Here’s a problem that sounds almost insultingly simple. On the internet, how do you make sure one person is just one person?

💥 Master AI & Tech Skills
🎓 Get Up to 50% OFF Premium Courses
⏰ Limited-Time Offer
👉 Enroll Now & Start Learning

Not their name. Not their identity. Just the raw count: that the “thousand people” signing a petition are a thousand people, that the “five-star reviews” came from five hundred different humans and not one bot farm, that the account arguing with you is a separate person from the ten others making the same point. One human, one account. How hard can it be?

As it turns out, it may be the single hardest unsolved problem in online safety, and once you understand why, you start seeing it everywhere. Spam, fraud, fake reviews, bot swarms, manipulated votes, astroturfed outrage, they are, underneath, all the same problem wearing different outfits. This is the case for why “one person, one account” is so brutally difficult, why it quietly underpins so much of what’s wrong online, and why the most promising answer is also still very much a work in progress.

The problem has a name, and a surprisingly old one

Computer scientists call this the Sybil problem, and it’s been understood for a long time. The term was coined by a Microsoft researcher in the early 2000s, named after a famous psychology case study of a woman with dissociative identity disorder, one person presenting as many identities. That’s the whole attack in a phrase: a single entity masquerading as a crowd.

A Sybil attack is simply when one actor creates many fake identities to gain outsized influence over a system. And the goals are familiar: flood a platform with spam, swing a vote, manufacture consensus, drain a reward meant to be shared, or drown out real people. The uncomfortable insight is that any system that treats “number of accounts” as a proxy for “number of people” is vulnerable the moment creating accounts is cheap. Which, online, it almost always is.

Why it’s so much harder than it sounds

The reason one-person-one-account is so stubborn comes down to a genuine dilemma, and every attempted solution so far has run into one side of it or the other.

To guarantee that each account maps to a distinct human, you need to verify something about the person that can’t be cheaply duplicated. But the moment you start verifying people, you hit a brutal trade-off:

• Verify nothing, and accounts are free to mass-produce. This is the open internet’s default, and it’s exactly why Sybil attacks thrive. Anyone can be anyone, and everyone can be thousands.
• Verify identity (demand a government ID, a real name, a phone number), and you solve the counting problem but create three new ones: you build a privacy nightmare and a honeypot of personal data, you exclude the billions without the right documents, and you turn a simple “are you a unique human” check into full-blown surveillance.

That’s the trap. For decades the only reliable way to stop one person being a thousand was to make every person prove exactly who they are, which is a cure most people, rightly, don’t want. So the internet mostly chose the other horn: stay open, and live with the Sybils.

And the patches we layered on top never quite worked. CAPTCHAs were meant to prove humanness, until bots got better at them than humans. Phone-number and email verification just raised the price of a fake account slightly, so attackers bought numbers in bulk. Behavioral detection and algorithmic monitoring help catch some fake nodes, but it’s an arms race, and the fakes keep getting more convincing, especially now that AI can generate plausible faces, text and activity at scale. Every defense treats a symptom. None addresses the root: the system still can’t actually tell whether there’s a unique person there.

Watch the whole problem collapse into one

This is the part worth sitting with if you work anywhere near trust and safety, because it reframes a dozen separate headaches as a single structural flaw.

• Spam is one entity posting as many. A Sybil problem.
• Fake reviews and ratings are one seller manufacturing many “customers.” A Sybil problem.
• Bot armies and engagement farms are one operator running thousands of “users.” A Sybil problem.
• Vote and poll manipulation is one actor casting many ballots. A Sybil problem.
• Astroturfing and influence operations are a few people simulating a grassroots crowd, even state-backed “bot farms” spreading political content have been documented doing exactly this. A Sybil problem.
• Airdrop farming and promo abuse is one person claiming a per-user reward many times over. A Sybil problem.

Every one of these is usually owned by a different team with a different tool, moderation here, fraud there, bot detection somewhere else. But they share a single root cause: the internet cannot reliably enforce one-person-one-account. Solve that, and you don’t fix one of these. You make the entire category structurally harder to pull off at once. That’s what makes it the hardest and the most valuable problem in the space.

A Structural Answer

If the dilemma is “verify nothing versus verify identity,” the breakthrough is realizing there’s a third option that nobody could build until recently: verify personhood and uniqueness without verifying identity.

This is the idea behind proof of human (sometimes called proof of personhood). Instead of asking “who are you?”, it asks only “are you a real, unique person?”, and answers it without collecting your name, documents, or personal data. It confirms the two things the Sybil problem actually turns on, that you’re real (not a bot) and that you’re one (not a thousand), and deliberately stops there.

What makes this genuinely new is the privacy engineering. Using techniques like zero-knowledge proofs, a person can prove they hold a valid proof of human without revealing anything else about themselves, and modern designs keep the underlying data on the person’s own device rather than in a central store. That’s the move that escapes the old trap: you get the uniqueness guarantee of identity verification without the surveillance, the honeypot, or the exclusion. Proof of human checks are increasingly recognized as a core defense precisely because, unlike a CAPTCHA, they verify a unique individual is behind each account without demanding sensitive ID.

The most prominent effort building this is World, whose proof-of-human network is designed to let a person confirm they’re a real, unique human, once, then prove it anywhere without handing over who they are. It’s one concrete attempt at the third option, a verification that’s human-first rather than identity-first, which is exactly the shape the Sybil problem has been waiting for.

The honest part: this is still early

Thought leadership that oversells its solution isn’t worth much, so here’s the straight version: proof of human is the most structurally promising answer to the Sybil problem, and it is also early, unproven at full scale, and genuinely hard to get right.

A few honest caveats:

• Adoption is the mountain. A one-person-one-account layer only works where it’s actually used. Getting platforms, apps and services to adopt a shared primitive is slow, political and far from guaranteed. The idea can be sound and still take years to matter.
• Verifying a unique human doesn’t make them honest. Proof of human stops one person from being a thousand; it doesn’t stop a verified person from behaving badly. It’s a Sybil-resistance primitive, not a trust-or-morality guarantee. It raises the cost of mass fakery from near-zero to real, which is the win, but it’s not a cure-all.
• The enrollment problem is real. Any system that verifies uniqueness has to actually reach people to do it, and doing that globally, fairly and accessibly is a genuine challenge, not a footnote.
• It has to stay private to be worth it. A proof-of-human layer that quietly became an identity-tracking layer would reintroduce the exact problem it was meant to solve. The privacy properties aren’t a feature; they’re the whole justification.

None of this is a reason to dismiss it. It’s a reason to be clear-eyed: we may finally have the right shape of solution to a decades-old problem, and we’re still in the hard, unglamorous phase of trying to make it real.

The bottom line

One person, one account. It’s the most deceptively simple-sounding requirement on the internet, and the fact that we still can’t reliably enforce it is the hidden reason so much of online life feels gameable, spammed and faintly fake. Nearly every large-scale abuse online is a Sybil attack if you squint, one actor wearing a thousand masks.

For most of the internet’s history, the only fix was to unmask everyone, which was a cure worse than the disease. Proof of human is the first approach that credibly offers the uniqueness without the unmasking, confirming you’re a real, singular person while letting you stay private.

It won’t arrive overnight, and it won’t fix everything even when it does. But if the hardest problem in online safety is counting people correctly without surveilling them, this is the first answer that’s even the right shape. And for a problem this old, that’s a bigger deal than it sounds.

Thank you for being a part of the community

Before you go:

👉 Be sure to clap and follow the writer ️👏️️

👉 Follow us: Medium

👉 CodeToDeploy Tech Community is live on Discord — Join now!

Disclosure: This post includes affiliate and partnership links.

The Sybil problem: why one person, one account is so hard to guarantee online was originally published in CodeToDeploy on Medium, where people are continuing the conversation by highlighting and responding to this story.

Open WhatsApp and you’ve probably seen it: that little grey line at the top of a chat saying “Messages are end-to-end encrypted.” You’ve seen it on Signal, on iMessage, in the settings of half the apps on your phone. You nod, you feel vaguely safer, and you keep typing.

💥 Master AI & Tech Skills
🎓 Get Up to 50% OFF Premium Courses
⏰ Limited-Time Offer
👉 Enroll Now & Start Learning

But have you ever actually stopped to ask what that means? Like, mechanically, what is happening to your “running late lol” so that supposedly nobody but your friend can read it? And, the bigger question, why do we all just trust that little line without any way to check it?

Let me explain the whole thing the way I’d explain it to a friend over coffee. No jargon walls, no math degree required.

Where you’ve seen it (and where it quietly went missing)

End-to-end encryption, “E2EE” if you want to sound fancy, is everywhere now, but unevenly.

