午夜福利1000集合

A FEW weeks ago, I noticed that a foul and offensive hashtag was trending on Twitter. Like a horror movie character who goes into the basement after hearing monster noises, I clicked on it.

Every post on the hashtag was like a parody of a political debate, with each side making the same screaming accusations. It was almost as if these people had learned to argue from bad algorithms.

That is when it hit me. Maybe these angry tweets were generated by algorithms. Or by operatives at a place like the Internet Research Agency in Russia, where they make memes to fan the flames of the political trash fire in the US. Not for the first time, I wished that I could check some kind of social media weather report on outbreaks of propaganda.

That dream isn鈥檛 so far from being turned into reality, it turns out. Meysam Alizadeh at Princeton University is making an automated system for identifying trolls on social media 鈥� and predicting what they will say next. He and his team say they want to create a public dashboard that shows 鈥渨hat鈥檚 happening on social media and whether there is coordinated activity sponsored by foreign states鈥�.

To do that, they have trained a set of algorithms to spot the telltale signs of so-called influence campaigns. The group started by working with data sets released by Twitter and Reddit, which contained distinct troll activities originating in Russia, China and Venezuela between 2015 and 2018.

The campaigns were all aimed at the US, but they had very different approaches. Trolls from China seemed mostly to target people in the Chinese diaspora, especially ones with an interest in Islam. Venezuelan trolls tended to be bots spouting political news and links to fake news websites.

The Russian trolls were the craftiest. They responded quickly to current events in the US. Their posts about Black Lives Matter spiked during protests, and ones about Islam peaked during President Donald Trump鈥檚 various travel bans on Muslims entering the US.

鈥淭hese weren鈥檛 bots spewing automated hate; they were Russian operatives, reacting to US news in real time鈥�

Alizadeh says there was a distinct, week-long Russian influence campaign aimed at actor Alec Baldwin, who has done many satirical impressions of Trump on Saturday Night Live. Alizadeh speculates that these weren鈥檛 bots spewing automated hate; they were trained Russian operatives, reacting to US news in real time.

Once the algorithms had learned these distinct patterns, Alizadeh and his colleagues set them loose on data sets that contained some troll posts and some 鈥渃ontrol鈥� posts from typical users. After several tries, the algorithms were able to predict whether or not a post was from a troll most of the time. The Venezuelan trolls were easiest to identify, with 99 per cent accuracy on some tests. When it came to Chinese and Russian trolls, the algorithms got it right between 74 and 92 per cent of the time (Science Advances, ).

That isn鈥檛 perfect, but it is a lot better than I can do with my armchair speculation about how a nasty hashtag might be an influence campaign.

The real question is, how do you separate real social media nonsense from fake, when the fake accounts are so nimble and constantly changing what they are discussing? Alizadeh says the answer is to train these troll-seeking bots on new data every month. Based on the previous month鈥檚 activity, he believes it is possible to generate accurate propaganda weather reports for the next month.

Here鈥檚 hoping that Alizadeh鈥檚 algorithms are coming to a social media platform near you. I can鈥檛 wait for the warnings: 鈥淎n 80 per cent chance of foreign government-sponsored disinformation about Islam this week, with a 40 per cent chance of conspiracy theories about voting.鈥