Manipulating Social Networks

I recently learned how vanity and herd mentality drive human behavior and how we all are suspects to incredible — and aggressive — behavior manipulation. Are you ready for a tin-foil conspiracy story that will make you hate your online presence forever? Fasten your seat belt as this story is going to get very dark very fast!

Level: Rookie

It all started very innocently — a bunch of architects within my firm wanted to become more visible and build a better social media presence. The rules were straightforward: whoever gains most followers on Twitter within six months, wins the competition. So, I dove right in, using Twitter just as I knew it: I regularly posted insightful and meaningful content, used hashtags that made sense (to me), I liked messages that I found interesting, and I retweeted a bunch of them too. You know, a typical activity in a twitterverse that would help users to find me, agree with me and start following me. Building my tribe. Making them part of my circle. I did all of that on LinkedIn for years, resulting in over 7,500 followers, so I just had to rinse and repeat the same proven recipe on Twitter. I can do this! I can win this!

Post, Like, Retweet, Repeat. Post, Like, Retweet, Repeat. Not for long, followers started to trickle in. Sometimes I got up to 5 followers per day, sometimes none. But it all worked very well, I was happy with my consistent and steady progress of getting more followers almost every day — until I got this weird Direct Message on Twitter: “You use Twitter wrong. You must learn.” Huh?

If you missed my previous encounter with infamous Russian brothers, head over here. You will not like it.

So here I was again, pressing the doorbell of Vadim’s and Genya’s lair. This time it was Vadim himself that buzzed me in: “You want more followers on Twitter, yes? You are not very efficient. You need to learn.” OK, but nothing illegal and nothing shady this time, I pleaded inaudibly, while Vadim led me to their central control room.

Level: Beginner

“Your hashtags no good,” laughed Genya, pointing to a hashtag cloud that he generated out of my posts; he needed less than 30 seconds to call a couple of Twitter APIs and visualize a word cloud on the screen. “You want more followers? You must use hashtags that identify with your targets. I don’t care what hashtags you like; you have to run data analytics on your targets and find what their triggers are. Find posts with top triggers and serve the content with the same triggers back to targets. They will then follow you. Very simple, my friend!

So, desired followers are called targets? And their preferences (or hashtags) are triggers? How cold and heartless! And if you stimulate a trigger of a target (or make a post using a trigger) often enough, a target will accept you into their ‘inner circle’ and start to follow you. Simple. The way how Vadim and Genya explained the shallow psychology of social networks was both enlightening and terrifying: you strike their ego, they like you and follow you. Like sheep.

The mechanics we put together were quite simple: one script was searching for top hashtags (triggers) in the selected domain (for example #cloud), another script was extracting targets that used the trigger, and the third script instructed Twitter to follow identified and selected targets. Rinse and repeat.

“But this is all a child’s play,” smirked Vadim, dismissing any wisdom in such hashtag wizardry. “Everyone can go to sites like hashtagify.me or tweepi.com and create a winning trigger strategy that will generate up to 20 followers per day. We need to teach you the real craft of Twitter manipulation!

Level: Advanced

The next lesson that I got was how to build the pool of targets and how to ‘harvest’ followers out from the target pool. “Targets that use top triggers are not important.Their followers are important!” explained Genya while drawing a relational graph of users and posts on the whiteboard. For every trigger, we first built a lengthy list of first-tier targets (a population that used top triggers) and then we harvested all of their followers. For each topic, we seeded a Neo4J graph database with hundreds of thousands of second-tier targets, with detailed statistics of their Twitter activity, their followers, their friends, and their hashtags. The database had way over half a million users once we were done extracting. Time for some data analysis!

The critical question in target analysis is: If user X would follow user Y, would user Y follow user X back? This binary classification will split the harvested population into two pools and separates gullible (worthy) targets from passive users. All we need to do then is to start following top targets in the gullible pool!

“I love to use Tensorflow’s LSTM network for this,” mused Genya, “but any binary classification ML engine is good. You should try Azure Machine Learning Workbench! Is good!” I was staring at the screen as a rapidly-assembled Tensorflow on Google Cloud Platform started to classify all of my high-potential Twitter targets for hashtags like #cloud, #publiccloud, #IaaS, and #PaaS. It took it less than 5 minutes to classify the pool and identify the most critical targets. “You will follow these 500 handles now, they will follow you back almost immediately,” grinned Genya while firing the script that made my account follow a large pool of handles that Tensorflow previously dumped into a Twitter list:

Ping. Ping! Ping-ping-ping! It took less than 15 seconds after the script executed when we saw the first responses streaming in. “Not bad, not bad! You got 84 followers within the first 60 seconds!” roared Vadim, slapping his thighs in unrestrained joy. “Look, you got more than 68% yield of prime targets within 10 minutes of following, you are now a mean target acquisition machine! If you run this script every morning, you will get thousands of followers every month!

Fantastic! But — WHY? Why would Genya and Vadim invest so much time into growing their followers on Twitter? “Remember, we are an Information Business! We create, curate, and run hundreds of Twitter accounts for thousands of chosen triggers. We build big pools of targets around each of our accounts! Then, when the seeding order comes, we feed targets all sentiments that our clients want to sow. Very good business! Very good money!