Apps that have it on by default: Signal (the gold standard, more on that later), WhatsApp, and Apple’s iMessage (for blue-bubble, iPhone-to-iPhone chats). Every message, on automatically, you don’t have to do anything.

Apps where it’s there but NOT on by default, which trips people up: Telegram is the big one. Most people assume Telegram is super-secure, but its normal chats are not end-to-end encrypted, they sit readable on Telegram’s servers. You only get E2EE if you manually start a “Secret Chat.” Facebook Messenger was the same for years before recently switching it on by default.

The famous one from the past: BlackBerry. Back when every executive and world leader carried one, BlackBerry’s claim to fame was its locked-down encrypted messaging, so secure that several governments threatened to ban BlackBerries because they couldn’t read what people were sending. It’s a big reason “encrypted phone” entered the public imagination.

The gaps that surprise people: iMessage to an Android phone? That green bubble drops to old-school SMS, which has no E2EE at all. And Telegram’s default chats, as mentioned. The label matters, “encrypted” doesn’t always mean “end-to-end encrypted,” and the difference is everything.

Who actually invented this (spoiler: not BlackBerry)

Here’s a genuinely great story most people don’t know.

End-to-end encryption for ordinary people was born in 1991, and BlackBerry had nothing to do with it. It came from one guy: Phil Zimmermann, a software developer and anti-nuclear peace activist who was worried that as computers took over communication, ordinary people would lose the ability to have a private conversation.

So he built a program called PGP, short for “Pretty Good Privacy” (yes, really, that’s the actual name), and gave it away for free on the early internet. His whole philosophy, in his own words, was that privacy is something people should be able to take into their own hands.

And then things got wild. The US government treated strong encryption like a weapon, literally, it fell under arms-export laws, and launched a three-year criminal investigation into Zimmermann for letting PGP spread outside the country. His genius workaround: he published PGP’s source code as a physical book through MIT Press. Software was restricted, but a book is protected free speech under the First Amendment. Anyone abroad could buy the book, scan it, and rebuild the software. The case was eventually dropped, and the idea was out of the bottle forever.

That fight, in the 1990s, is now called the “Crypto Wars,” and Zimmermann basically won it for the rest of us. Every encrypted app you use today is a descendant of what he started. (BlackBerry, for the record, came later and used this kind of encryption; it didn’t invent it.)

The thing that makes it work: two keys, not one

Okay, the actual mechanism. This is the clever bit, and it’s more intuitive than it sounds.

Old-fashioned encryption used one secret key to both lock and unlock a message. The problem is obvious once you see it: how do you get that key to the other person? If you send it over the internet, anyone snooping can grab it, and now they can unlock everything. It’s like mailing someone a locked box and having to mail the key in a separate envelope that could just as easily be stolen.

End-to-end encryption solves this with a beautiful trick called public-key cryptography, which uses two keys that work as a pair:

• A public key, which you hand out to everybody. It can only lock messages.
• A private key, which never, ever leaves your device. It’s the only thing that can unlock messages that were locked with your public key.

Here’s the magic: something locked with your public key can only be opened by your private key. Not even the public key that locked it can undo it.

So picture it like this. Your public key is an open padlock you hand out freely. Anyone can snap it shut on a box to send it to you. But you’re the only person on earth with the key that opens that padlock, and that key never leaves your pocket. People can send you locked boxes all day; only you can open them.

So what actually happens when you hit send

Let’s trace your “running late lol” from your phone to your friend’s, step by step:

1. Your app takes your friend’s public key (which your phone already grabbed) and uses it to scramble your message into unreadable gibberish, right there on your device, before it goes anywhere.
2. The scrambled gibberish travels across the internet, through your Wi-Fi, through your phone company, through the messaging company’s servers.
3. Every single one of those stops, your ISP, the app’s own servers, any hacker in between, sees only the gibberish. None of them has your friend’s private key, so none of them can unscramble it.
4. The message arrives at your friend’s phone, where their private key, sitting safely on their device, unlocks it back into “running late lol.”

The “end-to-end” part is exactly this: the message is only ever readable at the two ends, your phone and your friend’s phone. Everywhere in between, including on the company’s own servers, it’s locked. That’s the whole idea in five words: only the two of you can read it.

But how do we trust it? We can’t see any of this

This is the question I think about most, and it’s a fair one. You can’t see the keys. You can’t watch the scrambling happen. You’re taking that little grey line entirely on faith. So why is that faith mostly justified?

A few real reasons:

• The math is public and tested. The encryption methods aren’t secret, they’re published, and thousands of cryptographers worldwide have spent decades trying to break them. When something holds up to that much expert attack, that’s real evidence, not marketing.
• The best apps are open source. Signal’s code is public. Anyone can inspect exactly what it does and confirm it’s not cheating. You don’t have to trust their word; independent experts check it, repeatedly.
• They can be audited. The Signal Protocol (the encryption engine behind Signal and, quietly, WhatsApp and others) has been independently audited multiple times.

But here’s the honest catch, and this is the part the marketing doesn’t emphasize: you’re still trusting the app itself. E2EE protects the message in transit. It does nothing about what happens on the two ends. If the app secretly kept a copy before scrambling, or after unscrambling, encryption wouldn’t save you. This is why which app matters so much, and why open-source, audited apps like Signal earn more trust than a closed app that just says “trust us.”

And there’s a second catch worth knowing: metadata. Even with perfect E2EE, the company can often still see who you messaged, when, and how often, just not the contents. For a lot of situations, that “who and when” is revealing enough. Signal is unusual in going out of its way to collect almost none of it; most others happily keep it.

Is it actually strong? How it compares

Short version: proper end-to-end encryption is about as strong as everyday security gets. The encryption itself isn’t the weak point, it would take longer than the age of the universe to brute-force with today’s computers.

But “encrypted” comes in different flavors, and they’re not equal:

• No encryption (plain SMS): your text messages. Wide open. Your phone carrier can read them, and they’re trivially intercepted. Treat SMS like a postcard.
• Encryption in transit only: this is what a lot of services use. Your data is locked between your device and the company’s server, but the company can read it once it arrives (and so can anyone who breaches the server). Better than nothing, but the company holds the keys.
• End-to-end encryption: only you and the recipient hold keys. The company itself can’t read your messages even if it wanted to, or was forced to. This is the strong one.

The jump that matters is from “in transit” to “end-to-end.” With in-transit encryption, you’re trusting a company not to look and not to get breached. With E2EE, you don’t have to, because they can’t look. That’s the whole point, it removes the need for trust in the middleman.

Do apps still use it, or something of their own?

Mostly, they use one very good thing, sometimes with their own wrapper around it.

The dominant standard today is the Signal Protocol, built by the non-profit behind the Signal app. It’s so well-regarded that it became the quiet backbone of the whole industry. WhatsApp uses the Signal Protocol. Google Messages uses it for encrypted texts between Android phones. Lots of others build on it too. So when people debate “Signal vs WhatsApp,” a funny truth is they’re running the same core encryption, the difference is everything around it (WhatsApp, owned by Meta, collects far more metadata; Signal collects almost none).

Apple’s iMessage uses its own homegrown encryption rather than Signal’s, and it’s strong, with the well-known catch that it only applies Apple-to-Apple. The industry is slowly fixing the cross-platform gap: Apple and Google have committed to bringing E2EE to RCS (the modern replacement for SMS) so iPhone-to-Android chats can finally be encrypted too, though at the time of writing it hasn’t fully shipped.

So the trend is clear: rather than everyone inventing their own, most serious apps have converged on the same audited, trusted standard, which is genuinely good news for you, because it means the protection is battle-tested rather than some company’s untested homebrew.

The short version, to take with you

• End-to-end encryption means only you and the person you’re talking to can read your messages, not the app, not your ISP, not a hacker in the middle. It’s locked everywhere except the two ends.
• It works with two keys: a public one that locks (shared freely) and a private one that unlocks (never leaves your device). Think open padlock out in the world, only-key safe in your pocket.
• It was invented in 1991 by Phil Zimmermann with PGP, not BlackBerry, and the US government tried to prosecute him for it. He won, and we all inherited private messaging.
• You trust it because the math is public, tested, and (in the best apps) open-source and audited, but you’re still trusting the app on each end, so pick a good one.
• The gold standard today is the Signal Protocol, which even WhatsApp and Google use under the hood. “In transit” encryption trusts a company; “end-to-end” removes the need to.

Next time you see that little grey line, you’ll actually know what it’s promising, and roughly how it keeps that promise. It’s not magic. It’s just a very clever idea about padlocks, one that a peace activist gave the world for free, and had to fight his own government to do it.

