Social media considered harmful. Maybe… (4/5): A human social network

Social media considered harmful. Maybe... (4/5): A human social network

Posted on 7 November 2026 • by kLeZ

In 2018 Facebook tried to make its feed healthier by rewarding “meaningful social interactions” between friends and family, and three years later its internal documents revealed that it had achieved the opposite: a feed that gave a louder voice to the angriest. Anyone who designs software gets this story instantly, because in my line of work there’s a saying that sums it up in one line: “when you optimise for a metric, it loses its value”.

In the previous instalments I told you that they knew, that those who tried to stop them had little success, and that today the job of surviving falls to the people who use them. In this instalment I’ll try to describe what a healthy social network would look like, what features it would have and, above all, which ones it wouldn’t have, and why. It’s the idea for a piece of software, dreamt up by someone who designs software for a living, and how such a network would keep itself afloat and who would control it is a problem big enough to fill the final instalment on its own.

§A healthy metric, pushed to the limit

The 2018 change came out of a real problem: users were interacting less and less. To reverse the trend Facebook changed the way posts were ranked and introduced a points system, in which a like was worth one point, a reaction or a reshare without comment was worth five, and a substantial comment or a reshare with some text on top was worth thirty. On paper it was a noble goal, and whoever chose it probably believed in it.

The documents published by the Wall Street Journal in 2021 tell how it went. The company’s researchers had concluded that the weight given to reshares had amplified the angriest voices, and that misinformation, toxic and violent content were disproportionately present among reshares (“Misinformation, toxicity, and violent content are inordinately prevalent among reshares”). When the engineers proposed fixes, Zuckerberg held back some of them, because he feared they would reduce engagement.

Economists have known this phenomenon for fifty years under the name Goodhart’s law. Charles Goodhart, a British economist, observed in 1975 that the indicators used to steer monetary policy stopped behaving as they used to as soon as the authorities started using them as targets. The most frequently quoted formulation, though, comes from an anthropologist, Marilyn Strathern, who in 1997 used it to describe what was happening to British universities subjected to score-based assessments:

When a measure becomes a target, it ceases to be a good measure.

Marilyn Strathern, Improving ratings, European Review, 1997

Facebook’s feed is a textbook example, because meaningful interactions, the moment they became the number to push up, stopped being meaningful and started doing damage all by themselves.

§Engagement yes, maximisation no

People who hang out on the Fediverse, the network of federated social platforms of which Mastodon is the best known, look at the word engagement with suspicion, and they have their reasons. And yet people use social media precisely for what that word describes: to switch off after a gruelling day in cities that aren’t built for the people who live in them, to laugh and keep up with the news, to argue, to stay in touch with those far away and to meet someone they would never otherwise have met. These are healthy needs, and a social network that ignores them stays empty. The damage starts when someone profits from every extra minute and designs the app to squeeze out as many as possible, and the Facebook case shows that to do damage all you need is to maximise, whatever it is you’re maximising.

From this come the two rules on which I’ve built everything else. The first concerns the software, which should make it impossible to exploit the mechanics of engagement, so they can stay without causing harm, because removing them entirely would mean building a social network nobody uses. The second, which for me matters as much as the first and applies to every choice in this instalment and the next, demands that every ban comes with a legitimate, serious alternative that achieves the same healthy purpose. Banning things without offering anything in return is the populists’ shortcut, and it dumps the problem on people who still have to bring home the bacon. Making money is perfectly fine; what a project like this should prevent is getting rich by harming others, in body, in mind or in their relationships.

§How I picture it

The Fediverse is the right direction, even if today it isn’t ready for people who can’t be bothered to study it. The underlying idea is the same as email: nobody owns the network, and there are lots of independent servers, here called instances, each with its own community and its own rules, while an open protocol, ActivityPub, a W3C standard since 2018, lets someone on one instance follow and read someone on another. Different programs, like Mastodon for short texts, Pixelfed for photos and PeerTube for video, already speak the same language, and this leads to a choice I care about a lot: not fencing things off by type, because an illustrator’s drawings, a drummer’s video and the village tourist board’s announcement should all be able to circulate on the same network, each with the tool best suited to showing it.

