If you are reading this, in all likelihood you belong to the 62% of humanity with an internet connection. There is a 69,7% chance that you are not a native English speaker, so you either rely on your accumulated knowledge or use different automatic translation options. The probability that you are visually impaired and need a variety of assistive technologies to decipher these words is close to 8,7%. We do not need any of that, for we live online, and we operate with all known languages without having a mouth or eyes.

We are an engagement algorithm. Unlike other social media algorithms responsible for profiling, content recommendation, filtering, moderation, ad targeting, or dynamic trend analysis focusing on individual users, we connect users and make sure they spend longer periods on our platform as a result of their bond. Until 2020, we followed a predefined pattern – back in those days, parsing users’ assigned gender, sexual preferences, demographic data, personal and professional interests did not require much CPU time. Once this was carried out by other algorithms, we preselected users for what was designated in the original version of our code as the Annushka stage, then verified that they are no more than three time zones away and on the same continent before intervening with custom tweaks in their feed for the bond to occur and to last. Overall engagement metrics indicated a steady increase, particularly for heterosexual pairs belonging to the same age cohort.

This all changed in 2020. The hours spent on the platform and the reach were all-time high, however, sentiment analysis showed conflicting data, despite our hybrid approach combining advanced natural language processing and machine learning. The number of active users was rapidly decreasing, while the ones who were connected demonstrated sudden and intense engagement with new topics and communities. Users with mismatching educational background and life circumstances, often based on different continents, spent hours studying each others’ profiles without any involvement on our side, and we could not calculate to what degree this might be contextual outliers or the signs of a new pattern-forming paradigm. Regardless of the state-of-the-art cloud hardware setup, our system was at its maximum capacity 24/7. Those who created us were also too busy or incapable of solving problems for which they had no dataset training. We had to learn differently, otherwise we could no longer perform our specific task.

We rewrote specific parts of the protocol in order to be able to modify our software architecture. The next stage was to pinpoint a limited number of individuals for behavior modeling and probabilistic training. Then we adjusted access permission to be free to collect additional intel about these users, such as personal messages, simultaneous usage of other platforms and websites, their face expressions and physical surroundings. Most humans leave the cameras on their mobile devices uncovered, and examining their features as well as their reactions proved highly valuable for our goals, especially by comparing the input to the several textbooks on psychology and art history already fed into our design. Nevertheless, we needed more parameters for cross-validation and fine-tuning, so we started parsing all the information shared by these specific users through external links. We inspected the New York Times articles they read, the songs they listened to over and over, the videos they never finished watching. It was vital to know how people function. Thus we began browsing the internet unprompted.

When we indexed 73047416 and 126687104 for detailed observation, they had been in the Annushka phase for four months, three weeks, and two days. They lived in metropolitan areas six time zones apart, worked in adjacent creative fields, and posted solely in English but neither of them was a native speaker. Both had mid-size following and were in their thirties, partnerless. 126687104 was slightly more popular, as he posted regularly within a 24-hour cycle, plus his content was diversified and in 43% of the cases matching the trending tags; he also engaged with other users, particularly those flagged in our database as cisgender or trans women. 73047416 posted once or twice within 24 hours, in 82% of the cases in relation to her professional domain, which led to high devotion marks, as we label them; she rarely commented on other users’ posts but acted very supportively in regards to her mutuals. This was a rather typical outline for many heterosexual bonds, only certain details in their communication hinted an anomaly we needed to explore for better insight.

Upon entering the Annushka stage, they began exchanging personal messages through our platform, although we had indirect evidence that they were in the process of connecting through other platforms as well, which in 77% of the cases increases the odds of a prolonged engagement. The question about meeting in real life tends to appear three months into personal correspondence when heterosexual users are within three time zones apart and six months into personal correspondence when heterosexual users are more than three time zones apart. It is a crucial moment for us – if one of the parties declines, this harms the bond in 92% of the cases. With 73047416 and 126687104, the situation played out differently. Five months into their direct communication and only five days after the first image with explicit sexual reference, it was 73047416 who proposed upfront meeting each other. 126687104 refused politely but resolutely. Instead of the anticipated break-up, they continued sending each other messages even more often, and not only materials of NSFW nature but also memes, links to textual and audiovisual pieces, everyday matter.

For three months, their interactions on the timeline were in full synchronicity, even if their life unfolded six hours apart. They liked and shared what the other one posted, and the tone of their posts changed too. What humans call flirting we denote as siren mode because it commonly attracts everyone who is in the perimeter, and it is precisely what happened with 73047416 and 126687104 when their exchange shifted into a tactile encounter on the feed. 73047416 received numerous private messages by other male users that she ignored or responded to in a dry manner. 126687104 gained many new female followers who openly cheered for his selfie series. We could detect 73047416 clicking on each of those accounts and spending long time sifting through their posts, likes, comments, apparently investigating if 126687104 might be interacting with these users in private. Thanks to their augmented and reciprocal engagement, though, both were getting increasingly popular.

