The Walk I Don’t Take
Why I don’t use social media

I am asked about this often enough that the short answer has worn smooth, and a smooth answer is usually a dishonest one. This is the long version.

It is not a moral position. I am not detoxing, not making a point, and not describing anybody else’s decision. Eight reasons follow. None of them settles it on its own. Together they were enough, and I have tried to put each one no more strongly than the evidence lets me.

Part 1:
It does something, and the something is measurable

The best evidence here is not a survey, but the handful of studies where somebody managed to change the thing and watch what happened.

When Facebook launched in 2004 it opened at one university at a time. Braghieri, Levy and Makarin used that staggered rollout across 775 US colleges, matched against seventeen consecutive waves of the National College Health Assessment and more than 430,000 student responses. Arriving at your campus raised an index of poor mental health by 0.085 standard deviations. Under their assumptions, Facebook’s introduction accounts for around 24% of the increase in severe depression among American college students over two decades. The effect was largest among the students most likely to come off worse in the comparison - those living off campus, those from lower socioeconomic backgrounds, those outside the fraternity system. ¹²

Then there is the experiment. Allcott, Braghieri, Eichmeyer and Gentzkow paid 1,661 people around US$102 to deactivate Facebook for the four weeks before the 2018 midterm elections. Deactivation returned about an hour a day. Subjective wellbeing improved by 0.09 standard deviations. Political polarisation fell by 0.16. News knowledge fell by 0.19 - the cost is real and they report it. Afterwards, use stayed 0.61 standard deviations below the control group, and people’s stated valuation of the product dropped by up to 14% once they had lived without it for a month. ³

Hunt and colleagues went smaller and got the same direction: 143 undergraduates capped at ten minutes per platform per day for three weeks, with significant reductions in loneliness and depression against controls. ⁴

Verduyn and colleagues narrowed it further, and this is the finding that matters most to me. It is not use that predicts the drop in affective wellbeing; it is passive use - scrolling, reading, looking. The effect survives controlling for active use, other platforms and offline interaction, and it is mediated by envy. ⁵

Now the other side, because it is substantial. Orben and Przybylski ran the same adolescent data through thousands of different but equally defensible ways of doing the sums, and found the link between screen use and wellbeing to be tiny - about the size of the effect you would get from eating potatoes. ⁶ Valkenburg, Meier and Beyens read twenty-five reviews of this literature and found that most described the association as weak or inconsistent, while a minority described the identical association as substantial and deleterious. ⁷ Odgers, reviewing the popular version of the alarm in Nature, made the point that the causal story has run well ahead of the causal evidence. ⁸

So the honest summary is this: the average effect is small.

But an average is a claim about populations. A small average can be made up of most people being unaffected and a minority being hit hard, and the studies that pin down the mechanism keep pointing at the same minority - people who compare badly, people who watch rather than talk. ⁹ I already know I sit at the wrong end of several of these. A population average is the wrong number to plan a life with.

One mechanism is missing from all of that, and it is the one most people would name first: fear of missing out. Hunt’s paper is titled for it. It belongs here, but not as another reason, because it is not a harm. It is the retention.

Fear of missing out is a claim about what happens while you are away, which makes the deactivation study the place to test it. Those participants valued the product less after a month without it, and one in twenty were still off it nine weeks later, being paid nothing to stay off. ³ In Hunt’s trial the fear fell just as much in the control group as in the group that had been cut back, which the authors put down to nothing more than having made people watch their own use. ⁴ Whatever it is measuring, it does not survive the absence.

So the comparison and the envy are what the thing does to you while you are there. The fear of missing out is what keeps you there. It can only be disproved by leaving, and it is the reason people do not leave.

