Whether you’ve been left reeling from or celebrating the result of the US election, one thing is for certain – President-elect Donald Trump remained the underdog in the eyes of pollsters until the bitter end.
But unlike experts and political commentators, Havas Cognitive’s Artificial Intelligence project, Eagle AI, was predicting a Trump win from the off.
Data Doesn’t Lie
Commissioned by ITV News to create an AI bot, Havas worked with IBM Watson (an AI created by IBM and licensed by several companies including Havas), to analyse and understand human behaviour – and furthermore to predict human behaviour.
“Usually we do it with a marketing hat on,” explains Faye Raincock, European communications, Havas. “[We] try and predict people’s buying decisions and behaviour when it comes to relationships with brands; but what we’ve developed here is unique in that it’s the first time it’s been used to try and analyse voting behaviour.”
Using billions of data sources, Eagle AI analysed any social media posts and comments related to the election; but it also listened to all of the debates, speeches across the election campaign, and campaign videos. On top of this, it also monitored any press coverage and its virality.
— Brian Fraser (@bfraser747) November 15, 2016
Attitudes on social media were more positive than pollster predictions.
“What was particularly clever was that it was trained to understand emotion, sentiment and tone – and that was the first time that had ever been done relating to voting intention,” Raincock says. “Instead of traditional polling, it dug beneath the surface into how people actually felt, and tried to predict behaviour. ITV never asked us to predict the outcome of the election, largely because they have an existing relationship with traditional pollsters in the US. But as part of our job, to understand what mattered to people, we began looking at voting intent. We felt that was an important parameter to understand behaviour. And we found that the AI was predicting a Trump victory quite soon into its analysis.”
Whilst a great many of us may have never believed Trump was capable of succeeding in his presidential bid, the unbiased analysis of the AI was monitoring countless attitudinal responses to the campaigns. Raincock explains that “when all the human beings, all the experts, all the pollsters were saying no way, it won’t happen – the AI was predicting the opposite.”
How Did Pollsters Get It So Wrong?
What sets the AI apart from traditional pollsters is the removal of the human element – and therefore the removal of bias. The AI has no preconceptions or prejudgements, and that creates a distinction. When traditional pollsters do phone polling, they ask specific questions (eg. how do you plan to vote?), whereas the AI delved much deeper, analysing not what people were willing to admit to pollsters, but rather every post a person put out, and their emotional connection to that.
“It stops being about how you will vote, and it starts being about how a certain candidate or election issue makes you feel. It extrapolates that into what it expects that feeling to predict. It’s a more nuanced understanding than a straightforward answer to a traditional poll,” explains Raincock. Essentially, by using machines rather than human analysis, you completely leave out the innate human bias.
“With pollsters and political commentators, there was a part of them that just couldn’t see how human beings could vote for someone who was saying and doing the things that Donald Trump was saying and doing. And so there was a bias that was connected to their already preconceived ideas as to what was acceptable and what might be deemed to be acceptable by the voting public. And we think that’s ultimately what it comes down to – it’s precisely because it isn’t a human being that it can predict human beings better.”
Executive producer at ITV, Alex Chandler, said, “ITV News was joined by a raft of experts to shed more light on the complexities of this election. Eagle AI’s job was to show for the first time on live TV, the impact of a vitriolic presidential race on the mood of the nation. Thanks to Havas, we were able to deep dive into millions of data points, in a way no programme has been able to do before.”
What Does This Mean For Marketers?
Havas believes that the creation of such a capable AI is the beginning of a new wave of consumer analysis. “When we looked at the analysis after the effect, we found the AI predicted 41/50 states and 4/5 swing states. There’s always going to be margin for error in these things but we’re seeing applications for marketing,” explains Raincock.
“Our job now is to help people understand cognitive systems better. They start out thinking you’re talking about unstructured, big data challenges and even the dark web – all quite difficult concepts to grasp. In fact, it’s a lot simpler than that. It’s just taking a massive sample, bigger than would be humanly possible, and interpreting a brand’s position or power in a more meaningful way.”
To give an impression of the scale of data the AI can process – it would have taken 2000 researchers 30 years to analyse what Havas analysed in around a month for the US election. With this scale of understanding applied to brand penetration and buying decisions, the application for marketers can be astronomical.
“What we’re finding is that AI and cognitive can analyse on a truly massive scale and then predict behaviour more accurately than ever before.”
Lisa De Bonis, executive digital director, Havas, will be going into detail on how Havas Cognitive achieved these insights at our Digital Marketing Conference on 30 November. Register for your place now!