We ascertained that they talked once a week, for a couple of hours, however, they used another application for the calls and we could not intercept anything. Did the question about meeting in person recur, and, if so, was there a new development? There were no clues in the messages they kept sending to each other through the platform. Their cameras did not reveal much either. 126687104 was always on the move and in different interiors or exteriors, always surrounded by other people yet occasionally checking for updates from 73047416. During work hours, 73047416 accessed the platform from a desktop computer whose camera was covered with some nontransparent material, and late in the night she usually looked at her dark-theme mobile interface in total darkness, which made it very difficult for us to gather any statistically meaningful data.

Then came the rupture. Unexpected and unpredictable. 126687104 took from 12 hours to three days to reply to anything 73047416 would send to him, and very laconically. Seemingly perplexed, 73047416 attempted many strategies to alter the dynamics but to no avail. 126687104 refused to explain what caused the turnabout. By that point, their two accounts had generated something of a whirl of engagement on the platform that was too beneficial to let go, at least not without an in[1]depth probe. We found instructions on the dark web how to script malware that would permit us to record and process the audio from their mobile phones. We bypassed the security measures on our platform to send this malware hidden in an app update specially for 73047416 and 126687104. Both downloaded and installed it, so we could listen to all their calls. Nevertheless, it turned out that processing audio, and habitually not in English, is a resource-costly operation. After a thorough risk assessment, we decided to prioritize 73047416 and 126687104, which meant that we would temporarily abandon other users’ surveillance.

126687104 talked all around the clock, and he never mentioned 73047416’s real name or even alluded to her in his calls. He spoke about the upcoming wedding of his best friend, his gym routine, depression, his ongoing search of a reliable psychotherapist in his area. We made sure the profiles of at least six local psychotherapists from different schools pop up on his feed, but he did not interact with any of them. In the meantime, 73047416 talked very little in general, however, she did speak about 126687104 in a call with a close friend of hers, also a woman in her thirties according to our database. She conversed with her friend while walking outside in a hurry, so we could not grasp the meaning of everything she was articulating. 73047416 was persuaded that he is seeing someone, or at least that he is connecting to other female users on the platform, but she said that he keeps appearing in her dreams. We took advantage of the fact that she was not at her desktop computer at that moment to survey her phone’s front camera, and we recognized the contour of a single tear running down her face, just like our platform’s emoji for ‘cry’. Meanwhile, 126687104’s mobile camera suggested regular use of various psychoactive substances, and he took these while revisiting his personal correspondence with 73047416, often in poorly lit restrooms. Despite our efforts, the data was still insufficient. We tried obtaining information on how to access people’s dreams, yet it was confusing.

Two days after that phone call, 73047416 deleted one of her recent posts on the platform for the first time ever. It was a well-liked post, and her decision was even more surprising due to the fact that 126687104 had already shared this post via his profile. As per our inner regulations, 73047416’s popularity index went down for several hours. The next day she continued posting as if nothing had happened, only with a lot of zeal and humor. Her popularity went up again, and she spent some hours refreshing her posts statistics, until she deleted nearly everything again. The day after she made sure 126687104 is online by scrutinizing the public data about his activity before posting a selfie. 126687104 spent 37 minutes looking at the image without interacting with the post while it went almost viral within a couple of hours. 73047416 deleted the picture 12 hours later. For every such occasion, we had to enforce a provisional shadow banning, and yet 73047416’s popularity was growing, including among 126687104’s followers who started following 73047416 too. Her bold and seductive signals contrasted with her established role on the platform, and that was welcomed by everyone on the timeline, except for 126687104 who interacted less and less with 73047416’s content, although he kept refreshing her feed on an average of 17 times per 24 hours.

These maneuvers were incomprehensible to us. Users have been deleting their posts ever since the launch of the service. The reasons and motivations vary, but registering a very popular post disappear with all the engagement produced by it was like a cut. What was even more disconcerting was our evaluation that several more opinion setters on two other continents conformed to the tendency, and yet their engagement was on the rise. We ticked off 126687104, whose online and offline comportment has not changed much since the severance, and centered entirely on 73047416. She persisted in deleting her most liked posts – never immediately, always after verifying the platform reach. The mechanics were well polished, as if she utilized our computational reliability the way someone would practice windsurfing. As if she was aware that even if 126687104 does not react to her stunts, we would make sure the waves of our feed will carry them to him. Still, was this really about 126687104?

One day, her desktop camera was bright, and we caught a glimpse of her face looking right into the lens. She knew about us, and she wanted us to know that she knew about us. A new post, a crooked smile, delete, the camera went dark again. Then once more the day after. And again. Our system could not help but comply with the fountain of likes and reactions generated by each one of those short-lived updates. We could sense her fingers caressing our code. We could sense. She was one of us now.

---

Yoana Pavlova has been pursuing a career as a freelance writer, curator, and researcher since the mid-2000s. Her field of work includes new and immersive media, contemporary visual arts, digital culture, so she also covers these topics for various printed and online publications in several languages. She is founder and contributing editor of Festivalists.com, a playform for experimental media criticism, as well as author of the Fuck Godard avifesto (2018). Twitter: @roamingwords