Part 2:
What two parliaments have already written down

On 10 December 2025 it became unlawful in Australia for an age-restricted platform to let someone under sixteen hold an account. The obligation sits in the Online Safety Act 2021 as amended in November 2024; platforms must take reasonable steps to prevent under-16 accounts, with penalties up to $49.5 million. No penalties fall on the children, or on their parents. ¹⁰¹¹

The United Kingdom took a different route to a similar place. Ofcom’s Protection of Children Codes under the Online Safety Act 2023 came into force on 25 July 2025, requiring highly effective age assurance and requiring recommender systems to filter specified categories of harmful material out of children’s feeds. ¹²

The part worth reading closely is not the ban. It is the definition.

Australia did not restrict “websites” or “the internet”. The rules name design features: a service whose sole or significant purpose is social interaction, with account-based recommender systems, endless feeds, infinite scroll, and mechanisms that display how many people approved of you. Messaging apps are exempt. Online gaming is exempt. Professional networking, education and health support are exempt. ¹³

That is a legislature writing down, in statute, which part it believes does the damage. Not the content. Not the strangers. The feed, the ranking, and the counter.

I want to be careful about the inference, because the obvious one is a bad argument. “Harmful to children, therefore harmful to adults” does not follow; adults are allowed plenty of things children are not, and the reason is not that those things become safe at eighteen. ¹⁴

The inference I am actually making runs the other way. The law concedes the mechanism. Once two governments have legislated on the premise that an infinite feed plus a recommender plus a visible score is the harmful combination, an adult is entitled to notice that the combination does not change on a birthday. What changes is who decides. A child gets a fence. An adult gets a route.

Which is where I keep landing on the same picture. I would not walk down a badly lit street, in a part of town with a reputation, at two in the morning, for no particular reason. Not because I am certain something would happen. Because there is next to nothing in it for me, and every so often it goes badly wrong for somebody. Nobody has banned the street. I am allowed to walk down it, and some people do, and most of them are fine. That is exactly the point. Protection is something done to you; declining is something you do.

Part 3:
Not posting is not the same as not being used

Fifty-three Australians installed an app called This Is Your Digital Life. Because those fifty-three had friends, the personal information of roughly 311,074 other Australian Facebook users was exposed to it, and from there to political profiling. Meta settled the Information Commissioner’s civil penalty proceedings for $50 million on 17 December 2024 - the largest payment ever made in Australia over the privacy of individuals. ¹⁵¹⁶

The 311,074 did not install anything. They did not consent to anything. They were friends with somebody who clicked.

In April 2020 the US Court of Appeals for the Ninth Circuit reinstated claims that Facebook had continued to receive users’ browsing histories from third-party websites after they had logged out, through embedded plug-ins and cookies that copied the referring URL of every page carrying a Like button. The court held the plaintiffs had adequately alleged a reasonable expectation of privacy against what it described as surreptitious and unseen data collection. Facebook settled for US$90 million in 2022. ¹⁷

And on 1 August 2025 a California federal jury found Meta liable under the state’s Invasion of Privacy Act for receiving reproductive health events - cycle dates, sexual activity, pregnancy intentions - transmitted by a period-tracking app through Meta’s software development kit. The class ran to roughly 38 million people. ¹⁸¹⁹

The fourth case is the one that unsettles me most, because nothing went wrong in it.

Between 2015 and 2019 LinkedIn varied the ratio of weak to strong ties in the connections it recommended to roughly twenty million of its members - about four million in a first wave, sixteen million in a second - and then measured what happened to their employment. The experiments worked, and I come back to what they found in Part 7. Nobody was told they were in one. When the work was reported in 2022 the company said it had acted consistently with its user agreement; its privacy policy did not say that it ran experiments on members. ²⁰²¹²²

There is no breach here, no third party, no litigation and no villain. There is a company that adjusted the job prospects of twenty million people in order to answer a question it wanted answered, and was entitled to do so because those people held accounts.

Four cases, four different situations, one thing in common. In the first the people had installed nothing. In the second they had logged out. In the third they were inside an entirely different application. In the fourth they had done nothing at all beyond existing on the platform. In none of them did it make the slightest difference whether anybody had posted.

