A couple of weeks ago the BBC ran an experiment on BBC Four with AI. They used AI to help do two things; browse their archive to select programmes that would appeal to BBC Four viewers and more interestingly try to make a programme. We wrote about the possibilities this technology can offer in this blog post here. This blog will take a look at how AI performed in this experiment and whether the robots are ready to take over!
Across the two scheduled evenings, there were some great programmes about AI and robotics in general. Some of the archive material from the Fifties and Sixties trying to predict what the future would look like with robots everywhere were, with hindsight, quite comical. However, they were interesting to compare with where we actually are today.
Whilst we don’t have robots doing all the housework, we do have robotic vacuum cleaners and lawn mowers, voice-activated home automation and intelligent white goods that will help us cook and order food when we run out. So maybe we’re not so far out as we thought?
More interestingly from a TV perspective though, was the idea of using AI not only to choose programmes for the schedule but also to try and create a programme.
The choosing of completed programmes for the schedule is not so far-fetched and is really no more than the industry has today with content discovery and recommendation services. By knowing characteristics of the audience and what has proven popular in the past, the algorithm chose some 150 programmes from the 270,000 in the BBC’s digital archive that were a good match for BBC4 and scheduler’s made the final choices.
These same 150 programmes were then used by the BBC R&D team’s AI platform to create a series of short sequences. They used a number of techniques, including image recognition, subtitle analysis using natural language processing techniques, video dynamism and finally combining all the above techniques to create a short program of up to 15 minutes that had some sort of narrative. Their blog post about it is here.
Unfortunately, it mainly resulted in sequences that were not really watchable. Despite the AI melding the 150 programmes into over 150,000 chunks of between 15 seconds and 2 minutes with the singular techniques, it didn’t produce a meaningful narrative. In some cases, the link between clips could be ascertained but frequently it could not. The final sequence using all 3 techniques was better, but still lacked the finesse of an expert production team and editor.
However, whilst watchable TV was not produced by the AI system, the experiment cannot be dismissed as a failure. What it clearly shows is that we have a long way to go before AI is likely to replace humans in this process. It highlights the subtlety and nuance that journalists, producers, researchers, cameramen, sound recordists and editors have in producing a programme that engages and educates an audience.
We have much to teach an AI system before it starts winning a BAFTA!