In order to perform its tasks, Robin needs to use human-like judgement. To do so, we use a number of different technologies that get lumped under the term "AI". The important thing to know is that, like people, sometimes Robin will get things wrong.

We aim to minimise the effect of this, but we can't completely eradicate it. This is unlike the previous generation of computer software which, at least in theory, always got things right.

Accuracy is very important to the usefulness of Robin, so we track and measure it carefully - both by manually checking the reports we produce, and by building up banks of sample data which we can use to test whether changes have improved the overall accuracy of Robin.

Those improvements can come from our own software and prompting, or from improvements to the underlying AI model that we're using.

To give you a sense, in our latest round of testing we judged that the Review Dates in the policy audit are 93% accurate. That accuracy percentage has grown significantly since we started.

The accuracy also varies based on the clarity of the information on the school website. Where schools use a clear template for their policies, with explicitly stated review dates, Robin is near-perfect at extracting the information. But Robin is more likely to struggle where the sources themselves are unclear.

You can help us improve the accuracy of Robin by giving us feedback using the 👍 👎 buttons in the Robin app. These help us investigate issues and improve our tests.