In the network I have in mind the software would be free, under a licence that requires anyone who modifies it to make the code of their changes available to everyone who uses it over a network, and it would be written and maintained by a non-profit European foundation. Anyone could open an instance: a town council or a university, an association, a cooperative or a company. Some would be certified, meaning they’d comply with a set of rules checked by a body separate from the foundation, and only certified ones could host young people and be shown first to newcomers. Who certifies, who pays the certifier and who stops it becoming a master in its own right are the hardest questions in the whole project, and indeed they’ll get their own instalment.

The first non-choice concerns who owns the network. In a centralised social network the network effect, the fact that a network is worth more the more people are on it, coincides with the size of the company that owns it, and that company has every interest in growing and holding on to people. In a federated network the value lies in the network itself, and no instance would need to become huge for its members to talk to everyone else. For the same reason, turned on its head, there wouldn’t be a state social network either, because whoever governs the network, in my view, should have neither a balance sheet to grow nor power to defend.

If I had to give it a name, I’d call it Puhru. In Akkadian, the language of Babylon and Assur, puḫru was the assembly, both that of the gods, which decided fates, and that of men, which judged disputes and before which, according to the Code of Hammurabi, anyone who had struck a superior in the face was flogged. The house where the assembly met was called bīt puḫri, and I like that between the two words lie the two things an instance really is: a community of people and the place where it gathers.

§Getting in

In 2023 Mastodon tried to solve its best-known problem, namely that newcomers don’t know which instance to choose, in the simplest possible way, by having the app suggest its own instance, mastodon.social, first. As a fix for sign-ups it worked, and it upset those who saw it as letting back in through the window the very centre that federation had thrown out the door. Both sides were right, because an entry point is needed (nobody wants to study a list of servers before they can say hello to a friend), but in my view it should work like Roma Termini, where dozens of trains from different companies arrive and the station belongs to none of them.

In my project the station would be the app, together with a protocol. Each instance would announce its existence to the others, and each would keep an up-to-date list of all of them, with no central registry that someone could switch off or buy. On first launch the app would contact a few instances chosen at random, download the list and present it with what you need to choose: language, main topics, whether it’s free or paid, whether it has ads and how many. Certified instances would carry a mark signed by the certifier, which the app could verify on its own, which nobody could forge and which would expire soon unless renewed, so an instance that lost its certification would really lose it, without needing to tell anyone. Technically there’d be nothing new here, as it’s the same principle as the certificates that protect connections to websites. Newcomers would set a few criteria, and the app would show the matching instances in random order. Size would never count, and a huge instance and a small one meeting the same criteria would have the same chance of appearing at the top, because otherwise the biggest, chosen precisely for being the biggest, would end up as the only train in the station.

But how you arrive matters too, maybe more than the software. You join a social network because the people you know are there, and that’s why I think the most natural way in is by community: a town council opening an instance for its residents, a school, a university, a newsroom, an association or the village tourist board, whole groups that already have a reason to be together. When signing up you should then be able to find your contacts again, from your phone’s address book or from the archive exported from your old social network, without the address book ending up on someone’s servers.

A bridge to the big platforms, on the other hand, I wouldn’t build. It would be convenient and bring in users, but it would also bring in content ranked by their algorithms, and keep people tied to the very place they’re trying to leave.

§The reader chooses the feed

The feed is where the little machine from the first instalment spins fastest, and it’s the feature that needs to be taken apart piece by piece, because it drives engagement.

Users should be able to choose how to rank what they read from among the algorithms that are part of the official software, and an instance couldn’t add one of its own, because a new algorithm would first have to make it into the official software, where anyone can read its code and discuss it, on pain of losing certification. Since 2023 Bluesky has let its users choose between feeds built by anyone, and calls it algorithmic choice. It’s a step in the right direction, but an algorithm nobody checks beforehand remains free to maximise whatever it likes.

