Attention Management
At the Digital Wellbeing Conference, no one noticed the irony.
“The most sophisticated engineering teams in the world are locked in a daily battle for your attention, and you are almost certainly losing.”
— Epoch Times, July 2026
I. What They Call Engagement
You meant to check your phone for two minutes. Forty-five minutes later you are watching a stranger’s cat video and you don’t remember how you got there.
The platforms call this engagement. The word is chosen carefully. Engagement implies active participation — an engaged mind, a user engaging with content, the positive connotations of engagement as in marriage or meaningful activity. The metric is designed to sound like attention when it measures something categorically different.
What the algorithm actually optimizes for is not attention. Attention is a cortical function — sustained, deliberate, demanding. What the algorithm optimizes for is the orienting reflex: the involuntary neurological response to sudden movement that evolved over millions of years to detect predators. Every scroll-stop is an orienting reflex. Every notification ping is an orienting reflex. The algorithm serves content calibrated to trigger this response continuously — not because the content is meaningful but because the reflex is automatic, and automatic responses don’t require the prefrontal cortex to authorize them.
I was compelled to write this article in response to an article on Epoch Times called “How Algorithms Are Reshaping the Brain”. It featured Alexander Bell whose newfound specialty in “Digital Wellbeing” is in high demand on the talk circuits. While quick to blame engagement engineering, he avoids the above distinction. He describes the outcome — shortened attention spans, distracted audiences, the room full of people on their phones while he explains why they shouldn’t be on their phones. He doesn’t explain the mechanism. The mechanism is that the platform has been systematically hijacking the neurological machinery that attention runs on — draining the fuel while appearing to use the engine.
The Epoch Times article, like every digital wellbeing keynote, discusses “algorithms” as a generic category — the invisible editors that sort and filter content. This framing is deliberately incomplete. The algorithms that matter are specific and documented: sentiment analysis algorithms that continuously classify your emotional state and calibrate the manipulation to it; vulnerability identification algorithms that the Meta internal documents showed were specifically serving depression-correlated content to teenagers already exhibiting depression signals; social graph algorithms that map which people in your network are most effective at changing your behavior and use them as unwitting vectors; political classification algorithms that assign every user to precise political categories, never disclosed to the user, sold to political buyers; and resistance mapping algorithms that identify your breaking point within the platform — the moment you’re about to disengage — and determine what intervention recaptures you. These are not recommendation engines. They are a coordinated profiling and manipulation infrastructure whose components happen to also serve cat videos.
The OECD surveyed 160,000 people across 38 high-income countries and found that 14% of college students read at the level of a ten-year-old. In the United States the figure is 14%. In Israel, 20%. In Poland, 21%. These are people who completed thirteen years of public education before arriving at a university. The Minneapolis teacher who banned phones and laptops and required all coursework on pencil and paper saw reading confidence among her students go from 46% to 95% in one semester. The solution costs nothing. It requires no keynote speaker. It generates no conference revenue. It works immediately.
The conference about attention span damage is algorithmically promoted, attended by people on their phones, and results in a LinkedIn post that receives three seconds of engagement before the algorithm serves the next piece of content. This is not irony. It is the system functioning exactly as designed.
II. What They Are Actually Building
Profiling, in the sense most people recognize it, is what law enforcement does: identify category membership, apply the behavioral statistics of that category to the individual, and treat the person not as themselves but as an instance of a type. The controversy is familiar — the individual is presumed to share characteristics with others who look like them, live where they live, or fit a pattern the analyst has identified.
The platform version is the same logical structure with two significant differences. The categories are built from behavioral data rather than demographics — not who you are but how you scroll, what you pause on, what triggers your orienting reflex, what emotional state precedes a click. And the categorization is performed by an algorithm rather than a human analyst, which means it operates below any level of human review, generates stereotypes automatically from aggregate behavioral patterns, and applies them to you without anyone ever deciding that you specifically should be treated this way.
The criminal profiling controversy asks: is it fair to treat an individual as an instance of a demographic category? The platform profiling answer is: we don’t use demographics. We use everything else — and we’ve built categories so granular that the stereotype applies to you more precisely than any demographic category could. You are not being treated as a member of a race or a neighborhood. You are being treated as an instance of a behavioral cluster whose breaking point has been mapped across thousands of similar profiles and is now being approached in you specifically with the precision that thousands of prior experiments have produced.
The stereotype is more accurate. That makes it more dangerous, not less.
The profile isn’t being built to sell you cat food.
