Take five names from a school register in Glasgow: Murray, Ahmed, Campbell, Kowalski, Brodie. Read them out in alphabetical order, and no one in the classroom will complain. Ahmed comes first and Murray comes last. Computer scientists call this a sort, and an algorithm is simply the set of steps that produces it: compare two names, swap them if they’re in the wrong place, repeat until nothing moves. A child could do it with index cards. Alphabetical order is the most innocent sort there is.

Now change one word. Instead of sorting the children by name, sort them by how much attention each one attracts. The steps are the same, but you have crossed from arithmetic into politics. Someone has to decide what attention means, how to count it, and why it should matter. The computer scientist David Monniaux has warned that whenever a system claims “the machine decided,” people made the choice first.

Democracy has always understood this, because democracy is itself a sorting machine. An election takes millions of competing preferences and puts them in order, and that turns out to be fiendishly hard to do fairly. In the seventeen-eighties, the Marquis de Condorcet showed that a majority can prefer A to B, B to C, and C to A, all at once. In 1951, the economist Kenneth Arrow proved that no ranked voting system can satisfy every reasonable test of fairness at the same time. Every method of counting votes favors someone.

Democracies did not solve that problem. They made it public. They chose a rule, wrote it down, argued about it in parliaments and courts, and changed it when enough people objected. When Scots voted for their parliament this May, they did so under a carefully negotiated system that mixes local winners with regional party lists. Its formula is published, debated, and the same for everyone. Every voter received the same ballot. On it, the candidates’ names appeared in alphabetical order, the same innocent sort as the classroom register.

That same spring, two researchers at the University of Strathclyde, Dayei Oh and Chamil Rathnayake, watched the other sort at work. They created two new TikTok accounts in Scotland. One was trained to look like a voter who avoids the news, watching memes and hobby videos. The other was trained to look like someone who seeks news out. For eleven days before the election, they recorded what TikTok’s For You Page served each account. The difference was stark. Fewer than six per cent of the videos shown to the news-avoiding account had anything to do with the election. For the news-seeking account, the figure was more than eighty per cent.

In other words, the feed did not merely shape opinions about the election. It quietly decided who would notice that an election was happening. The researchers point out that, unlike a public broadcaster, the platform imposes “no civic information floor.” Even the news-seeking account was steered away from the outlets it preferred. Professional journalism made up a small share of what it was shown, outnumbered by influencers, ordinary users, and independent commentators, and the drift away from established sources grew over the eleven days. The study was small and exploratory, as its authors acknowledge. But it captures something larger studies keep finding. The ballot treats every citizen identically. The feed treats every citizen differently, by rules none of them can read.

What those rules reward is no secret. Feeds are sorted to keep people engaged, and anger keeps people engaged reliably. The software engineer Kenneth Reitz argues that this creates a kind of escalation. Once outrage becomes the way to get noticed, yesterday’s shocking claim becomes today’s background noise, and everyone has to shout a little louder. He also notes a distortion in how each side sees the other. The feed surfaces the wildest voices on both sides, so your opponents look more fanatical than they are, and so does your own team. Gloria Marchetti, a legal scholar, points to a quieter effect. When a feed keeps showing you people who agree with you, an unusual view can start to feel like the majority view. Politicians watching the same numbers can mistake a burst of approval online for the will of the public.

For years, critics could say all this, and platforms could reply that it was unproven. That defense is weakening, because researchers have begun building their own sorts. In a study published in Science last year, a Stanford-led team used a browser extension to rearrange the feeds of real users on X during the 2024 election campaign. When posts expressing hostility toward the other party were pushed down, participants felt measurably warmer toward their opponents. When those posts were pushed up, participants felt colder. A team at Northwestern went further. They randomly assigned two thousand Americans to custom feeds for eight weeks before the same election. According to their paper, now in Nature, the feed sorted by engagement amplified moralized, emotional, and hostile content compared with a simple chronological one, and left its users believing their political opponents disliked them more.

None of this makes the sort all-powerful. An earlier experiment with Meta, which switched some users to a chronological feed for three months during the 2020 election, found little change in their political attitudes. Our own choices about whom we follow and what we click narrow our news diets too. But the new evidence points the same way: change what the feed sorts by, and you change what people see, feel, and believe about one another.

That is why the usual remedies keep disappointing. Marchetti examines them in turn, from fact-checking and debunking to media-literacy classes, and finds that each falls short. They correct the results of the sort, one post at a time, without touching what it sorts by. The law is only now beginning to reach that question. Europe’s Digital Services Act requires the largest platforms to offer at least one feed that is not built on personal profiling, and courts have started to enforce it. Last year, an Amsterdam court ruled that Meta broke the law by automatically switching users back to its personalized feed. This spring, a German court ordered TikTok to make its recommendation settings easier to use.

These are real steps, but modest ones. An option buried in a settings menu is not the same as a public rule. A ballot works because everyone receives the same one and anyone can see exactly how it is counted. A feed that lets you opt out of profiling still sorts everyone else in the dark.

Picture a voter in Glasgow this May, standing in a polling booth. On the paper in front of her, the candidates are listed alphabetically, by a rule that took centuries of argument to settle and that applies to everyone equally. In her pocket, her phone holds the other list, of the same names and the same election, sorted by something else entirely. Nobody voted on that order. Most people never see it at all. It may be the most consequential ballot she carries.