiple-choice questions, a central part of many lessons and time-consuming to create.
We released it in October and enlisted the help of teachers to try it out to see if it meets classroom standards. They created over 6,000 quizzes and their feedback has helped us improve the tool as we go along.
We’ve also launched a second tool where AI helps to generate lesson plans. That is showing good potential, too, with 2,500 lesson plans created by teachers in just the first few days.
“Oak's AI lesson planner offers a really helpful starting point if stuck or low on inspiration and I was particularly impressed at how good it is at highlighting misconceptions for a topic.”
Alex Hawkes, English secondary teacher
It can never be a substitute for highly-skilled teachers planning lessons, but it can make this work more effective as well as more efficient.
Teachers can find both at Oak’s AI tools for teachers and have a go at creating quizzes and lesson plans.
Helping schools to innovate with generative AI in education
In both cases we’ve made sure that teachers are always in charge, that there is “a human in the loop”. All outputs generated can be edited and overridden by teachers, making any necessary changes and adaptations. There is simply no prospect that these tools could replace teachers’ professional judgement.
We also use moderation tools to help ensure that the content that is AI-generated is suitable for the age group and educational context. It’s early days and there’s lots we can do to improve the quality and build upon the feedback we’ve received.
These tools are experiments at an early stage in their development, but it’s important we gain feedback early on. We’ve currently removed certain subjects like Maths and MFL. There's a good reason for this, as they often don’t produce accurate results in those subjects, and require specific development if you are looking to produce educational quality content.
How can they be improved?
Building technology products requires breaking down things to the smallest elements, and optimising for what’s important as you gain feedback. If this happens, all our products should continuously improve. On Oak’s AI products right now we’re focusing on improving and generating high-quality distractors. We think it’s critical that AI products focus on pedagogical quality and alignment.
It’s straightforward to write a prompt to produce a number of multiple choice questions about a topic in isolation using ChatGPT. Many teachers are doing this already, but we should be careful not to anchor here. It’s certainly harder to make sure that these are really high-quality, use the knowledge and vocabulary required, and build upon the prior knowledge in the curriculum.
There are multiple levers we can pull at Oak to improve to do this:
Large Language Model (LLM) Choice: we’re currently using OpenAI GPT-4 but there are many other LLMs out there. For particular subjects, for example Maths, it may be important to look at alternatives.
Prompt optimisation: When any request to an LLM is made this is done via a prompt. We have shared our prompts, and these are updated as we iterate. For example, if you’re generating content for use in the UK, you will need to make sure that you include this in the prompt, but this sort of optimisation will only go so far.
Fine-tuning: This can be done by providing quality-assured datasets to LLMs based on what you expect it to return for a given prompt. Oak has a rich amount of quality content that we can use to tune the underlying LLM we use, for example, high quality human written multiple choice questions, and content from transcripts and slide decks.
Retrieval Augmented Generation (RAG): This is a critical tool for high-quality generative AI in education. It allows for large scale alignment to Oak’s curricula, pedagogy, and can help ensure that results include important contextual information, knowledge and vocabulary by retrieving and embedding the most relevant information in the prompts.
This significantly reduces the risk of hallucination by narrowing to use the pedagogy, facts and knowledge contained within Oak’s resources. For example, in science, it’s critical that multiple choice questions generated contain the correct vocabulary, both in their correct and in the distractors, so that there is alignment to the exam specifications. This involves ensuring that Oak content is available in the form of embeddings to LLMs. In time we can support others by sharing the embeddings we’re using.
We’re currently experimenting with the methods above, which essentially use the amazing power of LLMs but narrowed to the UK national curriculum and Oak’s content. We’re looking forward to seeing if and how this improves generated quality over time.
Enabling high-quality AI-powered tools for everyone
Recently, the Department for Education announced more investment in this work.
It will mean that every teacher in England is set to benefit from new resources powered by AI, for free. We’ll further improve what we offer, making quality improvements, sharing data, and responding to teachers’ feedback about what will help them the most.
But we know that the great potential in education AI is not just in what Oak will do. It’s in the hundreds of innovations that haven’t even been thought of yet.
So, vitally, our new investment will also seek to help others - schools, trusts, publishers and edtech developers. We’ll provide a safe, high-quality body of educational content so innovators in other organisations can build their own AI tools from it, including our embedding.
The future possibilities for others in edtech here are endless, from AI revision helpers to teacher training and feedback tools. We want to make sure that educational quality remains as AI adoption, and tools in the market, rapidly increases.
The aim is clear: to help a wide range of other organisations to create AI-powered tools that teachers have trust and confidence in, that optimise for quality, that save them time, and that allow them to focus more on teaching and supporting pupils.
That’s a big win for us all.
A version of this article first appeared in Tes.