Research & Development

Posted by Galen Reich on , last updated

At BBC Research & Development, we‘re keen on Making AI more Understandable. It can be helpful to have intuitions about the decisions that artificial intelligence (AI) systems make, and games are a perfect way to build these. Recently we’ve developed two prototype games to help children (and adults) learn more about AI.

Most of us use AI every day, often without realising it, and AI systems affect how we interact with and experience the world. It’s usually pretty difficult to tell what criteria an AI system uses to make decisions. But when things go wrong with these systems, it becomes difficult to untangle what’s gone awry and how to fix it. With an increasing use of AI in our day-to-day lives, these systems will significantly affect the next generation. It’s vital for those children to grow up as confident and capable users of this technology.

But AI uses a lot of complicated maths - how can we make this tricky technology accessible for young people?

Rather than trying to explain how AI works, we asked if we could help children build an understanding of the decisions that AI systems make. We explored different ideas about the form that these explainers could take. We considered short-form videos, interactive websites, and even installation artwork! We landed on building games as a practical way to incorporate interactive elements and encourage creative play. Creating a sandbox that lets children interact with AI, allows them to independently develop intuitions about AI systems.

A photographic rendering of a young houseplant against a neutral background, seen through a refractive glass grid and overlaid with an single neuron from a neural network diagram.

Photographic rendering of a young houseplant against a neutral background, seen through a refractive glass grid and overlaid with an single neuron from a neural network diagram.

Image by Alan Warburton / © BBC / Better Images of AI on Flickr, cc licence.

There are a few characteristics that lots of AI systems share, and we used these to set the learning goals for our games:

  • AIs are not intuitive and often can’t explain the logic behind their decisions
  • AIs exhibit bias, replicating prejudices in their training
  • AIs are not creative; they are programs that operate in a restricted way

Besides the pedagogical challenges, there are significant technical challenges to developing safe AI systems for children to use. There are many examples of AI systems producing output inappropriate for children to encounter. We decided that using an AI system to generate textual content would be too risky to include in a game aimed at children. Instead, we chose an AI system that gives the player a score. This approach still enables meaningful interactions but is less open-ended and much safer.

With these restrictions in mind, the team developed ideas for several prototype games, and we built two of these.

AI Art School

A laptop playing the AI Art School game.

AI Art School is a single-player game in which three AI assessors challenge players to draw particular shapes. Players receive scores based on the shapes that the AIs recognise. This game builds an understanding of how AIs view the world very differently from humans. They identify features rather than looking for a complete impression of an object or shape. It demonstrates that AIs can make choices that seem illogical from a human perspective. This game runs a lightweight AI model in the web browser using Tensorflow.js.

Missing Link

A laptop playing the Missing Link game.

Missing Link is a multiplayer game where players describe words and try to outwit an AI player for points. By writing short prompts, players try to communicate which of six words they have picked. Players try to give context for their fellow humans to guess correctly while misleading the AI player to choose incorrectly. This game highlights the differences between AI training and human experience. Players learn about the types of associations that AIs make and use this to win the game. This game uses an instance of Sentence-BERT for evaluating the similarity between words and prompts.

Why this matters

With the publication of the AI Council’s AI Roadmap and the National AI Strategy by the UK Government’s Office for Artificial Intelligence in 2021, there is a concerted effort to build the UK’s capability to use AI. The National AI Strategy plans to encourage “children to explore and be amazed by the potential of AI”, and the AI Roadmap aims to “ensure that every child leaves school with a basic sense of how AI works”. To do this, there will need to be a significant increase in the resources available to help children understand AI.

The BBC is in a great position to develop AI learning resources for children, and these prototype games are one small step towards achieving this ambitious goal.

Tweet This - Share on Facebook

BBC R&D - Better images of AI

BBC R&D - Making AI more undestandable

IBC TV - Interview with BBC R&D's Craig Wright of the AI Production project

IBC TV - Tech Talks: AI in Production

IBC - AI in Production: Video Analysis and Machine Learning for Expanded Live Events Coverage

IBC - Interview with BBC R&D’s 'AI in Production' team

BBC R&D - AI and the Archive - the making of Made by Machine

BBC R&D - The Business of Bots

BBC R&D - Content Analysis Toolkit

BBC R&D - Data Science Research Partnership

BBC Academy - What does AI mean for the BBC?

BBC Academy - AI at the BBC: Hi-tech hopes and sci-fi fears