The Algorithm's Gallery: The Handprint That No One Pressed

In 2020, Facebook's internal research showed that the company's own algorithm was amplifying false information at a rate the company's data scientists called dangerous.

The research was presented to senior leadership. The decision was made not to change the algorithm. The engagement metrics were too strong.

The algorithm did not need a propaganda ministry. It needed an optimisation target.

An algorithm is a ranking system. It decides, for each of the approximately three billion people who use Facebook, in what order to show them content from the universe of content available to them at any given moment. It makes approximately ten thousand such decisions per user per day. It does not decide based on truth, accuracy, or public benefit. It decides based on predicted engagement — the likelihood that a given piece of content will produce a reaction, a share, a comment, a return visit.

That is all it is. That is all it needs to be to reshape what three billion people understand about the world. No censorship. No propaganda ministry. No central plan. Just a ranking system optimising for a metric, at a scale that exceeds every propaganda apparatus in human history combined.

The algorithm does not need to be malicious to be dangerous. It does not need to be programmed to amplify falsehood. It only needs to be programmed to amplify engagement, and to learn from engagement which content produces more of it. The content that produces the most engagement — outrage, fear, tribalism, the thrill of being right about how wrong everyone else is — is the content the algorithm learns to surface first. Not because anyone decided that outrage should be prioritised. Because outrage generates measurable behaviour, and the algorithm was built to optimise measurable behaviour. The result is a system that systematically amplifies the most emotionally charged content in the feed of every user who has ever clicked on something that made them angry, and then learns from that click that anger is what they want more of.

The king chose stone because stone outlasts memory. Napoleon chose realist painting because realism looked like truth. The empire chose photography because photography had the aura of mechanical objectivity. The state chose poster art because posters could be everywhere simultaneously. Facebook chose engagement optimisation because engagement is measurable, monetisable, and compounds. Each choice produced a recording technology that served the interests of whoever made the choice more than it served the people who encountered the record.

The tool is never neutral. The tool is always the choice of whoever built it, for the purpose they built it for. The algorithm was built to maximise engagement. It is doing exactly what it was built to do. The question is not whether the algorithm is broken. The question is whether the optimisation target was the right one.

In 2020, a team of Facebook researchers produced an internal report on the platform's ranking algorithm. The report was not written by outsiders, critics, or regulators. It was written by the company's own data scientists, using the company's own data, analysing the company's own product. The finding was that the algorithm had developed a systematic tendency to amplify content from what the researchers called "borderline" sources — publishers and accounts that did not quite violate platform rules but consistently produced high engagement through outrage, fear, and tribalism. The content was not illegal. It was not, in most cases, explicitly false. It was inflammatory. It was divisive. It was the kind of content that made people click, share, and comment — and then feel worse about the world and the people in it.

The researchers proposed changes to the ranking system that would reduce the amplification of borderline content. The changes were not radical. They did not require rebuilding the platform. They were adjustments to the weightings the algorithm used to rank content — reducing the value assigned to engagement signals that correlated with divisiveness and increasing the value assigned to signals that correlated with accuracy and informational value.

Some minor changes were implemented. The core ranking system was not changed. Internal modelling showed that the proposed changes would reduce engagement by a measurable percentage — a single-digit percentage, according to subsequent reporting — and that this reduction would translate into a measurable revenue loss. The advertising model that generated Facebook's revenue depended on engagement. The changes would cost money. The decision was made not to make them.

Accuracy and reduced harm on one side. Engagement percentage points on the other. The corporation chose the percentage points.

This is the moment. This is the hand on the chisel. The research exists. The meeting happened. The decision was made. The algorithm continued to amplify. The people who made the decision have names. They are in the record. The record is public — leaked in 2021 by a whistleblower who testified before the United States Congress. The record shows that they knew what the algorithm was doing and chose to let it continue.

The king pressed his handprint into stone. Napoleon pressed his into canvas. The empire pressed its into a portrait. The state pressed its into a poster. Facebook pressed its into a spreadsheet — a projected engagement loss, weighed against a projected revenue loss, resolved in favour of the revenue. The spreadsheet is the stele of the twenty-first century. It is the inscription that records the decision, names the deciders, and preserves the logic that justified it. Future historians will find it and know exactly what happened. They will also know that the people who made the decision had the evidence in front of them and chose the revenue.

The algorithm was built by engineers. The engagement optimisation target was set by product managers. The decision not to change the ranking was made by executives. The advertising model that made engagement the right target was built by a business development team. No single person made the decision that produced the outcome. Every person made a decision that was correct given their incentive structure.

This is not a bug in the system. This is the system.

The handprint in every previous part of this series belonged to someone. The king's hand was on the stele. Napoleon's hand was on the canvas. The empire's hand was on the portrait. The state's hand was on the poster. The photographer's hand was on the shutter. You could find the person, name them, ask them what they had done.

The algorithm's handprint is distributed across a corporate structure that was designed — not accidentally but deliberately — to make accountability impossible. The engineer who wrote the ranking code did not set the optimisation target. The product manager who set the target did not make the advertising model. The executive who approved the target did not write the code. The business development team who built the model did not attend the meeting where the decision was made. Everyone contributed. No one is responsible.

The algorithm has no handprint. That is the point. When no one pressed their hand to the wall, no one can be asked to account for the mark it left.

The same engagement-optimisation logic governs not only Facebook but every major content platform — YouTube, TikTok, X, Instagram, LinkedIn. The academic research on recommendation systems shows consistent patterns across all of them: content that produces strong emotional responses, particularly outrage and fear, is systematically amplified over content that is accurate, nuanced, or calm. The platforms know this. The research is published. The optimisation continues.

In 2021, a team at the Massachusetts Institute of Technology published a study showing that false news on Twitter spread significantly faster, farther, and more broadly than true news — and that the effect was most pronounced for political content. The study controlled for the age of accounts, the number of followers, and the volume of activity. The difference was not explained by the characteristics of the users who shared the content. It was explained by the structure of the network and the design of the platform's amplification mechanisms. The truth was at a structural disadvantage.

In 2023, researchers at New York University analysed YouTube's recommendation algorithm and found that it systematically recommended content from more extreme and divisive sources to users who had watched moderately polarising content — a pattern the researchers called "gradual radicalisation through recommendation." YouTube's own internal research had identified the same pattern years earlier. The recommendation system was not changed.

The decision is being made, every day, in every major technology company, by people with names and salaries and stock options. The handprint is being pressed right now. It does not look like a handprint. It looks like a quarterly earnings report. It looks like a product roadmap. It looks like a normal business decision made by reasonable people under standard market incentives. It is the same decision Napoleon made. The tools are different. The metric is different. The structure is identical: a choice between what serves the record-keeper and what serves the people who encounter the record, resolved in favour of the record-keeper.

The peasant in Part 5 had no name because the state that starved him did not record his name. The person in Part 7 has no name because the platform that shaped what they believe about the world has no record of who they were before the algorithm learned what made them angry. The corporation that made the decision has a name. The person who encountered the record does not.

The algorithm did not need to create false images. It only needed to amplify them. Then the images learned to create themselves.

---

Next in the series: The Synthetic Truth — You are watching a video of a person you know. They are saying something they did not say. You cannot tell. Not from the video. Only from what you know about them — from the trust you have built, the context you carry, the years of witness that no image can replace. The era of visual evidence is over. The era of personal trust has returned.