If you enjoyed the read, do consider clapping and dropping a follow.

• Your AI assistant is about to start buying things
• Can you trust anyone on a dating app anymore?

Thank you for being a part of the community

Before you go:

👉 Be sure to clap and follow the writer ️👏️️

👉 Follow us: Medium

👉 CodeToDeploy Tech Community is live on Discord — Join now!

Disclosure: This post includes affiliate and partnership links.

How does end-to-end encryption actually work? was originally published in CodeToDeploy on Medium, where people are continuing the conversation by highlighting and responding to this story.

There is a photo comparison that floats around the internet every so often. On the left, a cast-iron street lamp from around 1900: fluted column, curling scrollwork, a little crown of ornament up top. Someone designed that. Someone sat down and cared about a pole whose entire job was to hold up a light.

On the right, a street lamp from today: a galvanized metal tube. Straight. Gray. Efficient. And to be fair, it is genuinely better. Cheaper to make, easier to fix, more of them for the same money. It holds the light up just fine.

But look at the two of them side by side and you feel it immediately. Something is gone.

I cannot stop thinking about those two lampposts lately, because the exact same thing is happening to the internet right now, except instead of taking a century, it is taking about eighteen months. We are living through the great flattening of web design, and the tool doing the flattening is the one everyone is busy celebrating.

You already know the look

You have seen it a thousand times this month without naming it. Let me name it.

A soft purple-to-blue gradient in the hero section. The Inter font, or something almost identical to it. A big centered headline making a vague, confident promise. Below it, four cards in a neat grid, each with a little rounded icon, a two-word title, and a sentence of filler. A faint hover animation if you are lucky. Lots of whitespace. A call-to-action button with rounded corners.

That is the house style of the entire modern web, and once you see it you cannot unsee it. As one writer put it, you can now identify one of these pages from across the room.

It is not a coincidence. It is not a trend the way flat design or skeuomorphism were trends, choices that spread because people liked them. This is something different. This is the statistical average of every website that came before, served back to us as a default, at a scale and speed nothing in design history has ever matched.

Where it came from: the “vibe code” era

The term for how these sites get made is vibe coding, coined in early 2025. The idea is simple and genuinely kind of wonderful: you describe what you want in plain language, an AI writes the code, and you accept it without reading every line. No CSS, no design school, no agency invoice. Anyone can build a website now. Anyone at all.

I want to be fair here, because that is a real gift. The barrier to putting something on the internet used to be high, and knocking it down let a lot of people build things they never could have before. That part is good. I am not nostalgic for the days when you needed to know flexbox to exist online.

But here is the problem nobody wants to say out loud.

When you ask an AI to build you a website and you accept what it gives you, you do not get your taste. You get the average of everyone’s taste. These models are trained on the existing web, so they reproduce the most common patterns they have seen. Ask for “a clean, modern landing page” and the model returns the median of a million clean modern landing pages. Purple gradient. Inter font. Four cards. Every time.

Researchers at the University of Washington actually studied this. In a March 2026 paper with the very polite title “Interrogating Design Homogenization in Web Vibe Coding,” they examined the major AI website tools and found exactly what you would expect: because the tools are trained mostly on English-language, Western design conventions, they reproduce those same patterns over and over, narrowing the range of what the web looks like.

They gave the trap a name, and it is the sharpest phrase in this whole discussion.

The “good enough” trap

The researchers call it the “good enough” trap, and understanding it explains everything.

Here is how it works. The AI generates something that looks polished. Clean lines, nice spacing, professional-looking. And because it looks polished, the person building it, who often does not have a design background, accepts it. Why wouldn’t they? It looks fine. It looks like a real website.

There is no moment of friction. Nothing in the process stops you and asks: does this actually reflect you? Does it fit your idea, your audience, the specific strange thing you are trying to say? The tool is optimized to remove exactly that kind of friction, because friction feels like failure. Fast and frictionless is the entire selling point.

So the median output ships. Not because anyone decided it was the best design, but because it was good enough and asking for better would have been work.

Multiply that decision across millions of people and you get the beige internet we now live on. Nobody chose it. Everybody just accepted it.

This is where the lampposts come back

The reason the lamppost comparison stuck in my head is that it is the same mechanism, just slower.

The ornamented 1900 lamppost was not replaced because people decided ornament was bad. It was replaced because a plain metal pole was more efficient. Cheaper, faster, scalable. Every individual decision to use the plain one was rational. The ornament was a cost, and costs get cut.

But nobody ever stood up and voted for a world with less beauty in it. It happened one reasonable, efficient decision at a time, and one day you looked up and every lamppost was a gray tube and the craft that made the old ones was simply gone.

That is what “good enough” does over time. Each individual choice is defensible. The aggregate is a loss nobody consciously chose.

Vibe coding is that process at hyperspeed. Every purple-gradient site is somebody making the rational, efficient choice to accept the good-enough default. And the aggregate is a web that is losing the strange, personal, hand-made quality that made the early internet worth wandering around in.

What we are actually losing

It is tempting to wave this away as an aesthetics-snob complaint. Who cares if websites look samey, as long as they work?

But it is not only about looks, and this is the part I most want to land.

Design is communication. A generic layout communicates something whether you intend it to or not. It says: I built this fast and hoped you would not notice. It says there was no particular person behind this with a particular point of view. When every site looks interchangeable, looking interchangeable becomes a signal, and the signal is “no one cared enough here to be different.”

Sameness erodes trust and attention. There is a real, measurable cost to being pattern-matched as machine-made. The moment a visitor clocks your site as generic AI output, they assign it less credibility and less effort, and they scroll on. The design being “fine” does not save you, because fine is now the background noise everyone has learned to ignore.

It flattens more than pixels. The homogenization runs deeper than color and font. When you let the AI make the structural decisions, the layout, the hierarchy, what goes where, you are outsourcing the thinking, not just the styling. And those decisions are based on statistical averages of what already exists. So you do not just look like everyone else. You are quietly nudged into thinking about your own project the way the average project thinks about itself. The creativity does not just leave the surface. It leaves the concept.

That is the real cost. Not ugliness. Ugliness would at least be interesting. The cost is a kind of confident, polished blandness that actively punishes anyone trying to do something genuinely new, because new does not match the training data and therefore never shows up in the default.

The tools are not the problem

I want to be careful not to turn this into “AI bad, go back to hand-coding.” That is lazy, and it is wrong.

The tools are extraordinary. The access they created is real and worth protecting. The problem was never that a machine can build a website. The problem is that we have collectively decided the machine’s first, most average guess is the finished product.

The same University of Washington researchers who named the “good enough” trap also proposed the antidote, and it is a good one. They call it productive friction: deliberately adding back the small moments of resistance that make you stop and actually choose, instead of just accepting whatever showed up first. Push back on the default. Ask for the weird version. Treat the AI’s output as a first draft to argue with, not a finished thing to ship.

How to not look like everyone else

Here is the genuinely practical part, the stuff you can do this afternoon to make your site stand out. None of it requires being a designer. It just requires refusing the first answer.

Kill the default font. Inter is fine. Inter is also the single biggest tell that a machine picked your typography. Swap it for almost anything with a bit of character, a distinctive serif for headings, a less common sans, even a slightly odd pairing. This one change alone breaks the spell more than any other, because font is the first thing your eye reads as “generic AI.”

Pick a color that makes you slightly nervous. The safe purple-blue gradient is safe precisely because it is everywhere. Choose a palette with an actual point of view: a weird green, a muted rust, a real black instead of the standard soft gray. If the color feels a tiny bit risky, you are probably onto something.

Break the four-card grid. The moment you have “three or four things in a neat row of rounded cards,” you are in template territory. Lay them out asymmetrically. Make one bigger. Stack them oddly. Let something overlap. The grid is comfortable, which is exactly why it is invisible.

Write like a person, not a landing page. The default AI copy voice, competent, balanced, faintly enthusiastic, is as recognizable as the visuals. Rewrite it the way you would actually say it out loud. Use a real sentence with a real opinion. Say the specific thing instead of the safe general one. Your words are the cheapest, fastest way to sound human, and almost nobody bothers.

Add one thing that is just… yours. A strange illustration. An odd little detail. A joke. A photo you actually took instead of a stock image. One genuinely personal choice signals a human was here more than a hundred polished-but-generic sections ever could.

Ask the one question the whole system is built to make you skip. Look at what the AI handed you and ask: is this actually me, or is it just the statistical average of everyone who came before me? If it is the average, that is your cue to change something. Anything. The goal is not perfection. It is a fingerprint.

