How could generative AI change early-years education?

Prototype 1: Explain like I’m three

This prototype responds to questions about complex concepts with age-appropriate explanations, aiding ‘contingent talk’, the practice of responding to a child’s behaviour and questions, which is vital for language development. We also instructed the AI to avoid gendered language.

Try it out

Imagine you are an early-years educator in a nursery, working with curious three-year-olds who ask lots of questions: why does the moon change shape? What happens when we sleep? Why do people speak different languages?

Type your questions below to see a response. Note that you can also ask follow-up questions to the chatbot, such as “explain in more detail” if the first response is too simple, or “can you come up with more questions?”

Watch a video of this prototype in use.

Our experts’ verdict

• The explanations generated by the prototype use clear language and creative analogies that will help children to understand answers.
• However, responses are unpredictable, sometimes factually incorrect, and may perpetuate existing biases, which is particularly concerning given children's developing brains. That said, a future iteration of this prototype could personalise content to improve experiences for children from minoritised groups.
• Using this prototype would interrupt a conversation with a child. Engaging immediately with a child to encourage curiosity is more important than giving accurate answers to their questions. Practitioners could instead use this prototype between conversations for inspiration.

Prototype 2: Personalised activity generator

This prototype creates ideas for activities that are tailored to a child’s interests, saving time and creative energy. It refers to the Department for Education’s Development Matters guidance for England, allowing users to strengthen areas of learning that align with the curriculum for that age group.

Try it out

Imagine again that you’re in the nursery, trying to think of activities for a group of children interested in space, snails or fairies. Type in your topic and choose a predefined learning goal to get suggestions for conversations and activities.

If you describe your own learning goal, explain it in a full sentence so that the prototype can give good results.

Watch a video of this prototype in use.

Our experts’ verdict

• The prototype successfully generates a range of creative activities that combine a child’s interests with tasks from the guidance. 
• Some activity descriptions are repetitive or vague, but this prototype would give a practitioner inspiration that they could build upon.
• Practitioners could become overly reliant on activity or advice generators, rather than developing their own knowledge. There is emerging evidence from other sectors that people working with AI tend to “fall asleep at the wheel” and become less skilled in their own judgement.
• The four nations of the UK use different guidelines, so tools should be adapted to work in each.

Prototype 3: Early-years parenting chatbot

Our final prototype is a chatbot for caregivers that provides advice on various topics about babies and toddlers with an accessible and conversational tone. Accessing timely and accurate parenting advice can be challenging, with internet misinformation and limited support services hampering caregivers’ efforts to get help.

We chose the NHS Start for Life website as the basis for the chatbot’s answers because it is an openly available and trusted source of information about topics such as sleeping, feeding, and learning to speak. We also added the ability for users to leave feedback for specific responses, to help us identify problem areas for the chatbot.

Try it out

Now imagine that you are a parent of a three-year-old, dealing with a tricky situation or question you have. Your child might need calming in a doctor’s surgery, be struggling with potty training or you might want some guidance on issues with their friends.

Type your questions below. Note that this chatbot is saving the text conversations to aid our evaluation process, and you can add your feedback to the chatbot’s responses.

Watch a video of this prototype in use.

Our experts' views

• This chatbot could be very helpful to caregivers, particularly those with limited time or fewer trusted contacts.
• The advice our prototype gives is usually accurate and almost always uses a friendly and supportive tone, but it sometimes refers to information from outside the sources we’ve selected. This means it could give advice that comes from less trusted sources on the internet and it isn’t always possible to tell when this is happening.
• Privacy and the use of children's data are a concern given the lack of transparency about how many generative AI tools use the information submitted.
• Such a chatbot would need rigorous evaluation to ensure that the advice it gives is accurate and described correctly. This is a particular challenge in generative AI, where models can sometimes include information from other sources.

Evaluating the chatbot

Effective evaluation against defined criteria is crucial in developing generative AI systems, and refining evaluation methods for large language models is a significant, ongoing research area. 

Guided by expert advice, we tested the chatbot’s outputs by having users compare its responses to human-written answers to parenting questions. The chatbot's responses were deemed sensible and sometimes even preferred, our initial results found. While further testing is necessary for conclusive results, these preliminary tests are beneficial for rapid prototyping, indicating the prototype's effectiveness and identifying areas for improvement.