Seeing the future: how AI predicts elections and horse races
The US election result came as a surprise to many – the President-elect included – but not MogIA. The predictive artificial intelligence system out-figured most professional pollsters, calling Donald Trump’s electoral success at the beginning of November.
A rival system called UNU picked Hillary Clinton to win the popular vote, which she duly did, and it correctly predicted another notoriously difficult to call American race – the Kentucky Derby. Both systems work by listening to people.
MogIA scans public social networks, while UNU surveys respondents, taking that human opinion and “amplifying” it into a system called “swarm intelligence”. AI-using neural networks seek to mimic the brain’s network of cells, explains Louis Rosenberg, founder of UNU-creator Unanimous AI, while “swarms” gather intelligence via a network of people. “It’s an emergent intelligence, moderated by computer software, but leveraging the wisdom, insights, and intuition of human agents,” he said. “And it turns out that these artificially intelligent ‘human swarms’ are very smart. They outperform individuals. They outperform experts. And they outperform simple methods of tapping the intelligence of groups such as polls and markets.”
It’s more than averaging human selections – as existing polls do. With the Kentucky Derby, UNU built a swarm of 20 people. If the system had simply gone with the most popular picks, it would have picked only one of the top four finishers. “We amplify the intelligence of individuals into a unified emergent intellect that exceeds natural capacity,” he said.
Dr Chris Brauer, director of innovation at Goldsmiths, University of London, explained that AI is essentially maths in action. “Weather forecasts improved dramatically with investment in weather stations, satellites, data, and processing power. “Similarly, forecasting will improve in any complex socio-technical system such as political elections when the system continually improves and looks out for new data sources and forms – and has the power to process and analyse these inputs in meaningful ways. Next to this, polling looks like a man standing in front of a map pointing at clouds with a stick.”
Humans could perform better, but our hopes and assumptions get in the way, he noted. “One benefit of these emerging AI prediction platforms is their ability to burst through filter bubbles,” said Brauer. “This may seem paradoxical, since many are blaming smart algorithms for creating filter bubbles in social media. But when it comes to predictions, too often our models accept and embed flawed assumptions. AI systems are better than human beings at dealing with what we’d normally term unknowns.”
Just how smart are these predictive systems? Rosenberg said his swarm intelligence outperforms between 90% and 99% of the individuals in the swarm. “We also compare swarms to standard polls, and outperform them in almost all situations,” he said, pointing out that most polls have more participants than a swarm.
But there are limitations. “Our system performs best when asked complex problems that have many possible outcomes,” Rosenberg said. “For simple questions that have only two possible outcomes, our performance is good, but becomes more similar to a poll.” And there are downsides to looking to machines for answers without understanding their limitations. “These predictive AI systems aren’t foolproof or objective by design,” Brauer said. “In many ways, the frailties of algorithmic bias are more threatening than human biases since there’s a presumption of innocence.”
Trust in human authorities – such as governments and journalists – has long been falling, but that doesn’t mean machine-learning algorithms are entirely benign or lack bias. “When we encourage learning systems to adapt autonomously in the wilds of data and human behaviour, we invite machine discretion and ethics,” he said. “One day, a machine is predicting the winner of an election; the next day, that victor is turning to the machine for advice.”
Aside from picking horse races and beating pollsters, UNU and swarm intelligence have been tapped for market research and consumer forecasting, such as reactions to movie trailers, as well as examining financial markets and medical diagnoses. Will such AI put the bookies out of business? “I doubt it,” said Rosenberg. “If swarm intelligence becomes widespread, it will simply change the odds, since it means the gambling public has become smarter.