The through-line here is simple: the tools are designed to remove friction, and friction is exactly where taste lives. So you have to add a little back on purpose.

There is still hope

Here is what gives me a little hope. The flattening is so total that difference has become cheap and powerful.

When everything looks the same, the smallest amount of genuine intention stands out enormously. A site with one real idea in it, one unexpected choice, one piece of actual personality, now reads as a small miracle simply because it broke the pattern. The bar for standing out has never been lower, precisely because so few people are bothering to clear it.

The early web was a mess. It was full of ugly, broken, gloriously personal pages that looked like nowhere else because a specific human made each one with their own hands and questionable taste. We lost some of that to templates, and we are losing more of it to the median-machine. But it is not gone, and it is not inevitable that it goes.

Those two lampposts are not really about lampposts. They are about what happens when we let efficiency make every decision for us, and forget that “good enough” is a description of a floor, not a goal. The web does not have to be a horizon of identical purple gradients. It only becomes that if we keep accepting the first thing the machine hands us.

So don’t. Argue with it. Make something a little strange. The internet has enough gray metal poles.

P.S: This is the coolest website in recently

Shopify Editions | Winter '26

Why does Every Website looks the same? was originally published in The Repo on Medium, where people are continuing the conversation by highlighting and responding to this story.

Most advice about identity theft skips the part that would actually help you: what the criminals are doing, step by step, and why. It jumps straight to “use strong passwords” without explaining the machine those passwords are supposed to defend against.

So let’s do it the other way round. Here is how identity theft actually works in 2026, from the moment your information is stolen to the moment someone opens a loan in your name. Then, once the mechanics are clear, the controls that genuinely stop it, and the reason a couple of them matter far more than the rest.

The uncomfortable starting point: your personal data has almost certainly already been stolen. The question is no longer whether it is out there, but whether the specific pieces needed to hurt you have been assembled and sold yet.

Step one: how your data gets harvested

Identity theft starts with raw material, your personal information, and there is an astonishing amount of it in circulation.

One security firm recaptured 53.3 billion distinct identity records circulating online in 2024 alone, a 22% jump on the year before. The average person now has around 229 of their own records exposed on the dark web: names, dates of birth, phone numbers, and in many cases Social Security numbers and bank details. That is not a risk of future exposure. That is the current state.

The data gets harvested in a few main ways, and the mix has shifted recently.

• Infostealer malware. This is now the leading way credentials get stolen, and it is nastier than a big corporate breach. A piece of malware quietly sits on someone’s own device and harvests everything: saved passwords, browser cookies, session tokens, autofill data. These “stealer logs” get sold on criminal markets within hours, sometimes before the victim has any idea they were infected.
• Data breaches. The classic route. A company gets compromised and millions of customer records spill out. What matters is that these never expire. A breach from years ago is still useful, because your date of birth and Social Security number do not change.
• Data brokers. This one is legal, which is what makes it insidious. Your home address, phone number, family members’ names and financial indicators are routinely packaged and sold by data brokers, feeding the same dossiers criminals use to impersonate you.
• Phishing. Still everywhere, but sharper now. A single click on a convincing fake login page can hand over your credentials instantly, or install the infostealer that harvests everything else.

The theme across all four: this is an industry, not a series of one-off crimes. The raw material is abundant, cheap and permanent.

Step two: how it gets monetized

Stolen data on its own is just a spreadsheet. The harm happens when it gets turned into money, and there is a whole economy built to do exactly that.

The dark web functions as the marketplace. As one Kaspersky analyst described it, stolen credentials and card data harvested by infostealers get aggregated, repackaged and sold, while ready-made phishing kits are offered as services. It is a self-sustaining loop: data theft feeds fraud, and fraud funds more data theft.

Once someone buys your data, they can do several things with it:

• Account takeover. They log into accounts you already have, using your real credentials, and drain them or use them to reach others.
• New-account fraud. They use your Social Security number and date of birth to open new credit cards or loans in your name. You often find out only when a debt collector calls about an account you never opened.
• Synthetic identity fraud. This is the fastest-growing and most insidious variant, and it deserves its own explanation, because it is where the system is failing worst right now.

The synthetic identity problem, and why systems fail to catch it

Synthetic identity fraud does not steal your whole identity. It takes pieces, most damagingly a real Social Security number, often belonging to a child or a deceased person, and combines them with a fabricated name, address and manufactured credit history to build a person who does not exist but can pass verification.

The losses are staggering and climbing. US unsecured credit losses tied to synthetic identity fraud are projected to exceed $3.1 billion, up from $1.8 billion in 2020, growing around 16% a year. Broader estimates of the total US economic hit run far higher.

Here is why the systems fail to catch it, and this is the part that should bother you. These synthetic identities are cultivated slowly, over months, across multiple financial products. Because there is no real victim noticing fraud early, and because the account behaves normally for a long time, the fraud-detection systems that watch for sudden suspicious activity see clean traffic where they should see a threat. The fraud is invisible precisely because it is patient.

And the people who get hurt are often the least able to fight back. When a child’s Social Security number is used to build a synthetic identity, the theft can go undetected for years, surfacing only when that child turns eighteen and gets denied their first loan or credit card for debts they never knew existed. The system was never designed to protect an identity that has not started being used yet.

How AI turned up the volume

Everything above existed before AI. What AI changed is the scale and the success rate, and the shift has been fast.

Generative tools now let fraudsters manufacture synthetic identities that pass document-based verification, at industrial volume. Services exist that generate high-fidelity fake government IDs in seconds. Pair one of those with a deepfake video feed that matches the fake ID photo, and an attacker can pass through every layer of a typical verification funnel, the document check, the selfie, the liveness test, as a person who does not exist.

The AI escalation shows up across the board:

• Synthetic document fraud jumped sharply in a single year as AI tools made convincing fake IDs trivial to produce.
• AI phishing sends hyper-personalized messages that reference your real name, employer and recent activity, making them far harder to dismiss.
• Voice cloning now needs only a few seconds of audio, scraped from social media, to impersonate a family member in a distress scam. Seniors alone lost billions to fraud last year, much of it voice-enabled.

The through-line is that AI did not invent new categories of identity theft. It removed the friction from the existing ones. What used to take skill and time now takes a subscription, which means the volume aimed at every one of us has gone up.

The controls that actually stop it

Now the useful part. You cannot un-leak your data, and you cannot personally fix broken verification systems. But the machine above has specific choke points, and a handful of controls hit them hard. These are ranked roughly by impact, because not all advice is equal and most articles pretend it is.

1. Freeze your credit. This is the big one. A credit freeze blocks lenders from pulling your credit report, which means that even if a criminal has your Social Security number and date of birth, they cannot open a new account in your name. Security experts consistently rank it the single highest-impact protection available to an individual, and it is free. You place it separately with all three bureaus, Equifax, Experian and TransUnion, and you can lift it temporarily when you actually need credit. It directly defeats new-account and synthetic-identity fraud, the categories doing the most damage.
2. Use phishing-resistant MFA, and move off SMS codes. Multi-factor authentication blocks the overwhelming majority of account-takeover attacks. But text-message codes are the weak version, defeated by SIM swaps and interception. Passkeys and hardware security keys are the strong version, because there is no code to steal or phish. Turn the strongest option you have on your email first, since your email is the master key that can reset everything else.
3. Use a password manager, with a unique password everywhere. Password reuse is what turns one leaked password into ten compromised accounts. A password manager makes “unique everywhere” effortless, which shuts down credential stuffing against you almost entirely.
4. Separate your anchor identity. Keep one private email address used only for banking and government accounts, and a different one for social media, shopping and newsletters. That way a breach of your public-facing address cannot cascade into your financial life.
5. Turn on alerts and monitor. Transaction alerts, new-login alerts, and free weekly credit report checks compress the single most important variable in identity theft: how long it goes undetected. The research is blunt that every day of earlier detection changes the outcome.
6. Reduce what you hand over. Check whether your email appears in known breaches, revoke access for apps you no longer use, and share less personal detail publicly, since that is the raw material for both social engineering and the security-question answers that protect your accounts.

If you only do one thing, freeze your credit. If you do two, add a passkey to your email. Those two moves alone close the doors the most damaging fraud walks through.

Where the harder problem is heading

None of the controls above fixes the deeper issue, and it is worth being honest about that.

The root problem is that our whole identity system runs on shared secrets, a Social Security number, a date of birth, a mother’s maiden name, that were never designed to be secrets and are now sitting in 53 billion leaked records. You cannot reset your date of birth. A system that authenticates you using facts that are permanently public is broken at the foundation, and no amount of individual vigilance fully fixes a broken foundation.