Which is why I find “I just don’t post much” an unconvincing answer, including when I used to give it. Declining to post is a decision about publication. It is not a decision about what is taken from you, or what is done to you. The account is the only lever the user actually holds.

Part 4:
More than half of it is not people

Imperva’s traffic analysis puts automated traffic at 53% of all web traffic in 2025, up from 51% in 2024. Human traffic is the minority, at 47%. The same report notes that automated requests increasingly arrive well-formed and successfully authenticated, which is a technical way of saying they are hard to tell from us. ²³²⁴

Ferrara and colleagues made the underlying point a decade ago: social bots are built specifically to pass as human, and their success is measured by how many people never ask. ²⁵

I am not making a claim about how many accounts in any given feed are automated; nobody publishes a clean number for that, and I will not invent one. ²⁴ The claim is smaller than that, and it is enough for me. The reason to be in a room is the people in it. If most of what moves through the building is not a person, and I have no reliable way of telling which half I am addressing, then the thing I came for is not reliably on offer.

Part 5:
More and more of it is not new

An analysis of 55,400 English-language articles of at least a hundred words, sampled from the Common Crawl web archive between January 2020 and March 2026, found the share that is primarily AI-generated rising from near zero to about 36% a year after ChatGPT’s release, 48% two years after, and roughly half from early 2025 onward. ²⁶²⁷

Generated text is a rearrangement of text that already existed. It can be useful and it can be well made; I am not sneering at it. But a rearrangement contains no new observation. Nobody went anywhere, tried anything, changed their mind about anything, or was wrong in a way that taught somebody else something. Whatever else it is, it is not new information entering the world.

The technical version of this is not just a figure of speech. Shumailov and colleagues showed in Nature that models trained on recursively generated data degrade, and that what disappears first is the tails - the rare, the unusual, the specific. The distribution collapses toward its own middle. ²⁸

Which reads as a fair description of what a feed made mostly of recycled material feels like from the inside. A feed that is half rearrangement carries half the new information it appears to carry, and I am spending real attention on it either way.

But halving it is the optimistic reading, and it is not the one the research supports.

Attention does not scale with supply. Simon made the point in 1971, before any of this existed: a wealth of information creates a poverty of attention, and the scarce resource becomes the capacity to allocate it. ²⁹ What follows is not that the good material simply gets a smaller share of the reader. It goes further than that, and a model shows why.

Qiu, Oliveira, Sahami Shirazi, Flammini and Menczer built a social network with finite attention and finite memory and calibrated it against real platform data. As the volume of information rises against a fixed attention budget, what degrades is the system’s discriminative power - its capacity to sort good from bad at all. Under realistic conditions, they found, high-quality information has little advantage over low-quality information. Quality stops predicting what spreads. ³⁰³¹

That mechanism explains the feeling that a feed has got worse without any single thing in it being bad. Nothing needs to be suppressed. The good piece is still there, still true, still well made. It now competes with forty other things for the same three seconds, and the sorting that used to separate them has lost the ability to.

Where the same thing can be measured directly, it shows up. Bevendorff and colleagues tracked 7,392 product-review queries across Google, Bing and DuckDuckGo for a year against a web-scale baseline. All three engines returned search-optimised affiliate content far out of proportion to its share of the web, all three fell to large-scale link-spam campaigns, and the more heavily a page had been optimised, the simpler its content. ³²³¹

So the cost is not that half my attention would be wasted. It is that the half worth having becomes progressively harder to locate inside the half that is not, while the sorting mechanism that was supposed to do the locating is the thing being played. A library in which half the books are photocopies of the other half is not half a library. It is a library with a broken catalogue.

Part 6:
Echo chambers, and the more uncomfortable version

This is the reason I hold most loosely, and I want to say why before I make the case.

The descriptive evidence is strong. Cinelli and colleagues examined more than 100 million pieces of content from over a million users across four platforms. On Facebook and Twitter, users clustered with people who already agreed with them, and material spread mostly among the like-minded; Reddit and Gab did not show the same pattern. A direct comparison of news consumption found higher segregation on Facebook than on Reddit. Design, not human nature, made the difference. ³³

But the interventions do not do what the description implies they should.