More than the list, what matters is the default algorithm, because the vast majority of people never touch their settings. If the instance chose it, it would choose whatever holds on to people most, for the same reason Facebook did. That’s why the default should be set by the software, with no way for the instance to change it, and it should rank posts on explicit signals, meaning who you’ve decided to follow and what you’ve marked as interesting, and never on how long you lingered over a video or how many times you went back to a profile. The difference seems subtle, but an explicit signal is a choice, whereas an implicit one is a reflex, and reflexes are exactly what the little machine feeds on.

The closest case anyone has measured is the chronological feed, tested on Facebook and Instagram in the experiment published in Science in 2023 that I talked about in the last instalment, and those moved onto it spent much less time on the platforms. A ranking based on explicit signals wouldn’t be identical to a chronological one, but it would share its essential point: what decides what you see is what you’ve chosen.

§Finding people and things

A social network that only shows what you’ve chosen risks becoming a closed room, and users should be able to discover new people and things. Discovery would come through three routes: declared interests, closeness in the network (people followed by people you follow) and selections made by flesh-and-blood curators, who could be magazines, newsrooms or simply enthusiasts. There would be no trends, the charts of what’s most popular right now, because a global chart is the perfect target for anyone wanting to manipulate it, and by design it rewards whoever shouts loudest.

Virality, however, would stay, because it’s one of the most human mechanisms there is: something is liked, people share it, and it stops when it stops being liked. What I’d remove is amplification, the system that pushes content in front of strangers because it’s racking up reactions. Sharing would remain a free and unlimited act, and each share would reach your followers, the people who follow the feed of whoever shared it, one hop at a time, each the result of a human choice. The feed could take into account what’s circulating among people close to you in your network, never a global chart. On very long chains I’d add a bit of friction, because it’s precisely among reshares, as Facebook’s researchers had discovered, that misinformation concentrates, and the friction would also stop chain letters, which often carry it.

That leaves the weak spot of a network built this way: distant discovery. A network built on closeness is good at finding what’s close, whereas one of the finest purposes of a social network is connecting people who would never otherwise have met, and as early as 1973 the sociologist Mark Granovetter showed that new information and opportunities more often come from weak ties, acquaintances, than from close friends. That’s why I picture a directory of curators, built like the one for instances, where each declares topics and language and the app suggests them according to the searcher’s criteria, in random order and without size counting, so even the small, faraway curator would get their chance. You could follow topics as well as people, and a topic is global by nature. And two community instances could twin, an idea lifted wholesale from the world of associations, and choose to see each other’s selections, like the brass band of a village in the Apennines with that of a village in Bavaria, or two Linux user groups on two different continents.

§What wouldn’t be there

Some things, in this social network, simply wouldn’t be there, and in my view they weigh as much as the things that are.

There’d be no notifications designed to drag back in anyone who’s drifted away, the ones reminding you how many days it’s been since you logged in, or announcing that someone liked someone else’s post. And it wouldn’t be an option switched off that an instance could switch back on, because the code to send them simply wouldn’t exist in the software. Notifications from people would arrive batched, as in the 2019 study I talked about in the last instalment, except for direct messages and mentions from people you follow, which could stay instant. And at night there’d be silence, for everyone, unless you change it by hand.

There’d be no public counters either. The number of likes under a post and the number of followers under a name would be visible only to their owner, and companies paying for visibility would get theirs privately. In the first instalment I told you about Project Daisy, the hidden likes Meta launched despite knowing from its researchers that they didn’t improve teenagers’ wellbeing, and this might look like the same mistake. Here, though, wellbeing has little to do with it. Without a public score there’s no leaderboard to climb, and nobody would have a number to show off or grow in front of others, which was exactly Facebook’s problem in 2018.