This is why a lot of the interesting work is shifting toward a different question. Instead of asking “does this person know the secret,” which criminals also know, the newer approaches try to confirm that a real, unique, live human is actually present at the moment of a sensitive action. A selfie-based check that confirms a genuine live person, without building a database of who they are, is one example of this direction.

Tools like Selfie Check aim to confirm a real, unique human is present rather than checking a secret that anyone might have stolen, which is a meaningfully different way to think about the problem than the one that got us here.

That shift will take years to play out across banks, governments and platforms, and it raises its own questions about privacy and access that deserve real scrutiny. But the direction is sound. The long-term answer to identity theft is probably not asking people to guard secrets that are already public. It is building systems that care whether a real human is actually there.

For now, though, the practical truth stands. Your data is out there, the machine to exploit it is mature, and AI has raised the volume. You cannot change that. What you can change is whether the specific doors that fraud walks through, new credit accounts, your email inbox, your reused passwords, are locked.

Freeze your credit, harden your email, stop reusing passwords, and you move yourself from an easy target to a hard one. Against an industry that runs on volume and easy targets, being hard to rob is most of the battle.

How identity theft actually happens was originally published in The Repo on Medium, where people are continuing the conversation by highlighting and responding to this story.

Spend an evening swiping and a strange new feeling shows up.

You match with someone. Good photos, quick wit, warm messages. And instead of excitement, a small suspicion surfaces first: is there actually a person on the other end of this, or am I talking to something?

A few years ago that question would have sounded paranoid. In 2026 it is just sensible. The tools to manufacture a completely fake but completely convincing person are cheap, fast and everywhere, and dating apps have become one of the places that fakery hurts most.

Which raises a practical question worth understanding: how does a dating app actually tell a real user from a fake one, and does the newest attempt, Tinder’s “real human” badge, actually work?

The lie got too good

Fake profiles are as old as online dating. For years the giveaways were reliable: a too-perfect model photo, a stolen image that turned up in a reverse search, stilted English, a story that did not add up.

Generative AI erased every one of those tells.

Consider what a modern fake looks like now:

• The face is AI-generated, belonging to no real person, so a reverse image search returns nothing.
• The conversation is AI-scripted, fluent and emotionally responsive, capable of stringing someone along for weeks.
• The operation runs at industrial scale, with one person quietly managing dozens of these personas at once.

The data backs up how bad it has gotten. A Bloomberg analysis found that 12 to 18% of new Tinder US accounts in early 2026 showed signs of AI-generated photos, and Match Group has named this its number-one trust-and-safety concern.

The human cost is not abstract. Americans lost more than $1 billion to romance scams in a year by the FTC’s count, and losses in just the first nine months of 2025 hit $1.16 billion, with a typical victim out around $2,200. Roughly one in four people on dating apps say they have been targeted by a scam.

Why the usual defenses came up short

Dating apps did not sit still. But most of their tools were built for an older kind of fake, and they solve only part of the problem.

Think of verification as a series of steps, each tougher to fake than the one before it:

The weakest: email or phone signup. All it proves is that you can receive a message. Both are available in bulk for pennies, so this stops almost nobody.

Better: the video selfie. Tinder now asks users to record a short selfie to confirm a live face is present. This helped, and it is where a lot of apps landed. But AI-driven face-swapping is steadily eroding it.

Stronger, but costly: ID verification. Matching a government document to a selfie is genuinely hard to fake. The trade-off is steep, though. You surrender your legal identity, and the app accumulates a pile of sensitive documents that becomes a breach risk.

Here is the catch that ties all three together. Every one of them can confirm that a real face showed up once. None of them stops a single verified person from quietly running many accounts. And the strongest option, ID, only closes that gap by taking something most people would rather not give a dating app: their real name and documents.

The missing capability is oddly specific. Confirm that someone is one real, unique person, without learning anything else about them.

Enter the “real human” badge

That specific gap is what Tinder’s newest verification layer goes after. In 2026, Match Group, Tinder’s parent company, partnered with World to offer it.

The mechanism reframes the question. Rather than “does this face match this document,” it asks: is a real, unique human behind this profile?

For the user, it is three simple steps:

1. Verify once with World that you are a real, unique person.
2. Link that verification to your Tinder profile.
3. A verified human badge appears for potential matches to see.

The badge quietly tells the other person: this is a genuine, single human, not a bot and not one of many profiles worked by the same scammer. Match Group’s trust and safety lead framed the partnership as a “natural next step” as AI keeps raising the stakes.

The two things that make it different from ID checks are worth spelling out:

• It proves you are human, not who you are. It relies on a zero-knowledge proof, a bit of cryptography that lets you prove a single fact, “I am a verified unique human”, while revealing none of the personal details behind it.
• It is built to protect privacy. Tinder does not receive your name or a profile of you. It gets a yes-or-no on humanness, and nothing else.

World’s own write-up of the dating rollout lays out the specifics of the badge and how it launched.

It started in Japan, then grew

Worth being precise about the rollout, because it shows a deliberate approach rather than a rushed launch.

The integration launched first as a pilot in Japan. Match Group and World used that market to test a real question: does a proof-of-human badge actually make people trust matches more and lead to better connections?

The pilot’s results were encouraging enough that the feature expanded to more markets, including the United States, where verified users also get a short-term perk of free profile “Boosts” for signing up early.

So if you see this described as Japan-only, that is out of date. Japan was the starting line, not the whole race.

The same fix, far beyond dating

It is easy to read this as a dating story. The more interesting angle is that the exact same check, confirm one real, unique human, reveal nothing else, solves a pile of problems elsewhere on the internet.

A few examples where “one person, one account” changes the game:

• Event tickets and limited drops. Blocking a single operator from running hundreds of accounts to hoard tickets or sneakers for resale.
• Reviews and ratings. Making one honest review per real person mean something, instead of a listing drowned in fakes from one source.
• Communities and social platforms. Shutting down sock puppets, ban-dodging and bot swarms spun up by one individual.
• Games. Curbing the multi-accounting and bot farms that wreck fairness and in-game economies.
• Promotions and free trials. Making “one per customer” actually hold when anyone can otherwise mint unlimited accounts.

Dating apps are moving first only because that is where the damage feels most personal. The underlying tool is the same everywhere.

One honest note that applies across all of these: confirming a real, unique person is present is not the same as confirming that person is honest. A verified human can still lie or scam. The check answers “is this a real person,” not “should you trust them.” Read the badge as a strong signal, not a guarantee.

The real story here

Step back and the dating angle is almost a footnote to something larger.

For thirty years, the internet ran on a quiet assumption: the thing on the other end is probably a person. AI has broken that assumption, and it broke it fast. Text, voices, faces, entire personalities can now be synthesized well enough to fool careful people.

Rebuilding that assumption, deliberately this time, is turning into one of the central problems of the modern web. And the most promising tools do it in a specific way: they confirm a genuine human is present without demanding that human’s identity as the price of entry.

Tinder’s badge is one early, visible version of that idea. Dating is simply the first place where “prove you are actually a person” stopped being optional and became the entire point. It will not be the last. As synthetic content keeps improving, being able to quietly confirm a real human is there, privately and reliably, looks less like a dating feature and more like something the whole internet is going to depend on.

Picture running an online store. One morning your traffic is up forty percent, which looks like great news until nothing else adds up. Conversions are flat. Your best-selling item sold out in ninety seconds but the sales numbers do not reflect it.

Your support inbox is filling with angry messages from real customers who could not check out. And your server bill just jumped for a month where you did not actually sell more.

That is what the bot problem looks like from the ground. Not a dramatic breach with a ransom note, but a slow, confusing drain where the numbers stop meaning what they used to, and you cannot quite point to why.

Here is the uncomfortable reality behind it. Automated traffic now makes up the majority of activity online, and ecommerce is one of the hardest-hit industries of all. The store owner’s core problem is no longer “are there bots on my site.” It is “I genuinely cannot tell which of these visitors are real.”

The traffic does not look like an attack anymore

The old mental image of a bot is a crude script hammering your server, easy to spot in the logs by its sheer clumsiness. That bot is mostly gone.

Today’s bad bots are built to look exactly like your best customers. According to the 2026 Imperva analysis, 41% of malicious bot traffic disguised itself as ordinary Chrome browser traffic. It arrives from residential IP addresses belonging to real homes, rotates between thousands of them, mimics human mouse movements, and moves across your website, mobile app and APIs the way a genuine shopper would.