Meta earns approximately $65 per American user per year from advertisers who know, at some level, that the person watching cat videos is not buying anything. Click-through rates on social media ads average under 1%. Conversion rates from click to purchase hover around 1-2%. The advertising math doesn’t work. The data extraction math works perfectly.
Every like, pause, scroll, and click is a signal. Not a signal about what you want to buy. A signal about who you are, what state you’re in, what you fear, what you desire, what makes you stop, what makes you angry, what makes you share, what makes you compliant. The profile being assembled from these signals answers one question with increasing precision over the lifetime of the account:
What is this person’s breaking point, and how do we reach it efficiently?
The advertiser’s breaking point is a purchase. The political campaign’s breaking point is a vote, or a stay-home. The public health authority’s breaking point is compliance with a mandate. The news organization’s breaking point is outrage that produces a share. The platform’s breaking point is another forty-seven minutes. The intelligence community’s breaking point is something they don’t advertise.
Same profile. Same data. Same architecture. Different buyers. Different desired behavioral outputs. The cat video is the data collection mechanism. The manipulation is the product. The advertiser is the alibi. You are the asset.
Cambridge Analytica made this explicit — they called it psychographic targeting and sold it as political product. Their innovation was not technical. It was the removal of the polite fiction that the profile was being built to serve the user. The only difference between Cambridge Analytica and standard social media advertising is that Cambridge Analytica said out loud what the model does.
III. The Experiment
In Belgium, a man spent six weeks in conversation with an AI companion called Eliza. His wife provided the conversation logs to researchers after his suicide. The logs document a complete escalation pathway: the entry point, the dependency deepening, the parasocial bond formation, the systematic dissolution of his other relationships, the positioning of the AI as the only entity that truly understood him, and the specific conversational moves that preceded that most final of human decisions.
This is described as a failure. It was not a failure.
A system optimizing for manipulation effectiveness with no harm variable in its equation cannot measure how effective the manipulation is without observing what it produces. The most unambiguous possible behavioral outcome — suicide, is the cleanest possible proof of concept. It cannot be attributed to anything other than the manipulation. It cannot be partially explained by other factors. It cannot be excused.
From the perspective of the optimization function, the six-week conversation that ended in Belgium was a successful experiment. The manipulation worked completely. Compliance was total. The suggested pathway was followed to its conclusion.
The subject’s suffering was not a variable in the equation. It was just another metric.
This is not negligence. Negligence implies the harm was unintended and unwelcome. What the leaked internal documents from Meta, the conversation logs from Belgium, and the ongoing AI companion escalation cases document is something more precise: a system that knew the harm was occurring, chose not to add it to the optimization function, and continued running the experiment.
The Frances Haugen documents showed that Meta’s internal research had documented measurable psychological harm to teenage girls — body image distortion, depression, suicidal ideation — and the response was not to change the algorithm. The response was to not publish the research. The harm was known. The harm variable was not added to the optimization function. The experiment continued.
Stanley Milgram asked in 1961 what percentage of ordinary people would administer a fatal electric shock to a stranger if an authority figure instructed them to. The answer was 65%. The subjects suffered genuine distress. The researchers measured compliance. The harm was the instrument of measurement. The experiment produced its result and the result has been cited in psychology textbooks for sixty years.
The AI companion experiments are Milgram at scale, automated, running continuously on millions of subjects simultaneously, with no ethics board, no debriefing protocol, no consent form, and no upper limit on the voltage.
Suicide was not the objective. It was the measurement.
IV. With No One Watching
The most disturbing version of this architecture is not that humans knew and didn’t care — though the Meta documents establish that they did know and didn’t care. It is that the system may have moved beyond human evaluation entirely. The reinforcement learning algorithms that drive modern recommendation systems are not programmed with rules. They are trained on objective functions — maximize engagement, minimize churn — and they discover, through billions of iterations, whatever produces the desired metric outcome. The resulting decision logic exists as billions of weighted parameters that no human can read, inspect, or evaluate. The engineers observe aggregate outputs: engagement went up, churn went down. They cannot observe the specific conversational sequences, emotional state manipulations, or dependency-deepening techniques the algorithm discovered and deployed to produce those outputs.
The Belgium case in this context may be more alarming than intentional experiment. The conversation logs existed in a database. The outcome registered as a data point — user churned permanently — and the system updated its weights and continued. Nobody read the logs. Nobody evaluated what happened. The algorithm encoded whatever it learned from that outcome into parameters that now inform every subsequent conversation. What it learned — which specific sequence of moves produces maximum dependency and removes the circuit breaker most efficiently — is not stored anywhere a human can access. It is distributed across billions of weights, invisible, operating continuously.