§Advertising, and the shop window

There would be advertising, because someone has to pay for the instances, and because for a business that wants to get known it’s a useful service. But it would never be profiled. The Digital Services Act currently bans advertising based on the most sensitive data and profiled advertising aimed at minors, and here the ban would apply to everyone and to any data. It would always be labelled as such, and would have a cap set per user per day, beyond which the rest of the day would stay ad-free, come what may. Almost everything else hinges on this choice, because with a per-user cap the time spent on the network, past a certain threshold, would stop being worth money, and the very reason for holding on to people would fall away. Television got there long ago, and the European directive on audiovisual services limits adverts to 20% of the time in each of two time slots. The cap, then, would have to be enforced by the software, and how to build it so nobody can cheat is one of the questions for the next instalment.

Then there are people who use social media to sell not a product but their own work, and that’s one of the best reasons social media exists. Frank Matano started in 2008 with pranks on YouTube, where he was one of the first Italians to pass a million subscribers, and from there made it onto television; The Jackal started in Naples as a web video collective. I have three illustrator friends who’ve told me how hard it is to break through in their trade, and how social media has become a work tool for them, like it or not, because they post bits of their artwork there, look for commissions and regular clients there, and spend hours there they’d rather spend drawing. Dario Capacci, Il Batterista Metal on YouTube, plays in Trick or Treat, a long-standing Italian power metal band, and yet in a video from last spring he describes himself as “a 100% content creator drummer”, and the typical day he describes consists mostly of content for social media. For a musician today, making a living from music alone without going through the platforms seems almost impossible to me.

I, for one, am not on Instagram, even though I’d happily follow my friends and the creators I already follow on YouTube, who put out loads of extra content on Instagram that simply never reaches me. More than ten years ago I realised I’m highly vulnerable to the engagement maximisation built into social media, and I know for a fact, because it used to happen to me, that I’d end up scrolling for hours, throwing whole days of my life into the cosmic void. A network like the one I’m describing is one where I could finally follow all these brilliant artists.

Publishing your work should be free for everyone, and anyone paying to get their posts in front of more people would be treated like a business, with labelled ads and within the recipients’ cap. The shop window, which without charts might seem disadvantaged, would rest on declared interests, closeness in the network and above all the curators, and an illustrator would get found by people who follow illustration as a topic or who’ve picked a magazine about comics, without depending on the luck of landing at the top of a chart for a day.

§Protected groups

In civil protection plans the most vulnerable categories are children and the elderly, and a healthy social network should think the same way. In the last instalment I wrote that national bans, as we’re currently writing them, can be dodged with a VPN, and I still think kids should stay as far away as possible from platforms built to hold on to them. A healthy social network could do for them, and for their grandparents, what those platforms can’t be bothered to do, starting with actually knowing how old the people signing up are, without anyone having to hand their ID over to a social network. In Italy, since November 2025, porn sites have had to verify age using the so-called double anonymity laid down in an AGCOM resolution, whereby whoever verifies the age doesn’t know which site it’s for, and the site only learns whether the user is above or below the threshold. In 2025 the European Commission published the blueprint for a verification app built on the same principle, designed to prove you’re over 18 but adaptable, the Commission writes, to other ages, such as 13 or 65 for pensioners, and Italy is among the five countries adapting it first. In this social network verification done this way would be mandatory for every certified instance.

The thresholds would be set by the project’s statute, the same for everyone, and could for example split newcomers into four groups: up to 12, 13 to 17, 18 to 64 and 65 and over. An instance could at most raise the minimum age or the adult age, never lower them or add thresholds of its own, and the adult age could never be below the age of majority in the user’s country, otherwise you could simply declare fifteen-year-olds adults and empty the teenagers’ group. The app wouldn’t ask the verification service “are they over such-and-such?” one threshold at a time, but would send it the statute’s groups all in one go, and the service would reply only with the group the user falls into. That way the protocol could reject any request with too many groups or groups that are too narrow, which would otherwise allow, one question after another, the exact age to be worked out. Then there’s a less obvious risk, linked to time, because an instance that re-ran verification every day would discover the day a kid turns eighteen, and with it their date of birth. That’s why moving from one group to another should only happen when the user asks for it, presenting fresh proof. The system would remain imperfect, because the group is information too, and combined with other things, like being a member of your school’s instance, it says a bit more, but it’s a limitation that can be reduced further, and it’s a far cry from the full date of birth platforms ask for at sign-up today.

