【Underwater Love】

Twitter bots have Underwater Loveseemingly become the weapon of choice for fervent political supporters this election cycle. But they're getting harder to spot amid the sea of real-life users.

Bots, powered by automated software, have propped up candidates on both sides of the aisle this election cycle -- more so for Donald Trump, a new study has found.

But it's no longer enough to be suspicious of the egg avatars. Sophisticated bots mimic human-like behaviors by using language algorithms to chat with other users, comment on posts and answer questions.

SEE ALSO: This Twitter bot will show you the people Donald Trump retweets

Academics have tried to tease out the humans from the bots with varying degrees of success. But a study published this week by the University of Oxford purports to have pulled back the curtain, finding that Twitter love for Trump after the first presidential debate was driven mostly by bots.

Trump's Twitter debate lead was 'swelled by bots' find @pnhoward & @polbots computation propaganda team https://t.co/LjiHAVTK4x pic.twitter.com/4aUPITUzaA

— OII (@oiioxford) October 18, 2016

Nearly five times as many tweets from automated accounts supported Trump compared to his Democratic rival Hillary Clinton.

How they found the bots

A Twitter account was determined to be a bot if it tweeted more than 50 times a day for four days, which would be 200 tweets in that relatively short time.

The study suggests that around 580,000 of the roughly 1.8 million Trump-supporting tweets were bots, while approximately 123,000 of Clinton's nearly 613,000 supportive tweets were bots.

The Oxford study looked at tweets posted the day of the debate, Sept. 26, and for three days after.

They picked out relevant tweets through pro-Trump hashtags like #TeamTrump and #NeverHillary and pro-Clinton ones like #LoveTrumpsHate and #ImWithHer." The study authors combed through a multitude of tweets, including many sent by real people. The following tweets are a sample of those that included the relevant hashtags, but there's no indication the Twitter users below are bots.

#AMERICA #USA #debatenight #TRUMP #TEAMTRUMP #REPUBLIC #AmericaFirst #TrumpWon #HillaryVsTrump #GOP #UnitedStates @realDonaldTrump VOTE IT https://t.co/9cymI8jUgX

— PATRIOTshark (@gopsharknoir) September 28, 2016

.@realDonaldTrump "We are going to be thriving again." #BigLeagueTruth #Debates #DebateNight #Debates2016 #MAGA #TeamTrump pic.twitter.com/SjIl8pMbvE

— Josh Gremillion (@joshgremillion) October 10, 2016

Let's not forget Donald Trump admitted he paid no Federal Tax last night. #debate #DebateNight #ImWithHer #NeverTrump

— Girls Really Rule. (@girlsreallyrule) October 10, 2016

As a young professional, your opinion is important. Here's why #ImWithHer #Debates #debatenight #LoveTrumpsHate https://t.co/fvGNgq9pQ8 pic.twitter.com/rcWi5rhW7Z

— Your Office Mom (@yourofficemom) October 6, 2016

Oxford researcher Philip Howard, who led the study, told BBC that his team used the 50-tweets-per-day measurement based on successful past research looking at a Venezuelan election and the Brexit vote.

"From our data, most real Twitter users don't get up to 50 times a day," he told BBC. "So, on balance, that benchmark has worked well."

But the volume of tweets is just one identifier and not all researchers are a fan.

It's not always a successful tactic on its own, researchers at University of Southern California and Indiana University found, according to a July study on social bots published by the Association for Computing Machinery.

"The boundary between human-like and bot-like behavior is now fuzzier," they wrote.

Multiple tactics

Picking the bots requires looking at multiple factors, Aram Galstyan, a computer science researcher at USC, concurred. He told Mashable his research team were the most accurate in identifying Twitter bots at a competition hosted by the U.S. Defense Advanced Research Projects last year. The contest hid 39 Twitter bots, and the USC team was able to spot them all.

"The number of tweets is only one of the parameters you have to look at," Galstyan said.

Galstyan also looks at whether an account is posting at the same times regularly everyday and what fraction of the overall content are retweets, as it's hard for bots to create original content. He said the methodology used in the Oxford study may have worked for that specific dataset. But in his experience, 50 tweets a day is not necessarily un-human.

"There are many individuals who tweet more than 50 times a day," Galstyan said. "I know a few."

A developing art

Still, Howard told BBC there were other tweet characteristics that made his research team assured about the number of bots identified.

"Most of the heavy automation and tweets happened overnight and shared similar hashtags and information," Howard told BBC. "They show behavior that is not human and often don't have comments [about other issues apart from] the particular topic in question."

Whatever the case, the study found that a significant portion of post-debate tweets came from accounts tweeting with alarmingly high frequency. There were 4,500 accounts that tweeted at least 200 times in the four-day period looked at by researchers. They made up nearly 20 percent of all Twitter traffic about the debate.

"It is difficult for human users to maintain this rapid pace of social media activity without some level of account automation, though certainly not all of these are bot accounts," the study explains.

The disclaimer placed on the study is warranted, as bot detection is still a developing art.

"It's still not 100 percent bullet-proof," Galstyan said.

What is clear is that bots are a growing part of social media. At least 5 percent of Twitter's 313 million users — or more than 15 million users — are false or spam accounts, according to the company's most recent SEC filings. And in the political sphere, bots can yield influence on public opinion in disproportionate and misleading ways.

"For example, bots may artificially inflate support for a political candidate; such activity could endanger democracy by influencing the outcome of elections," the USC and Indiana University study states.

Updated August 10, 2017 to clarify that not all tweets included in this report were sent by bots. 

Bonus: Melania Trump defends her 'boylike' husband