This is the shift that makes the whole thing so maddening for a store owner. The bot no longer stands out. It blends in. Security researchers now describe the core challenge as no longer “bot or not,” because that question has become nearly impossible to answer from traffic alone. A visitor that looks completely legitimate on the surface can be quietly hoarding inventory, testing stolen cards, or scraping your entire catalog.

And the scale is not marginal. Studies put the average revenue lost to bot-related fraud and operational costs at around 3.6% of revenue for online merchants. For a business running on thin retail margins, that is not a rounding error. That can be the difference between a profitable year and a bad one.

What the bots are actually doing in the store

“Bot traffic” sounds like one problem. On the ground it is several very different problems that happen to share a cause, and each one hurts a different part of the business.

• Inventory hoarding and scalping. Bots rush limited or high-demand items into carts the instant they drop, or buy them outright in seconds, before any human can check out. The stock shows as gone. Real customers see “sold out” and leave. The goods reappear on a resale site at a markup. One study of a limited-edition sneaker release showed a single retailer preventing over $500,000 in scalping-related loss in a 30-day window once they got it under control.
• Denial of inventory. A crueler variant. Bots add your stock to carts and just hold it there, never buying, so the item shows as unavailable to everyone else. Your shelves look empty, sales stall, and there was never any purchase to explain it.
• Credential stuffing and account takeover. Bots test millions of stolen email-and-password pairs from unrelated breaches against your login page. Wherever a customer reused a password, the bot gets in, drains loyalty points and gift card balances, and places fraudulent orders using saved payment methods. During major sale events, one security firm observed up to a 6x spike in account takeovers within hours.
• Card testing. Fraudsters use your checkout as a free validation tool, running tiny transactions to see which stolen card numbers still work before using the good ones elsewhere. Roughly a third of global merchants face active card-testing attacks, and each one leaves you with the chargebacks and fees.
• Fake account creation. Bots flood your signup flow to farm referral bonuses, stack promo codes, claim first-order discounts and grab early-access slots, all things meant to win real customers, drained instead by scripts.
• Price and content scraping. Competitors and resale operators harvest your prices, descriptions and images at scale, both to undercut you automatically and to feed fake marketplace listings.

Notice that only some of these are “fraud” in the classic sense. Several are just abuse of features you built for real people, which is exactly why they are so hard to wall off without blocking the real people too.

The damage the dashboards do not show

The direct theft is bad enough. But talk to people actually running stores and the quieter costs are what wear them down, because they do not show up cleanly in any one report.

Your data stops being trustworthy. When a large share of your traffic is automated, your analytics lie to you. Conversion rate looks terrible because the denominator is stuffed with bots that never intended to buy. Marketing attribution breaks. You make decisions, adjust ad spend, restock, plan promotions, on numbers that are quietly fiction.

You pay to serve traffic that will never buy. Every bot request hits your infrastructure. Fake checkout bots inflate session counts and force your systems to scale up, driving real hosting and bandwidth costs for zero revenue. You are subsidizing the people attacking you.

Your real customers absorb the friction. This is the one that does lasting damage. When credential-stuffing bots pound your login page, genuine users hit slowdowns and lockouts. When you panic and bolt on aggressive CAPTCHAs, real shoppers rage-quit at checkout, and cart abandonment climbs. The bots created the problem, and your loyal customers pay for it in a worse experience.

Your team burns out chasing ghosts. Fraud and support staff spend hours reviewing suspicious orders that were never real, investigating login-spike alerts, and fielding complaints from customers who lost a fair shot at a product. As one merchant put it, it is wasted money, chargeback risk, and hours spent reviewing false orders that were never meant to go through.

None of these show up as a line item called “bots.” They show up as weird conversion numbers, a higher server bill, more support tickets and a vague sense that the store is harder to run than it should be. That is why the problem stays hidden for so long.

Why the old fixes keep failing

Most stores reach for the familiar tools first, and discover they no longer hold.

CAPTCHA was the classic answer, and modern bots solve CAPTCHAs faster and more reliably than humans do, so the main thing a CAPTCHA reliably does now is annoy your real customers. Blocking by IP address fails because attackers rotate through thousands of residential IPs that look like ordinary home connections. Rate limiting helps against crude floods but not against a patient bot that behaves like a careful shopper. And web application firewalls tuned to known bad signatures miss the AI-driven bots that change behavior the moment they are detected.

The deeper reason these fail is that they all try to answer “is this traffic abnormal.” Modern bots are specifically built to look normal. As one security team framed it, retailers now have to detect intent rather than velocity, behavior rather than IP, patterns rather than raw volume. That is a much harder question, and it is why bot defense has quietly become a specialized field rather than a checkbox.

Where this is heading: proving there is a person there

Step back and the common thread across almost every one of these attacks is the same. The damage comes from a small number of operators cheaply pretending to be a large number of customers. One scalper wearing a thousand shopper costumes. One fraud ring spinning up ten thousand fake accounts. The whole model depends on being able to fake being many different people.

That reframes the defensive question in a useful way. Instead of endlessly asking “is this traffic suspicious,” a growing line of thinking asks something more direct: is there a real, unique human behind this action at all?

This is the idea of proof of human. Rather than inspecting behavior for signs of a bot, which is a race that gets harder every year, it aims to confirm at key moments, account creation, a limited drop, a high-value checkout, that a genuine and unique person is present, without forcing that person to hand over their identity. The appeal for a retailer is specific: if each real human can be confirmed as one human, the economics that make scalping and fake-account abuse profitable collapse, because you can no longer be a thousand people at once. Approaches in this space, from behavioral and device intelligence vendors through to newer privacy-preserving proof-of-human methods, are converging on that same goal from different directions.

It is not a magic fix, and it is worth being honest about the limits. A verified human can still behave badly, so proof of human raises the cost of mass abuse rather than ending fraud outright. Any check added at checkout risks adding friction, so where and how it is applied matters enormously. And a real defense still layers several signals together rather than betting on one. But the direction is sound, and it targets the actual root of the problem rather than its symptoms.

The honest takeaway for anyone running a store

The bot problem in ecommerce is not the dramatic thing most people imagine. It is quiet. It hides inside traffic that looks completely legitimate, drains a few percent of revenue here, corrupts your analytics there, frustrates your real customers, and inflates your costs, all without ever announcing itself.

The first real step is simply recognizing it for what it is. If your traffic spikes without matching sales, if limited items vanish impossibly fast, if your login page sees waves of failed attempts, if your conversion rate looks inexplicably poor, those are not random glitches. They are the fingerprints of automation moving through your store.

The uncomfortable truth is that you cannot out-filter this problem with the tools of five years ago, because the bots learned to look human. The more durable path is to stop trying to spot what is fake and start being able to confirm what is real. The stores that figure that out get back the thing bots quietly stole first, which was never just inventory or revenue. It was the ability to trust their own numbers.

The two phrases get used as if they mean the same thing. They do not, and the difference is not a technicality. It decides whether a system needs your consent, whether it builds a database of your face, and whether you can walk away from it.

I started paying attention to this after noticing how often a news story about creepy government surveillance and a mundane story about unlocking your phone used the exact same words: “facial recognition.” Those are opposite things wearing one label. Once you see the distinction, a lot of the confused public argument about face technology suddenly makes more sense.

Here is the difference, why it matters, and why almost everyone mixes them up.

The one difference everything else flows from

The entire distinction comes down to a question of matching, and engineers describe it with a simple shorthand: 1:1 versus 1:N.

Face authentication is a one-to-one match. The system already has one reference template for you, and it checks whether the face in front of the camera matches that one template. It is answering a yes-or-no question: is this the specific person they claim to be? When your phone unlocks with your face, it is not searching a database. It is comparing the live face to the single stored template of the phone’s owner and returning yes or no.

Facial recognition is a one-to-many match. The system takes one captured face and searches it against a database of many faces to figure out who it is. It is answering an open question: out of all these people, who is this? A camera scanning a crowd and checking every face against a watchlist is doing 1:N. Nobody claimed an identity. The system is assigning one.

That single distinction, one-to-one versus one-to-many, is the root of everything else. As one industry explainer put it, use cases, regulatory treatment, privacy implications and fraud risk all flow from that one difference. Get the matching model right and the rest of the confusion clears up.

Same building blocks, opposite purposes

Part of why people conflate them is that under the hood they use related technology. Both map a face into a mathematical representation, a template, using similar AI techniques. The plumbing overlaps.

But the purpose is reversed, and that reversal is the whole story.

Authentication is something you do, deliberately, to prove a claim you made. Recognition is something done to you, often without your knowledge, to work out who you are. That is why the same underlying math ends up in both a convenient login and a surveillance camera, and why treating them as one thing muddles every conversation about face technology.

