We’re always being told that our jobs are on brink of being automated, but creativity is usually considered to be safe for longer than most skills. Well, maybe we don’t have as long as we think. Watching the new Lexus ad, which was written by an artificial intelligence and directed by Oscar-winner Kevin Macdonald, some agency creatives might have cause to worry.
That’s probably an overreaction. Although, on the face of it, the ad looks like a car commercial, telling the story of one of Lexus’ ‘Takumi Master’ designers setting his creation free, on closer inspection it’s a little odd. It’s more than a little spooky that a machine decided to treat the car as a character in itself and the edit doesn’t follow the conventions of film storytelling that viewers are used to watching.
Intrigued by the ‘man and machine’ creative process that produced these results, LBB’s Alex Reeves spoke to The&Partnership creative partner Dave Bedwood and head of account management Benedict Pringle, as well as co-founders of Visual Voice Alex Newland and Will Nutbrown, who built the AI.
LBB> When did you realise that this was a possibility, that an AI could write a good ad?
Benedict Pringle> We’d seen examples of artificial intelligence adapting existing footage to create new stories, so we knew it was possible to train an AI to produce creative content. However, the idea that we could train an AI to create an original, high-quality piece of storytelling from scratch was a little bit of a leap of faith. The first step was to find the sort of data to feed into the AI that would result in an interesting story.
LBB> How did you 'train' the AI? What was the process?
Alex Newland and Will Nutbrown> The first stage in the process was to ingest the award-winning adverts into a format suitable for use as training data. This entailed using a number of AI-based visual analysis tools including IBM Watson to extract details such as the objects, actions, locations, activities, words and sounds from the award-winning ads. We were then able to use this data as training material for the AI to learn not only the elements of award-winning car adverts – but also the sequence in which they occur and also the relationship to each other.
LBB> What was the 'brief' you gave to the AI?
Dave Bedwood> The whole idea behind this advert was to mimic the product truth of the car - it has ‘intuitive’ technology aiding the driver, it is ‘man and machine' working together for a better driving experience. So the ad is also 'man and machine' working together. That was the core of the brief, and it ran through the whole process, and was the guiding principle as we worked with the AI team. The only thing we didn’t know at the start was just how much the ‘machine’ would be able to do.
LBB> How did you feel at the moment you saw the script come out?
Dave> I was shocked. A machine telling the story of a machine coming to life does give you pause for thought. But once that initial shock wore off, I was then suspicious. I honestly thought Will or Alex were doing a ‘Wizard of Oz’ number on us.
I thought we’d just get fragments, small details, oddities, that we would then turn into a full script, almost like a creative writing exercise. That, for me, was going to be the ‘man and machine’ element. But we got an entire script, so it turned out that the ‘man’ wasn’t going to be me after all, it would rest solely on Kevin [Macdonald, the director] to bring the AI words to life. Which turned into a simpler and a much more interesting story.
LBB> What insights from the training material were clear when it started writing?
Alex & Will> The data from the MindX study into the nature of intuition within humans [from applied scientists at the University of New South Wales - the basis of the AI’s training] strongly indicated that one of the most important aspects for the advert was for it to create an emotional experience. We see this manifested in various aspects of the output, where the AI sees opportunities to heighten the emotional experience for the viewer. We see what a human writer might consider to be overly dramatic shows of emotion such as the crying Takumi and dramatic tension when watching the events unfold on TV. We also, however, see more subtle elements which the AI sees as increasing the emotionality of the output, such as the seemingly arbitrary inclusion of a child.
One of the other major insights was the extent to which the AI had determined that the advert need not contain any particularly significant amount of driving. There were then other quite specific details about the display of the car, such as the avoidance of showing anyone driving it, and that the focus should be on demonstrating the benefit of a single feature of the car, rather than attempting to recount a list of product features and benefits.
LBB> What were the big technical challenges in the process of getting it to write to the standard you needed it to?
Alex & Will> The first challenge we were aware of from the start of the process was the risk of creating a mash-up of previous award-winning adverts. The hope was that these award-winners should be used to train the AI in order to create something new, not just to facilitate the creation of a greatest hits sequence of common elements. In order to mitigate against this, it was necessary to extract data from the adverts down to a more granular level than might have otherwise been the case. By separating out elements of the adverts into many parallel parameters such as emotion, activity and location, for example, it was much less likely that we would see the recreation of exact scenes from previous work. The AI found the reoccurring aspects of award-winning ads and then composited these elements together as it best saw fit, rather than drawing existing scenes together.
The other challenge, which was not readily apparent, was the extent to which awards are dominated by a small number of mass-market automotive brands. In some of the early output from the AI, we saw strong traits which could be linked directly back to those specific brands and adverts, which didn’t feel very Lexus. After discussions with the project team, it was decided that the training material for the AI needed to be broader and it was at this stage that a selection of award winning non-automotive luxury adverts were added into the process.
LBB> What quirks did the AI turn up that you might not have got from a human copywriter?
Dave> All of it. I don’t think any of it would have come from a human copywriter. But this wasn't about the AI writing an amazing ad in its own right. It is only interesting if you preface the ad with ‘This was written by AI’. It’s an experiment about seeing the truth of what it can come up with and I think that truth is pretty interesting. The result is more ‘normal’ than I would have guessed - which is a little bit unnerving.
LBB> How did Kevin respond to the script and what were the key decisions for him bringing it to the screen?
Dave> He went from being only on board to do the behind-the-scenes doc, to wanting to direct the actual ad. I think everyone at the beginning was sceptical, and also just really unsure what the process would produce. How do you get a director on board before there is even a script? It’s a real leap of faith into an area which can sound, to the layperson, a bit dubious – smoke and mirrors. At the beginning we presumed the final ad would probably be a 60-second cut of the process of making the ad. Our guess was the final ad would end up being really bonkers or unusable. But the moment the script came out Kevin was really interested. It was odd, but not that odd; it was trying to be emotional, but not quite landing it. It was a computer second-guessing human emotional triggers, but not having the subtlety to achieve that yet. In terms of bringing it to the screen, Kevin stayed true to the script, with director nudges here and there to help the story. Even in the edit, he stayed true to it. No human would have edited it in this way – it just doesn’t conform to how we would tell a story. And that’s what makes it interesting.
LBB> Where will you take this ad-writing AI next? Is it going to continue growing and progressing?
Benedict> This artificial intelligence was built to be successful at a very tightly defined task: write a script for a technologically advanced Lexus vehicle that would appeal to intuitive people. It is not a general AI which can be deployed against any brief. Having said that, as well as the script, the AI did output six pages of ‘success criteria for Lexus advertising’, which has already informed (and will continue to inform) our creative output for the brand.