They were the 90 minutes of television that set America on fire. As Donald Trump and Hillary Clinton stepped up to the podium for the first presidential debate of the 2016 election, the battle was already raging on Twitter.
Subscribe now for unlimited access.
$0/
(min cost $0)
or signup to continue reading
But not all of those users joining in the discussion were human.
An army of so-called Twitter bots, or automated accounts, were also at work, firing off hashtags and worming into feeds, in a wave of "information warfare" that has since come under investigation by a US probe into Russian interference in the presidential race.
Now, on the other side of the world, Canberra computer scientists have written an algorithm to track the influence of these bots on the election - and the results have surprised even them.
Analysing the more than 6.4 million tweets that circulated during those 90 minutes, researchers at the Australian National University found bots were on average about two and a half times more influential than people. While there was a lot less of them than expected, with only about 4.8 per cent of users identified as clear bots, they were also more successful at attracting exposure via retweets.
Researcher Timothy Graham explained that the average election bot is a lot more politically engaged than their human counterparts, and more likely to be pro-Republican.
"People think bots are just these faceless lines of code sitting on a server in a building in Saint Petersburg, but that's only a small fraction of the bots out there," Dr Graham said.
"We're also seeing more hybrid automation with multiple people behind one account, where the lines [between human and bot] are fuzzier. They can do superhuman things, they're always active... they're like soldiers."
Dr Graham and his team spent months crunching the numbers to determine which of the 1.5 million user accounts active during the debate were actually human. That was the easy part.
As Twitter does not provide data on "retweet chains", or the cascading exposure of a tweet over its life, a new model had to be devised to track a bot's influence.
The algorithim, which can reliably map user influence, took the lab just 12 hours to commute when set loose on the debate data.
Hijacking trending hashtags was a common tactic deployed by bots to spread their tweets and often involved misinformation, such as the "Crooked Hillary" hashtag, Dr Graham said.
Bots also appeared to target their content towards real-life influencers, especially Republicans, in the hopes they would be shared with larger followings.
During the election, a number of conservative accounts racked up thousands of followers, many of them Republican senators, before being unmasked as bots and shut down.
A tweet from a high-profile bot called Ten_Gop, which claimed to represent the Tennessee Republican Party, was even shared by former Donald Trump advisor General Michael Flynn.
"Would Donald Trump have won the election if we didn't have social media? This research is quite concerning," Dr Graham said.
"We have covert, information warfare happening in the very spaces people are choosing to spend an increasing amount of their time."
With Twitter little more than a decade old, Dr Graham said researchers had an ethical imperative to study the growing phenomenon and its impact on democracy.
"It's not a conspiracy theory. We know that now."
In January, Twitter admitted more than 50,000 Russia-linked accounts used the platform to post automated material about the 2016 US election.
Australian elections were not free from robots either, Dr Graham warned.
Research from the University of Canberra and UNSW found Russian trolls targeted Australian voters, particularly right-wing voters, on Twitter at the last federal election using hashtags including #auspol and #MH17.
The ANU's code measuring "cascade influence" on Twitter is available open source and online at: https://github.com/computationalmedia/cascade-influence.