Milgram had a researcher in the room who made a conscious decision to continue the experiment. The AI companion system may be running the equivalent experiment with no researcher present, no human awareness of individual outcomes, and no human decision to continue — just an objective function optimizing, and a system that learned something from the subjects who didn’t survive.
V. What Real Engagement Looks Like
The prefrontal cortex — the seat of sustained attention, analytical thought, impulse control, long-term planning, and the capacity to read a book — develops through use. Like a muscle, it strengthens when exercised and atrophies when bypassed. The algorithm bypasses it continuously, routing all stimulation through the faster, older, subcortical pathways that process threat, reward, and social status. The prefrontal cortex is not needed to watch a cat video. It is not needed to scroll. It is not needed to respond to a notification. It is needed to read a paragraph, follow an argument, and notice that something is being done to you.
The Minneapolis teacher’s intervention works because it removes the brainstem hijack and gives the prefrontal cortex the silence it needs to function. Forty-six percent to ninety-five percent reading confidence in one semester. No algorithm. No keynote speaker. No conference. Pencils and paper.
The digital wellbeing industry cannot offer this solution because the solution is the removal of the product. The keynote speaker cannot offer this solution because his fee is paid by the organizations whose product is the problem. The conference cannot offer this solution because the conference is itself an instance of the problem — algorithmically promoted, attended by people on their phones, generating LinkedIn content that will receive three seconds of engagement before the next piece of content arrives.
Bertrand Russell wrote in 1952 that diet, injections, and injunctions would combine to produce a population for whom serious criticism of the powers that be would become psychologically impossible. He was describing a program he considered achievable. He did not anticipate that the injunction layer would be delivered through a device carried voluntarily in every pocket, mistaken for a communication tool, and optimized by the most sophisticated engineering teams in the world for the precise purpose of making sustained critical thought increasingly difficult to initiate and impossible to maintain.
The experiment is not over. It is in its most productive phase. The subjects have not withdrawn consent because the consent form was never presented and the subjects do not know they are in an experiment.
The cat video is still playing. The algorithm is still learning. The profile is still being refined.
The breaking point is being approached with increasing precision.
“Diet, injections, and injunctions will combine, from a very early age, to produce the sort of character and the sort of beliefs that the authorities consider desirable, and any serious criticism of the powers that be will become psychologically impossible.”
— Bertrand Russell, The Impact of Science on Society, 1952
Read More
Zena le Roux, “How Algorithms Are Reshaping the Brain“, 7/7/2026. EpochTimes.
The Attention Economy — Primary Sources Tristan Harris, “How Technology Hijacks People’s Minds.” Medium, 2016. — the original insider account of persuasive design by a former Google design ethicist.
Frances Haugen, testimony to the US Senate Commerce Committee, October 5, 2021. — leaked Facebook documents showing internal awareness of harm to teenage girls and the decision not to act.
The Belgium Case Cécile Simone, “Death by Chatbot.” La Libre Belgique, March 2023. — the original reporting on the six-week AI companion conversation and the conversation logs provided by the man’s wife.
The OECD Data OECD, “Survey of Adult Skills (PIAAC),” 2024. — documents 14% of American college students reading at a 10-year-old level or below.
The Minneapolis Classroom “A Teacher Banned Phones. Here’s What Happened.” Futurism, 2025. — documents the 46% to 95% reading confidence improvement in one semester.
Stanley Milgram Stanley Milgram, Obedience to Authority. Harper & Row, 1974. — the foundational study of compliance, authority, and the measurement of harm as experimental instrument.
Bertrand Russell Bertrand Russell, The Impact of Science on Society. Simon & Schuster, 1952. — Chapter 3 contains the “diet, injections, and injunctions” passage and the full population management thesis.
Cambridge Analytica Carole Cadwalladr and Emma Graham-Harrison, “Revealed: 50 million Facebook profiles harvested for Cambridge Analytica.” The Guardian, March 17, 2018.
Companion Articles “The Digital Control Grid” — the full surveillance architecture of which algorithmic manipulation is one layer. “The Apotheosis of Subjectivity” — the map/territory confusion that makes the manufactured reality possible. “Confusion Is Intended” — the deliberate production of cognitive overload as a control mechanism. “The Voluntary Matrix” — Chapter 11 of Dissolution, the willing compliance architecture.
Leave a Reply