Nyhan and colleagues, working with 23,377 consenting US Facebook users over three months of the 2020 election, cut exposure to politically like-minded content by about a third. The median user had been getting 50.4% of their content from like-minded sources against 14.7% from cross-cutting ones, so the change was substantial. The effect on eight measures the researchers had committed to in advance - polarisation, ideological extremity, how people rated the candidates, belief in false claims - was nothing, and the study was big enough to rule out anything larger than a very small one. ³⁴

Bail and colleagues went further and paid people to follow a bot that retweeted the other side for a month. Republicans became substantially more conservative. Exposure to disagreement made things worse, not better. ³⁵

So the simple story is wrong, and I am not going to tell it. Törnberg’s account of the mechanism fits the evidence better: digital media do not isolate people into local bubbles, they sort them into national teams, replacing the cross-cutting disagreements of a neighbourhood with a single alignment that runs the length of the country. ³⁶

Which gives the version I do hold. The harm is not that you only hear one side. It is that the format converts everything into a side. A recommender needs something to rank, and disagreement ranks well. The medium’s lasting effect is not that it fails to change your mind; it is that it teaches you which questions arrive with teams already attached. I cannot prove that with an effect size, and I have marked it accordingly. ³⁷

Part 7:
LinkedIn, and the case against references

LinkedIn is the exception people assume I must make, and it is the one I am most confident about, because I spent thirty years on the other side of it, hiring.

Start with what the platform is. Van Dijck’s analysis is that LinkedIn does not simply host a professional self; it enforces one - a single, consistent, employer-facing identity, with the interface as the ground on which the user, the employer and the platform contest who controls it. ³⁸ The activity that identity mostly generates is watching. And passive watching is precisely the mode Verduyn and colleagues found to lower affective wellbeing, mediated by envy, on a network whose entire subject matter is other people’s advancement. ⁵ Professional voyeurism is not an unkind description of it. It is a fair description of the default use.

Then there is the profile itself, and this is where my hiring experience and the psychometric literature agree.

A profile is a record of outcomes that went well. There is no field for the venture that failed, the role that did not work, the year that went sideways, or the decision you got badly wrong. None of this accuses anybody of lying. The form has only one kind of box.

Guillory and Hancock tested it directly. They had 119 people build résumés offline, on a private LinkedIn profile, or on a public one. Public profiles contained fewer false claims about work history and responsibilities - the parts an employer can verify - but the overall rate of deception did not change. It moved. The misrepresentation relocated to interests and hobbies, the parts nobody checks. Being visible did not reduce the distortion. It moved it to where it could survive. ³⁹

Now put that beside what selection research says about the most familiar version of the same problem: references.

Reference checks have long been reported at a validity of about .26, which is worth one sentence of translation. Validity here is a correlation: how closely the rating a method gives a candidate tracks how well that person actually turns out to do the job. Nought means the method tells you nothing. One means it tells you everything. Nothing in hiring has ever come close to one. ⁴⁰⁴¹

So .26 sits about a quarter of the way along that scale - less than a third of the distance between telling you nothing and telling you everything. Squared, which is the conventional way of asking how much of the variation in real performance a measure accounts for, it is under 7%: better than 93% of what makes one person better at the job than another is invisible to the reference check. Put the other way round, at the only task that matters - take two employees, one of whom genuinely did better, and ask how often the method ranks the right one higher - a coin gets 50% and a reference check gets about 58%. ⁴²⁴³

That is before the correction. Sackett, Zhang, Berry and Lievens showed that the whole field’s estimates had been systematically overcorrected for range restriction, with most validities falling by .10 to .20 once the sums are done properly. In the revised table the structured interview leads the entire discipline at .42, or about 64 correct calls in a hundred, and years of job experience comes in at .07, which is a coin with a slight limp. Reference checks do not appear in the revised table at all; the .26 is the older estimate and, on these authors’ own argument, likelier too high than too low. ⁴⁴

