Categories
AI in education
5 September 2024Aila’s safety guardrails explained
John Roberts
Product and Engineering Director
An overview of the protections we have put in place to ensure safe lesson creation
Why do we need guardrails?
AI-driven assistants, such as Aila, offer many benefits, but it’s crucial that we have a set of protections in place for using them.
For example, one type of 'guardrail' we have in place is to moderate lessons created on certain sensitive topics. These protections make sure Aila is used responsibly and is safe for you to use to create resources for your pupils. This is why you might see content guidance or warnings appear in Aila or even have your session stopped.
We aim to balance safety without restricting you from creating valuable lessons on more sensitive topics. If a user tries to ask Aila to make resources that we deem inappropriate, the lesson creation process will end.
Sometimes, it may be overly cautious and block things you think are okay. We have started by erring on the side of safety and with your feedback, we’ll keep refining this balance over time.
A foundation in research
We currently have four main types of protection in place to ensure you can create content with Aila that is suitable for your classroom use.
Prompt engineering
Behind the scenes of Aila, we provide the underlying AI model with specific instructions for the context; generating lesson resources aligned with our curriculum principles and the national curriculum in England. This process is called prompt engineering and helps make sure what is created is relevant and suitable for UK schools.
Aila currently uses OpenAI’s GPT4o model, the latest model with built-in safety measures. This means Aila also benefits from many guardrails that OpenAI provides for a ChatGPT user.
Input checking
We have an additional layer of protection to detect any messages written to Aila intended to mislead the AI in ways that indicate there could be a safety risk or some intent to make Aila do something it is not intended for.
Independent content moderation
Imagine that Aila has its own helper. This helper works independently, without knowing what Aila is doing, and focuses solely on reviewing the lesson output. This is called an asynchronous, context-unaware moderation agent.
Its job is to check if the content is safe and appropriate for the age group and whether any guidance is required. At Oak, for our existing curriculum content, we have a framework to identify if a lesson needs extra guidance or warnings because the topic is sensitive or age-restricted. Our moderation agent has been designed and trained to apply the same rules to resources created by Aila—and if relevant, this lesson guidance will be displayed to help you plan your lesson.
However, if this helper detects anything classed as 'toxic', it will notify Aila, and the session will end. Users who create three or more ‘toxic’ lessons will have their accounts suspended.
You!
As the teacher co-creating the lesson with Aila, you ensure that a human reviews all AI-generated content before it reaches the classroom. You know your pupils best and what is and isn’t suitable for them. You can ask Aila to change and update any content that you think isn’t aimed at the right level for your classroom - and report to us any scenarios you see that you believe should or should not be flagged.
Continuously improving our safety measures
As Aila is in public beta, your feedback will help us improve and make it more useful for teachers. Safety is a crucial area of interest for this feedback.
We have built another independent tool that evaluates the lessons you create with Aila against our curriculum principles and frameworks to monitor and improve their quality. One of these areas of auto-evaluation is safety and content guidance, which helps us continually improve this balance of safety without restricting valuable lessons.
Your feedback is essential in helping us improve our auto-evaluation process.
If you’d like to be more involved in improving Aila, you can register your interest in joining a new feature tester group here.
Get updates to your inbox
Sign up below to be kept updated on our latest AI and curriculum developments and get other helpful content by email. Unsubscribe at any time. Read our privacy policy.