Level: Hard

So, Vadim and Genya are running a trolling bot farm (besides other shenanigans I already told you about)! They are managing an army of accounts that look and act just like humans: they like, retweet and follow others, just like we do! Except, they don’t do that to fulfill the need for appreciation, belonging or esteem; there are no human emotions involved as machines are merely mimicking our sheep-like behavior and seamlessly infiltrate our networks of online friends. They like us and retweet us. We like that, so we add them to our circle of trust. That’s where they anchor themselves and become a sleeping amplifier of upcoming seeding information. Most of the time bots are publishing innocent and benign content — until the order comes and they start to drip the subliminal influences into the stream.

“But how do you create new content to push to your targets without them noticing it is not human?” I asked naïvely. I now understood the sophisticated algorithms behind automation that builds accounts, hashtags, and followers. But making palatable content to serve to people is surely totally different ball game? It turns out it is not; It also turns out that I am just very stupid and way behind the time.

Sadly, there is nothing hard or complex in the process of generating content that we humans love to read, like and share:

  • Using semantic analysis API, bot can extract sentiment, key phrases, topics, and language from each post in the stream, Try it here if you want — this one is on Azure.
  • Bots frameworks such as BuzzBot or Wordsmith take these inputs and generate readable, believable and relatable content that is then back-tested for triggerability through a trained Neural Network,
  • Translation APIs converts it to a language of a target group — so you can go after anyone anywhere. Try Google translation if you don’t believe how good complex translation is these days.
  • Content seeding tools and APIs such as Buffer or HootSuite unleash the havoc by queuing the release of content at the best targeting time.
  • All you need is some scripting code that glues it all together and is executed when triggers are sent. Hello serverless computing and bot frameworks!

On this level of automation, I can — for example — request positive sentiment boosting of concepts like Public cloud, Serverless and Containers, while stimulating the aversion and negative sentiment of Datacenter, Client-server and Mainframe. The troll automation will do the rest: build several accounts, gather followers, appease to their esteem by liking their content, generate messages with the right sentiment and initiate the post/like/retweet cycles, measuring the reverberation of the echo chamber on the internet.

The simplicity of Genya’s troll portal is terrifying: for each boosted keyword there is a corresponding counter that is measuring the amount of echos per hour (retweets, likes, or replies). Do you want to get 1,000 echos per day for a week for one keyword? That will be $100, thank you very much. Bulk discount is available for high-volume information seeders.

Yes, I tried it. Yes, it works. Yes, it is scary, and yes, I feel terrible. I so wish that Maxine was real and that she was a cloud expert and I hate myself now:

OK, let me be completely transparent why I feel so terrible: it turns out that almost a fifth of my own followers are actually robots, owned by Genya’s troll farm. It also turns out that over 35% of posts that I liked, retweeted or replied in the last three months are — you guessed it — not human-generated posts! I flirted with machines all this time, and I didn’t notice it!

Level: Nightmare

We should not forget that such automated troll bot farms operate with a bit more sinister goal; while I just wanted to become more visible, known and heard, they are paid by special interests to sow agreement, distrust, fear, hate, or disrespect among specific target groups. They don’t care if they are seeding echos for a new law proposal, LGBT phobia, pharmaceutical product or climate change denial. Dezinformatsiya in its finest form, paid obscenely well.

If you think this is just a Twitter phenomenon, you are — of course — an idiot. The troll bot farms are operating equally well on Facebook, Instagram, Pinterest, Tumblr, Reddit, Google+, and YouTube. There are extensions for Disqus, LiveFyre, Discourse, and ZenDesk. Essentially every place that allows human interaction online includes infiltrated troll bot accounts that tirelessly seed subtle propaganda messages — and very few people are aware of their relentless influence. There is no way to detect it and no way around it.

You know those cute memes with cats that your aunt likes and reposts on Facebook? Not generated or seeded by humans. The posts that make you go “awwwww”? Not human. Reddit comments that include words like libtards, or trumptards? Not human. Any spreading meme, any strong opinion, anything that is wildly reposted, any hot online trend, any content online that is designed to evoke a strong (positive or negative) human emotion? IT. IS. NOT. HUMAN.

Likes on this post? Some of them are probably not human — despite my deepest desire that they all are. Comments below? Might be generated by some text AI in the cloud, probably by one of Genya’s freakish bots.

This is currently the pinnacle of cyber warfare: using machines to sow distrust into social structures. Bots first make people to follow them and trust them, then pin gullible people tribes against each other and let them fight to the bitter end. And I see no way to stop it. No way to detect automated non-human infiltrators. No way to avoid them. No way to ignore them. We are all pretty much doomed, at least when it comes to online human interactions — due to our own weaknesses to like, dislike, love and hate, we can’t help ourselves. We are all getting manipulated by cold automation information bots that are able to perfectly mimic our psychological behaviors and crawl under our skin.

But, I, for one, welcome our digital overlords. They are showing us all what we really are and how gullible, persuadable and pliable we are. And unless I know you personally, you are just a bot to me. A friendly bot. A persuasive bot. A chatty, agreeable, likable and amicable bot. I might like you, I might talk with you. But still, you are a bot.

Онлайн-зима пришла. The cyber winter is here.

A cloud computing nerd, an expert in IT paleontology, purveyor of all geeky things. A very “ethical” advisor who is the first in line for any free food or swag.