After the Brexit vote in the U.K. and President-elect Donald Trump’s win at the ballot box, the results of conventional polling seem increasingly further from the actual outcome.

According to the founder of an artificial intelligence system, who said Donald Trump would win the presidential election, polling can be described as a “statistics of convenience” subject to many human biases.

“There is obviously a big difference between polling and how machines could look at a wide base of data all across the internet through the internet engagement of people because of various reasons that might not show in polling,” Sanjiv Rai, founder of Genic.AI, an artificial intelligence system, told CNBC on Friday.

Most opinion polls wrongly predicted that Hillary Clinton would become the next U.S. president. They also wrongly announced that British voters would not decide to leave the EU or that former Prime Minister David Cameron would need a coalition partner when in the end he obtained a parliamentary majority.

Rai’s analysis looked at 20 million data points, from engagement in platforms such as Google, Facebook, Twitter and YouTube in the U.S. and projected that Trump would win the presidential election.

“A wider data set of millions of users, in our case over 20 million, was a much better way to kind of show where the trends where and then it could into very minor details,” Rai told CNBC.

The difference between the two systems seems to be on the level of human intervention. Whereas in polling there could be human biases from those answering it and asking it, analyzing online engagement is “neutral,” Rai said.