AI at the BBC: the promise and the challenges
The potential for streamlining services is huge, but so are the hurdles says lead strategist Noriko Matsuoka
The potential for AI and machine learning in streamlining processes at the BBC and creating the basis for new services is matched only by the challenge of implementing automation in such a complex organisation, said Noriko Matsuoka, lead strategist - emerging experiences product, technology strategy and architecture at the state broadcaster.
The BBC has legacy systems like no other, and its public service remit and taxpayer-funded status means it faces more constraints than most.
Matsuoka is part of a new team at the broadcaster with a remit ‘to transform BBC technology for future'. This covers areas from post-production, studios, connected data, understanding the audience, emerging technology, next-gen audio, conversational assistants, conversational AI experience and children's AI experience.
The remit is largely exploratory, seeing what's possible and finding out would be needed to take it forward. In the newsroom, for example, fully unmanned automated studios would help presenters announce breaking news without having to assemble a camera crew and sound engineers.
"There's a significant need for something like that so investing the time and manpower might well be justified," she told delegates at the Connected World Summit in London. Other possible interventions include "replacing talent with AI", which might be a tempting prospect for the cash-strapped broadcaster but is unlikely to get the green light. "It's just putting stuff out there to see what could work."
Matsuoka herself works mostly at the customer experience side of things. "What's coming in, how voice is maturing moving from NLP to voice prediction and voice patterns."
Get that right, and the BBC wouldn't have to deploy nearly so much software in order to interact with the general public. The biggest challenge, though, is the data.
AI is terribly interesting but terrifically complicated to do
"AI is terribly interesting but terrifically complicated to do. You need the right algorithms, and you need consistent data. That's a massive problem. The BBC has 100 years of broadcast history. Records are not consistent, with the same thing reported in different ways."
What's more, there is "software stacked on software stacked on software", with multiple variants of each customer-facing application - there are 43 different versions of iPlayers, for example, covering all the platforms and devices the BBC is obliged to serve. Given this complexity, any intervention must be carefully thought through.
Ther are 43 different versions of iPlayer covering all the platfoms the BBC is obliged to serve
"You have to find what is genuinely valuable. You need to train algorithms, make sure data is consistent, understand your failures, inbuilt biases, duplications - there are so many caveats."
So is it all too onerous?
Not at all, she said. For example, she is involved in building AI services to ensure children's areas of BBC sites are kept safe without restrictive password protection "which kids can easily get around", and without "being too worthy". In this, it hopes to avoid YouTube's recent experiences with ‘evil' Peppa Pig videos and other content flagged as safe for kids that was anything but.
But the power of AI means you need to think through the repercussions carefully before letting it loose. Matsuoka welcomes GDPR and similar legislation as it defines boundaries, but says this can create challenges of its own when working with commercial providers such as IBM Watson and Google. The commercial use of data by these players can "muddy the waters" when it comes to building trust she said. It is therefore necessary to keep them at arm's length - what she describes as a "push-me-pull-you" relationship. This means that all training data that goes into the BBC's models must stay with the BBC and not be shared with its implementation partners.
The potential of AI is huge, she said, but so is the amount of work and investment you need to make it work.