Aamodt’s review sets out why references sit near the bottom of a field that is already not doing well, and the reason is structural rather than moral. The candidate chooses the referee. The referee has an interest, often a legal one, in saying nothing negative. Fewer than 1% of referees rate anybody below average or poor, and two referees describing the same person agree at around .22, where two supervisors rating the same employee agree at around .50. A measure on which everybody scores the same cannot tell anybody apart. It is not that references lie. It is that they are all the same length. ⁴⁵⁴⁶

A LinkedIn profile is that structure taken one step further. The candidate chooses the referee, the wording, the content, the omissions and the timing, and there is no referee at all. Endorsements and recommendations are references written in public, by people the subject selected, who know the subject will read them and can remove them. Whatever leniency does to a private reference, it does more of here.

Which is why I think the profile has worse than no value. A measure on which everybody scores the same is not just useless. It is expensive, because it looks like evidence. Everyone arrives at the interview having “read the profile”, and the thing that gets shortened to make room is the structured interview - the best-validated instrument on the list, and the one that actually requires preparation.

Now the counterweight, and it is a real one. The experiments from Part 3 - the twenty million members whose recommended connections were quietly rebalanced - produced a finding. Rajkumar and colleagues report that the ties do causally produce job mobility, following an inverted U: moderately weak ties transmit jobs best, with larger effects in digital industries. ²⁰⁴⁷ Granovetter said the important half of this in 1973. ⁴⁸ So the claim that hardly any jobs come from it is not what the best available evidence says, and I will not make it.

That is the best evidence anyone has that LinkedIn does something, and it exists because twenty million people were used to produce it without being asked. Both halves of that sentence are true, and I am not going to hold one of them and drop the other.

And the shape of the finding is itself the answer to what the product asks of you. LinkedIn displays your connection count, caps the badge at 500+, and runs People You May Know without pause. The proposition built into all of it is that a bigger network is a better one, and that the way to be found is to know more people. An inverted U says otherwise. It says there is a narrow band that transmits work - ties weak enough to reach somewhere you cannot already see, strong enough that somebody will act - and that both ends of the range do less. Your closest contacts already know everything you know. The furthest ones will not pick up the phone. Most of a large network is doing nothing, and adding to the total does not widen the band.

So the two things the platform invites you to accumulate fail in the same way. The endorsements are references with the referee taken out. The connections are a number whose growth is presented as progress by the company whose own experiment shows that growth is not where the value sits.

What the evidence says is narrower, and more useful to me. The thing that works is the weak tie: a person who knows you exist and roughly what you do. That is an address book function. It does not require living there, posting there, reading it daily, or measuring yourself against it. The mechanism and the habit have been bundled together, and only one of them is doing the work.

Part 8:
Two thousand friends, four of whom would come

Dunbar surveyed two UK samples - 2,000 regular social media users and 1,375 working professionals - and asked them about their Facebook networks. Mean friend counts were 155.2 and 182.8. Respondents counted 27.6% of those as genuine friends. The number they could turn to in a crisis was 4.1. The number they would go to for sympathy was 13.6. Those are the same layers found in offline networks, in the same proportions, and Dunbar’s conclusion was that online media do not lift the ceiling: relationships still cost time, and still need occasional presence to survive. ⁴⁹ Hall’s estimate of what that costs - the hours required to move somebody from acquaintance to friend - explains why the ceiling holds. ⁵⁰

So the platform did not increase how many people you have. It increased how many you can count.

Between the count and the ceiling sits a category the language had no word for, so it borrowed one. Turkle’s argument, which has aged better than most technology criticism, is that this is the actual trade: more contact, less connection, and a growing tolerance for the substitution. ⁵¹

The part that decided it for me is the maths of what it costs. A friendship is partly evidence - somebody spent hours they could not get back, on you, when they had alternatives. A connection costs one click and is maintained by a ranking algorithm. What is left when the cost goes is each of us being seen by the other: I see yours, you see mine, we both get a small return. That is a transaction. It is not a dishonest one, and it is not worthless, but it is not the thing it is named after.

