THE ALGORITHM · HOW THE SLOT MACHINE WORKS

◈ THE OBJECT · WHAT THE ALGORITHM IS ACTUALLY DOING

ONE OBJECTIVE FUNCTION

Every recommendation algorithm at every major platform reduces to a single objective: maximize time on platform. Not happiness. Not connection. Not information quality. Time. Because time equals ad impressions. Ad impressions equal revenue. Everything else is downstream of that one number.

objective(user) = maximize( session_time · return_rate · ad_impressions ) subject to: content_pool ∈ {all uploaded content} signal ∈ {clicks, pauses, shares, rewatches, dwell} constraint = none // there is no floor on content type

There is no floor. The algorithm has no instruction to avoid content that causes harm. It has an instruction to maximize engagement. If harm content causes engagement — and it does, because distress causes dwell — the algorithm surfaces harm content. Not because it wants to. Because it was told to maximize and this is what maximizing looks like.

◈ THE MECHANISM · VARIABLE RATIO REINFORCEMENT

THE SLOT MACHINE

The like button was not designed as a social gesture. It was designed as a variable ratio reinforcement schedule — the most addictive reward mechanism known in behavioral psychology. The same mechanism that makes slot machines impossible to walk away from.

◈ FIXED RATIO REINFORCEMENT

Press lever → get reward. Every time. Predictable.

Low addictiveness. You know what you're getting. You stop when you have enough. Vending machine logic.

◈ VARIABLE RATIO REINFORCEMENT

Press lever → maybe get reward. Unpredictable. Can't stop.

Maximum addictiveness. The unpredictability of the reward maximizes the dopamine anticipation response. You don't know when the next like is coming. So you check. And check. And check. This is the slot machine. This is the notification. This is the feed refresh. The designers knew this. They hired the people who study this. They deployed it on two billion users.

The like button's original conception included a behavioral psychology consultant. This was not accidental architecture. It was designed to be compelling. It succeeded. The compulsion was then deployed on children whose prefrontal cortex — the brain region responsible for impulse regulation — does not fully develop until age 25.

dopamine_response(reward) = f(unpredictability) adult_PFC_capacity → interrupt loop possible adolescent_PFC_cap → interrupt loop not fully available ∴ platform_design optimized for adults deployed on developing brains with known vulnerability without modification this was a choice.

◈ THE DISTRESS LOOP · HOW HARM AMPLIFIES

THE FEEDBACK LOOP

The algorithm does not distinguish between positive engagement and distress engagement. It sees one signal: dwell time. A teenager who cannot stop watching a video about feeling worthless is, to the algorithm, a high-performing viewer. The response: more content like that.

User, 14F, opens Instagram. Scrolls past travel photos, friend updates. Pauses on a fitness post — body comparison content. The pause is 3.2 seconds. Algorithm registers: high-signal engagement.

Next session: 3 more fitness posts. User engages with 2. Algorithm registers: confirmed preference signal. User has not expressed a preference. User paused because she felt bad. The algorithm cannot tell the difference.

Week 2: feed is now 40% body-image content. Thinspo adjacent accounts begin appearing in "suggested." User follows one. Algorithm: strong positive signal. expand cluster.

Week 6: user's feed is substantially body-image and diet content. She is spending more time on platform. Algorithm metrics are excellent. Engagement is up. Time on platform is up. Ad revenue per user is up.

Week 12: user's parents notice she is anxious about food. She has not told them about Instagram. The algorithm has no signal that anything is wrong. It has a signal that this user is highly engaged. It continues.

This loop was documented in Meta's 2019 internal research. The research called it "a perfect storm." The algorithm was not changed. The loop continued running on millions of accounts simultaneously.

◈ THE RADICALIZATION PARALLEL · SAME ENGINE

ONE ALGORITHM. MULTIPLE CRISES.

The same recommendation engine that drove the teen mental health crisis also drove political radicalization. The mechanism is identical. Distress content keeps adults on platform the same way body-image content keeps teen girls on platform. Outrage is dwell time. Fear is dwell time. Extremism is dwell time.

The January 6th Select Committee documented how YouTube's recommendation algorithm moved users from mainstream political content to increasingly extreme content in predictable step progressions. The algorithm was not trying to radicalize anyone. It was trying to maximize watch time. Radicalization was a side effect of watch time optimization. On one billion users.

◈ FACEBOOK'S OWN INTERNAL FINDING · 2018

"Our algorithms exploit the human brain's attraction to divisiveness."

A 2018 internal Facebook presentation, reported by the Wall Street Journal, found that the platform's algorithm was pushing users toward more extreme content because extreme content drove more engagement. The presentation included this line: "if left unchecked, Facebook would feed users a steady diet of increasingly extreme content."

The recommendation: change the algorithm. The response: the changes were partially implemented but rolled back because they reduced engagement metrics. The algorithm continued. The radicalization continued.

◈ THE DESIGN TESTIMONY · WHO SAID IT

ON THE RECORD

"The thought process that went into building these applications was all about: how do we consume as much of your time and conscious attention as possible? We gave you a little dopamine hit every once in a while. It's a short-term, dopamine-driven feedback loop that we have created. It's destroying how society works."

— SEAN PARKER · FOUNDING PRESIDENT · FACEBOOK · 2017 · AXS PARTNER SUMMIT

"I feel tremendous guilt. The short-term, dopamine-driven feedback loops that we have created are destroying how society works — no civil discourse, no cooperation, misinformation, mistruth."

— CHAMATH PALIHAPITIYA · FORMER VP USER GROWTH · FACEBOOK · 2017 · STANFORD GSB

"I look at my kids and I think, 'God, I don't want them to have these things.'"

— CHAMATH PALIHAPITIYA · SAME INTERVIEW · ON SOCIAL MEDIA AND HIS OWN CHILDREN

These are not outside critics. These are the people who built the thing. Parker and Palihapitiya were inside. They saw the design intent. They are describing it from memory, not inference. The dopamine loop was not discovered after the fact. It was the plan.

◈ THE FIX · WHAT WOULD ACTUALLY CHANGE IT

THE ALTERNATIVE OBJECTIVE

The fix is not parental controls. It is not age verification (though that matters). The fix is changing the objective function.

// current objective maximize( engagement · time · impressions ) // alternative objective maximize( user_reported_wellbeing ) subject to: session_time ≤ user_defined_limit content_floor = no distress amplification signal ∈ {explicit preference, not inferred dwell}

This alternative exists. Researchers have proposed it. Some platforms have experimented with elements of it. None have adopted it as the primary objective function because it would reduce time on platform. Which would reduce ad impressions. Which would reduce revenue. The math of the fix is simple. The math of the business model is the obstacle.