Why the distinction matters for privacy

This is where the difference stops being academic.

Consent. Authentication is consensual by design. You choose to enroll, you present your face to make a claim, and you usually have another option, a passcode, if you would rather not. Recognition frequently happens without consent or even awareness. You do not opt into being scanned as you cross a train station.

Databases. Authentication needs only your own template, checked against your own claim. Recognition requires a database of many people’s faces to search against, and that database is the thing privacy law worries about. It can be shared, sold, breached or repurposed. A one-to-one system has no such pool to leak.

The law already treats them differently. This is the clearest signal that the distinction is real. Laws like Illinois’ BIPA, the EU’s GDPR and various state rules have moved to restrict non-consensual facial recognition specifically, the 1:N surveillance kind, while consensual 1:1 authentication remains the accepted standard for logging in and opening accounts. Lawmakers drew the line exactly where the matching model changes.

Permanence. Both use biometrics, and a face cannot be reset like a password. But a 1:1 system that keeps your template on your own device carries far less exposure than a central 1:N database holding millions of faces, which is a permanent, high-value target that can never be made un-leaked.

The short version: the privacy risk is not really about “face technology” in the abstract. It is about whether a system is matching you against yourself, or against everyone.

The threat models are different too

Because they answer different questions, they get attacked differently, which matters for anyone building or trusting these systems.

Authentication’s main threat is someone pretending to be you: holding up a photo, a mask, a screen or a deepfake video to the camera to pass the 1:1 check. This is called a presentation attack, and it is why serious authentication pairs the face match with liveness detection, a check that a real, live person is present rather than a spoof. Without liveness, deepfakes and AI-generated faces can fool a face match. With it, the bar is much higher.

Facial recognition’s risks are more societal than personal: misidentification (the system fingers the wrong person from the database), demographic bias in accuracy, and the simple existence of a searchable face database that can be turned to purposes the people in it never agreed to.

Different questions, different failure modes. Lumping them together hides both.

Selfie tools that check you are human, without knowing who you are

There is a newer category worth knowing about, because it sits in an odd spot between the two ideas above. A bunch of tools now use a selfie not to figure out who you are, and not always to authenticate an identity you have claimed, but just to check that a real, live human is on the other end. Usually the point is keeping bots and fake accounts out.

They all start from the same basic move, a selfie plus a liveness check, but they make very different choices about what happens next, and that is where the privacy questions from earlier come back. The main names are Incode, Jumio, Sumsub, Persona, Mitek and CLEAR, and Selfie Check from World.

The first six share a shape. They verify a live face, most of them match it against a government ID, and the selfie travels to a company server to get processed. In the language from earlier, that is mostly consensual authentication, usually bolted onto an identity database. That is exactly what you want when the law says you have to know who someone is, opening a bank account, say. It is a lot more than you need if all you are trying to do is stop ten thousand bots from signing up.

Selfie Check is the one doing a different job, so it is worth explaining plainly how it works. You take a selfie in the World App. The liveness check runs on your own phone, the photo stays there, and the app you are actually signing into never sees your face, just a yes-or-no that a real, unique person did the check, plus a rough signal of whether that person has shown up before.

The clever part is a privacy technique called anonymized multi-party computation, which is a mouthful, but the practical upshot is simple: it can tell that two accounts are the same person without anyone keeping a central pile of faces to compare against.

It is a deliberately lighter check than a full in-person one, tuned for a specific job rather than maximum strength. The point is not to lock down a bank vault. It is to answer one narrow question, is this a real, unique human, and to answer it without learning who the person is or letting their selfie leave their phone. For keeping bots and duplicate accounts out while leaving ordinary people anonymous, that is the exact question worth answering.

So the whole group splits along the exact line this article is about. The identity tools are doing authentication, and some of them are wired into recognition-style databases behind the scenes. The proof of human approach tries to skip identity altogether and answer only the humanness question. Same selfie, completely different question being asked.

So why does everyone mix them up

A few reasons, and they are understandable.

The vocabulary is genuinely messy. “Facial recognition” gets used as a catch-all for anything involving a face and a camera, even when the actual task is 1:1 authentication. Even careful writers slip.

The technology looks identical from the outside. In both cases you point a camera at a face and something happens. The difference is invisible: it lives in whether the system is checking you against your own template or searching you against a crowd.

And the marketing muddies it further. Plenty of products say “facial recognition” because it sounds powerful, or avoid it because it sounds sinister, regardless of which thing they actually do.

But the distinction is worth holding onto, because it is the fastest way to judge any face-based system you encounter. Ask one question: is this matching me against myself, or against everyone else? If it is one-to-one, you claimed an identity, you consented, and there is no database of strangers. If it is one-to-many, you are being identified, possibly without asking, out of a pool that has to exist somewhere.

The words will keep being used loosely. You do not have to. Once you can tell 1:1 from 1:N, you can see past the label to what a system is actually doing, which is the only thing that ever mattered.

Where this is quietly heading

It is easy to finish a piece like this feeling uneasy about face technology in general. I came away more optimistic, and it is worth saying why.

The law is starting to draw the line in the right place: cracking down on the non-consensual, one-to-many surveillance use while leaving the consensual one-to-one authentication use alone. The distinction this whole article is about is being written into regulation, and the creepy version is the one under pressure.

The engineering is moving the same way. The most interesting new tools are not building bigger face databases. They are building smaller ones, or none at all. Matching happens on your own phone. Liveness detection is getting good enough to blunt deepfakes. And proof of human approaches now do something that would have sounded contradictory a few years ago: confirm a real person is there while learning nothing about who they are.

None of this is finished, and the surveillance side still needs watching. But the direction is good. A face check that proves you are you, or simply that you are human, without adding you to a database anyone can search later, is a future worth aiming at. The tools mostly exist. The rest is choosing to use them that way.

Face authentication and facial recognition are not the same thing was originally published in The Repo on Medium, where people are continuing the conversation by highlighting and responding to this story.

I fell down this rabbit hole because of a ramen place.

It had 4.6 stars from something like 1,800 reviews, which sounds great until you actually scroll. Page after page of bare five-star ratings. No text. No photos. Just a star and a username, over and over. And a handful of furious one-stars from people describing a thoroughly different restaurant, wrong neighborhood, wrong menu; clearly never been there.

So I started wondering how that happens. Google runs some of the most advanced bot detection on the planet. I have been made to click fire hydrants enough times to know they are watching. So how does a listing end up stuffed with reviews that so obviously mean nothing?

The answer turned out to be more interesting than “the system is bad.” The system is actually pretty good. It is just built to catch the wrong thing.

The gap everything else is trying to paper over.

Before any of the clever technology, there is one plain fact that explains most of this.

A Google review is not tied to a purchase. You do not need a receipt, a booking, a check-in, or any proof at all that you set foot in the place. If you have a Google account, you can rate any business on Earth right now, this second, having never gone near it.

That is the whole foundation, and it is worth sitting with. Every fancy detection system that comes later is essentially trying to compensate for this one missing check. None of them can, because you cannot detect after the fact whether someone walked into a shop when you never asked them to prove it in the first place.

Why is it built this way? Coverage, basically. Google Maps wants a rating on every business everywhere, and the surest way to kill participation is to make people prove things. Frictionless means more reviews. It also means the door is wide open.

What Google’s bot protection actually does

Here is the part I actually wanted to understand: what is the technology doing, if not checking whether you were a customer?

It is doing a genuinely sophisticated job of a different task: telling humans and bots apart. That happens in a few layers.

The famous one is the CAPTCHA lineage. It started as distorted text you had to retype, back when computers were bad at reading it. Then computers got good at reading it, so it became the “I’m not a robot” checkbox, and then the pick-the-traffic-lights grids. The newest version does not show you anything at all.

It runs silently in the background and watches how you behave: how your mouse moves, how you scroll, timing between actions, your browser and device fingerprint, and whether your history looks like a real person’s. From all that, it produces a single risk score, roughly zero for “probably a bot” up to one for “most likely human.” The site decides what to do with borderline cases.

Moreover, for reviews specifically, there is a machine-learning layer looking for manipulation patterns. It tracks things like IP addresses and device fingerprints, so ten reviews from one network raise a flag. It watches for velocity, a sudden spike of five-star or one-star ratings on a listing that normally trickle.

And it cross-references accounts, so a profile that keeps showing up to review the same set of unrelated businesses starts to look like part of a ring. Google has said it blocked or removed hundreds of millions of fake reviews in a year using systems like this, so this is not a token effort. It genuinely catches a lot.