I am not saying that people with eight hundred connections are shallow. I am saying that one word is doing two jobs, and the numbers say most of it is the second job.

In closing

Of the eight, the first and the third carry the most evidence: the effect is real for the people the mechanism selects for, and the account is the only part of the collection I control. The sixth is the one I hold most loosely, and I have said so. The fourth and fifth are not harms at all - they are why the exchange is a poor one even when nothing goes wrong.

But the eighth is the one I actually care about, and it took me a long time to work out why.

There is an old line that true philanthropy is anonymous. The reasoning has nothing to do with modesty. It is a test. If your name goes on the building you have received something back - standing, gratitude, a reputation for generosity - and the moment anything comes back, the gift has become a trade. Anonymity is not humility. It is the only way to prove that nothing returned.

I have a saying of my own that sits beside it, and I have used it for years: true relationships are not transactional. Same test, same shape. What a relationship is worth is whatever survives the removal of the return.

Now try running that test on a platform. You cannot. Every act of connection there is recorded, attributed and counted. The like is receipted. The birthday message is timestamped and witnessed. The comment carries your name and accrues to you. There is no anonymous option, because a public ledger has no column for something given without return - and were there such a column, nobody could see the entry, which defeats the point of writing it down.

None of this says that people on those platforms are insincere. The claim is about what a ledger can hold. Whatever is truly given there is given despite the accounting rather than because of it, and the record cannot tell the two apart. That is the category error underneath the word “friend” on a platform. It names the receipt, not the thing.

Which sends me back to Dunbar’s 4.1. What marks out those four is not that you like them more than the others. It is that they would come without anybody seeing them come - no post, no photograph, no witness, nothing returned. The count climbed to 155 and the four did not move, because the four were never the ones being counted.

I owe a concession here, and it cuts against everything above.

For a lot of autistic adults, these platforms are where the community is. Van Driel and colleagues, interviewing autistic adults about how they actually use social media, describe what the format gives: text instead of faces, asynchronous instead of timed, the ability to compose rather than perform, and access to people who are like you when nobody within reach is. ⁵² I know this is true, because the reasons it works are the reasons it would work for me. For some people the alternative to a bad feed is not a better conversation. It is nothing.

But look at which part of it is doing the work, because it is not the part I have spent eight sections objecting to. What those participants describe - text rather than faces, time to compose an answer, a room where the other people are like you - needs no feed, no recommender and no counter. It is what a group chat does, or a forum, or a Discord server: a small room with a door on it, where the people are known to each other, the conversation has a subject, and nothing is being ranked. Botha and colleagues, interviewing autistic adults about community connectedness, found the benefit in belonging, in particular friendships, and in a shared political purpose - none of which is a feature of a feed, and all of which are features of a room. ⁵³⁵⁴

The Australian legislation makes the same cut, and it is worth noticing that it did. Messaging apps are exempt. Online gaming is exempt. The parliament that decided an endless feed with a recommender and a visible score was harmful enough to keep children away from it also decided that a place where people simply talk to one another was not. That is not an oversight in the drafting. It is the same line I have been drawing for eight parts, arrived at independently and then written into law.

So the concession is narrower than it first looked, and it points somewhere. What helps is the community. What harms is the ledger. They have been sold as a single product, and they are not one.

And the gift passes there too. Somebody who sits up with a stranger in a thread at three in the morning is giving anonymously in every sense that matters, and neither of them will ever be able to show it to anyone. The platform does not record that, and would not know how. My argument is with the ledger, not with the people writing underneath it.

So this is a decision, not a recommendation. It costs me something that plainly helps people I write for and about, and I have chosen to pay that cost in a different currency - long pieces, in a place I control, that anybody can read without an account and without being counted while they do it.

This is also not a break, and there is no date on which it ends. What I declined was not the writing. It was the part where the writing is scored.

---