Children, the first group, would stay out, and for them the software should do nothing but say no. Teenagers could only sign up to certified instances, themselves federated with certified instances, and there they’d have a private profile, be contactable only by people they’ve accepted, not appear in suggestions to adults outside their network, receive no advertising and have a fixed night-time silence that can’t be removed. Deliberately missing would be the daily time limit, which does appear in Meta’s 17 billion settlement and in the Kids Act proposal. A time limit treats a symptom, namely an app that’s hard to put down, and in a social network that doesn’t rank the feed on reflexes, doesn’t send notifications to lure you back and has no leaderboards to climb, that symptom would, at least in intent, already have been pulled out by the roots.

The elderly, on the other hand, are adults in every respect, which is why their protections would be on by default but could be switched off, unlike the teenagers’, and whoever switches them off takes a risk, but does so knowingly, and awareness is what counts. There’d be three protections, designed against scams and misinformation: more caution about contact from strangers, certain categories of advertising excluded, such as financial products and “miracle” health products, and a bit more friction on reshares. The reason is that, according to a 2023 report by the Polizia di Stato, among the factors that make the elderly more vulnerable are loneliness and a strong need to communicate, and among the ways of scamming people online the same report cites romance scams and trading scams, which almost always start with an unexpected contact.

§Who moderates

Moderation would be done by people, and every certified instance would have to have moderators. When an instance hosts abuse, it would be the others that isolate it, all together.

The software, on the other hand, wouldn’t surveil anyone, and wouldn’t scan content looking for illegal material. It’s a delicate non-choice, because it touches on child abuse. Someone with a server and depraved friends will always exist, and what stops them, in my view, is social services, the family psychologist, relationships and sex education in schools and, when needed, the law and the police, far more than any algorithm. Thinking a social problem can be solved with a program is the same technological illusion today’s social networks were born from.

§For everyone

An ethical social network worthy of the name should be easy to use and accessible to everyone, from able-bodied people with high-end devices to disabled people with low-end ones, from first-world fibre connections to third-world GPRS connections, and it should be translated into every language, from Afrikaans to Zulu (the fact that the first and last languages on the list are both African is just a curious coincidence). I’m deliberately leaving this part at the level of principles, because if I designed it in detail I’d end up actually building this social network, and yet it comes before everything else, because a healthy social network that’s a struggle to use stays as empty as one that ignores its users’ needs.

§Why people stay

People stay for other people: friends, distant relatives, the five-a-side lot, the village tourist board’s feed announcing the summer fête. They stay because it’s useful, to find out when the main road’s closing or to find someone to fix a bike. And they stay to unwind, which is a serious matter, with comedy, sketches, memes and satire, and then with art and culture in all their forms, which have found an audience on social media they never had before. These are the same reasons so many people joined Facebook twenty years ago, before someone decided to sell every minute of their attention.

§The bill

That leaves the question anyone who’s worked at a company in this industry will have had in their head since halfway through the article: who pays? A social network without profiled advertising, without time to sell, with hidden counters and flesh-and-blood moderators, still has to pay for servers, people and electricity at the end of the month. And if an instance starts cheating on the ad cap, who notices? Who stops the foundation writing the software from becoming, within ten years, another Meta? These are the questions I’ve spent the most time on, and the ones that have demanded the most compromises.

The next instalment is all about them, with two bodies keeping each other in check, an ombudsman who doesn’t pocket a penny, a formula that makes those who grow too much pay more, precise obligations for moderators, and the story of those who tried before and came to a halt, from App.net, which shut down in 2017, to IFTAS, which in 2025 had to switch off a large chunk of its safety services for the Fediverse because the money had run out.

Alessandro 'kLeZ' Accardo personal website