Picasso Labs knows Donald Trump and Hillary Clinton’s best sides
At PromaxBDA’s “The New Normal” conference, former Googler and Picasso Labs CEO Anastasia Leng is demonstrating the company’s USP with the potential man of the hour: Donald Trump. There are three pictures used by the Republican presidential nominee’s campaign on screen – the first is his daughter Ivanka and her baby Theo; the second is Trump surrounded by armed police officers; and the third is a quote attacking Hillary Clinton.
Audience members are invited to vote with a show of hands on which is the most effective – and I duly raise my hand for the Hillary Clinton attack.
But no, I’m wrong. Picasso Labs’ AI has done the number-crunching and found that it’s picture #2 that resonates with his audience. “Any time Trump was seen next to someone in uniform, the engagement rate of his users went up,” Leng explains. “The interesting thing here – or terrifying, depending on how you look at it – is that our system further isolated that within imagery of men in uniform with Trump, those images that contained a gun further enhanced performance by another 6%.”
“Any time Trump was seen next to someone in uniform, the engagement rate of his users went up”
And that, in a nutshell, is what Picasso Labs does. It’s an AI that examines images for shared characteristics, and then plots those against how well an image performs for brands. “I see us as the Google Analytics for visual decisions and visual content,” Leng tells me when I sit down with her to discuss her talk. “When people talk about creative, oftentimes they’re guessing. And they’re using a lot of their own opinions and confirmation bias about whether or not they like the way something looks physically to decide whether it’s something their users will engage with.
“Marketers are realising – 65% of them agree, a big shift on the year before – that visual is important to how their brand is being communicated. But what they don’t know is why certain images are cutting through while others are falling by the wayside.”
Primarily, this is on social media, but it can work anywhere where there’s an image and data to back up its relative success. “If there’s some objective way of assessing how an image is done, we can do it,” Leng tells me, citing an offline example where her company examined a few years’ worth of newspaper front pages cross-referenced against sales information to give a client an idea of what sells on the newsstand. “Inevitably, you do find statistically significant trends and patterns around things that might prompt others to pick up the newspaper,” she explains. “Images of the royals always sell newspapers,” she quips.
It probably doesn’t take an AI to tell you our tabloids are obsessed with royalty, but back across the pond in her adopted home of America (she’s Russian by birth), Leng has more examples from the US election. This time it’s Hillary Clinton getting a free demonstration of Picasso Labs’ AI – and specifically, which running mate works the best for her. The first picture is her alongside Michelle Obama; the second with her husband, Bill Clinton; and finally it’s her with the incumbent Barack Obama.
I raise my hand for Barack Obama, based on the idea that perceived presidential legitimacy can pass from one to the other via osmosis. Once again, I’m wrong: it’s Michelle Obama that gives Hillary the biggest boost. Bill Clinton is the weakest of the three, though, suggesting that Trump’s persistent focus on his opponent’s husband may be a shrewd, if somewhat blunt, tactic.
So I’m wrong on both presidential examples. Do the AI’s findings ever surprise Leng? “I’m surprised all the time. We think we know, but ultimately we’re just guessing.” The AI, on the other hand, has no preconceptions. It’s simply “codifying images and finding patterns”, as Leng puts it. Sometimes, that pulls up the unexpected, as with Trump.
“When I first saw that his best-performing imagery was him next to men in uniform… it’s not even something you’d think to look for, it’s just something our system picked up on. And what does that tell you about what his audience wants? They want a dictator, right? – Iit’s very autocratic-type imagery.”
This pattern, by the way, doesn’t work for Clinton. “If she’s next to a cop or someone in law enforcement, it doesn’t make a difference at all.”
Picasso Labs has now analysed “millions and millions” of images, and is doing very nicely for itself, which makes its origin story all the more surprising. It was, to use Leng’s own words, born out of “pure serendipity and sheer unwillingness to admit that I had failed”.
“Picasso Labs was, to use Leng’s own words, born out of “pure serendipity and sheer unwillingness to admit that I had failed”
At the time, Leng – an ex-Googler – was struggling with her first ecommerce business. “We were trying to do anything we could to get revenue up,” Leng explains. “We started noticing that all our revenue was coming from image-based marketing networks like Instagram, Pinterest and Facebook. We couldn’t understand why one image was doing well while another that looked quite similar to us didn’t.”
So they started developing the technology that would one day become Picasso Labs, purely to give the company a competitive advantage. It turned out investors were less interested in the ecommerce site than the technology behind it. “With $30,000 left in our bank account for a team of eight very expensive engineers, a number of investors decide to dig under the hood,” she explains. “They essentially pull us aside and say ‘hey we’ll give you money, but only if you start a second company.’”
It’s pretty clear that this is a far better fit for Leng and her team than a straightforward ecommerce site, and those analytical data-driven habits drilled in at Google die hard – even in how Leng chooses to present herself now. “I sat down after every meeting to jot down what I’d worn, and observed investor reaction and how feedback changed as a result of how I presented myself.” The glasses she wears to this day is an example of the unofficial feedback she managed to gauge.
“Always wear glasses; always wear trousers; usually put your hair up,” she summarises from her brief experiment. “With investors, a lot of them make decisions based on pattern recognition, and a lot of that dataset is trained towards men.” Not statistically significant, she concedes, but an insight into how her mindset helped her found the company.
The advice of Picasso Labs AI is, of course, not always helpful to a brand. Another example given in her presentation was for the fashion industry, and it found that runway imagery was the least effective by a long way – considerably cheaper street-fashion shots were far more impactful. This is something high-fashion brands are reluctant to give up, as “vital to the brand,” but even in those contexts, Leng contends, Picasso Labs can help by searching for the best way of presenting this essential but ineffective content. “Does the way the model is shot help? Having multiple people on runway versus one? Light colours versus dark? How do we improve this, because we know it’s important to you.”
It’s not that humans are bad at this necessarily, but there are patterns they can’t see that an AI can. Human judgement is fickle and inconsistent. “Sometimes they’re right, sometimes they’re not. It doesn’t mean you’re bad at your job, it’s just that human psychology is a very complicated thing. We add a little bit of science to something which has predominantly been an art.
“What we’re trying to do is get you to a point where next time you’re in a meeting and someone says ‘well, I just don’t like it,’ you can say ‘well, I don’t care – because I have the data to show other users do’.” Perhaps by the 2020 election, Republican and Democratic campaign managers will have had similarly frank conversations with their candidates.