And it is still evolving. In 2026, Google folded reCAPTCHA into a broader system called Cloud Fraud Defense, explicitly because the old model was showing its age. One of their product leads put the reason plainly: you cannot reliably tell a human from an AI-generated bot using static challenges anymore. So they are moving from “solve this puzzle” toward watching the whole interaction for signs of coordinated fraud.

All of this is real, and much of it works. Which makes the next part stranger.

Why does none of it catch the fake review that matters?

Here is where it clicked for me.

Every layer above is built to answer one question: is this a bot? And the fake reviews doing the most damage are not written by bots.

Think about who actually writes them. A person in a review farm, paid a few cents per review, types on a real phone with a real Google account that has a normal history. They move their mouse like a human because they are human. They pass the behavioral check perfectly. They pass the device and history checks because the account is aged and looks ordinary. There is nothing automated to detect, because a real person is sitting there doing it by hand.

Or think about the angry one-star from someone who never visited. A competitor. An ex-employee. Someone who saw the owner say something they didn’t like online. That review is written by a genuine human, from a genuine account, expressing a genuine (if malicious) opinion. No bot detector on Earth flags it, because nothing about it is fake in the way the system understands “fake.” It is a real person lying, and the technology has no test for that.

That is the whole thing, in one line. CAPTCHA and bot detection verify that you are human. They cannot verify that you are honest, and they cannot verify that you were ever a customer. For reviews, humanness was never the missing piece. Everyone assumed the problem was bots, and built brilliant defenses against bots, while the actual problem was humans vouching for experiences they never had.

And the bot side is getting harder too

To be fair, even the part Google is good at is under pressure.

Machine-learning models can now solve image CAPTCHAs with unnerving accuracy. There are services, sometimes called CAPTCHA farms, that route challenges to rooms of low-paid human workers who solve them for a fraction of a cent each, so even a “human-only” test just gets handed to an actual human. And AI now writes review text that reads perfectly naturally and slips past AI detectors after a light edit.

So Google is squeezed from both directions at once. On one side, humans writing fake reviews it was never designed to catch. On the other hand, bots are getting better at looking human. The puzzle-based approach is quietly losing relevance, which is why they are pivoting to watching behavior and coordination instead. But coordination detection still mostly catches the clumsy, scaled-up operations. The patient's single fake account, one human, one honest-looking account, one lie, remains close to invisible.

Which is why the empty five-star exists

Once you see it this way, the bare five-star rating with no text and no photo stops being a mystery. It is the fingerprint of the missing check.

Think about it. If you actually ate somewhere, you have something to say, even a sentence. The pad thai was great, the wait was long, the room was loud. A real experience leaves you with detail. A manufactured rating leaves you with nothing to write, because there was no experience, so the fastest path is to tap five stars and move on. Multiply that by a farm working through a list, and you get exactly what I saw on that ramen listing: a wall of naked stars.

The empty five-star is not laziness. It is the visible symptom of a rating that was never attached to a visit.

Why do some people quietly trust other places more

This is also why you will hear people say they trust certain other platforms more, even without being able to explain why.

The common thread is a link to an actual transaction. When a shopping site marks a review “verified purchase,” it is saying this person actually bought the thing, which is precisely the check Google skips. When you can only review a hotel through the service you booked it on, the review is gated behind a real stay. Some platforms lean hard on aggressive filtering that hides reviews their system does not trust, which frustrates businesses but does raise the credibility of what remains.

None of these are perfect. Verified-purchase tags get gamed, filters hide legitimate reviews along with fake ones, and every one of these platforms has its own fights with fraud. But notice the pattern: they all add either friction or a transaction link that Google deliberately left out for the sake of coverage. The trade is fewer reviews, more trustworthy ones. Google made the opposite trade: more reviews, less certainty behind each one.

What this actually means for you

The practical upshot is a small shift in how you read a listing, and it is genuinely useful.

Stop trusting the number at the top. The average star rating is the single most manipulated and least reliable thing on the page. Read the actual written reviews instead, especially the detailed ones and the ones with real photos, because those are the hardest to fake and the most likely to come from someone who was really there.

Be suspicious of a big cluster of five stars that all landed in the same short window; that pattern is what a burst of purchased reviews looks like. And when a place matters, cross-check it somewhere that ties reviews to a purchase or a booking.

Basically, read past the score to the texture underneath it. The score is the part designed to be gamed. The specific, detailed, photographed human account is the part that is difficult to fake.

The thing I keep coming back to

What stuck with me is that Google mostly won its war. The bot detection really is impressive, and it really does block enormous amounts of automated junk. It is a genuine engineering achievement.

It just turned out to be the wrong war. The question that decides whether a review is real was never “Is this a person?” It was “Was this person actually here?” and that is the one question none of the technology asks, because the whole system was built to avoid asking anyone to prove anything.

So the star rating, the thing that was supposed to be a two-second shortcut to trusting a stranger’s experience, quietly stopped meaning that. Not because the bots won, but because being human was never proof you were a customer, and being a customer was the only thing that ever made a review worth reading.

If you loved reading through this, then do check out my other articles. Also, do drop a follow and stay tuned for more articles.

Bot protection for the next generation of the web

Why a 5-star Google review means almost nothing was originally published in ILLUMINATION on Medium, where people are continuing the conversation by highlighting and responding to this story.

There is a statistic that reframes how you should think about the internet, and most people have not caught up to it yet.

As of 2025, bots make up the majority of web traffic. The 2026 Imperva Bad Bot Report, which analyzes traffic across thousands of domains, found automated traffic reached more than 53% of the total, the second year running that machines outnumbered people online. Human activity has fallen below half and keeps sliding.

Sit with that for a second. When you post something, sign up somewhere, vote in a poll, or read a comment section, the statistical majority of the “people” in that space are not people. They are software.

That is the backdrop for a phrase you are going to hear a lot more: proof of human. It is being pitched as the missing piece the internet needs to keep working now that humans are the minority. This is an attempt to explain what it actually means, why it suddenly matters, the different ways people are trying to build it, and the honest limits of the whole idea.

What proof of human actually is

Strip away the jargon and proof of human answers one deceptively simple question: is there a real, unique person here?

Not “who is this person.” That is identity, and it is a different thing. Proof of human is narrower and, in a way, more modest. It tries to confirm two facts and nothing more:

• Real: a living human, not a script or an AI agent
• Unique: one person counts once, so the same human cannot show up as a thousand accounts

That second word is the important one. Most of the damage bots do comes from scale, one operator pretending to be a crowd. Fake reviews, manipulated votes, hoarded tickets, spam, coordinated influence campaigns. They all rely on one person, or one program, cheaply becoming many. A working proof of human breaks that, because it makes each real person count exactly once, no matter how many accounts or agents they spin up.

Crucially, the good versions try to do this without learning who you are. The goal is to prove you are a real, singular human while revealing nothing else about you. That combination, confirming humanness while preserving anonymity, is the hard part, and it is where the different approaches diverge.

Why it suddenly matters so much

Proof of human is not a new idea, but three things turned it from a niche crypto concept into something mainstream in about two years.

CAPTCHAs stopped working. The old test, prove you are human by reading squiggly text or picking traffic lights, is effectively over. Modern AI solves those faster and more reliably than you can. The tool the internet leaned on for two decades to filter bots no longer filters bots.

AI made fakes cheap and convincing. Generative tools erased the old tells. Fake profiles have flawless photos, scam messages read fluently in any language, and deepfake video can pass a video call. The signals people used to rely on to spot a fake are gone.

Agents arrived. This is the newest twist. A fast-growing share of automated traffic is not malicious at all, it is AI agents doing tasks real people asked for. One analysis found that only half a percentage point separates the rate of benign automation from malicious automation, with agentic traffic growing explosively across 2025. The old “block all bots” reflex breaks completely when a lot of the bots are working on your behalf. You can no longer sort traffic into human and machine. You have to ask a different question: is a real person behind this.

Put together, the filters failed, the fakes got good, and the line between good and bad automation blurred. Proof of human is the response to all three at once.

The different ways people are building it

Here is the part most coverage skips. There is no single proof of human system. It is a whole field of competing approaches, and they make very different trade-offs. Understanding the main families tells you more than any one product does.

Social-graph verification. Projects like BrightID verify you through your connections. If enough already-trusted humans vouch that you are real, your trust score rises. No biometrics, no documents. The catch is that it needs an existing network to bootstrap, which makes it weaker in places without an established community, and social graphs can in principle be gamed by coordinated groups.

Cognitive-puzzle ceremonies. Idena takes a radically different route: at synchronized moments, everyone solves puzzles that are easy for humans and hard for machines, all at once, which makes it very hard to automate at scale. It is clever and privacy-friendly, but it asks real effort and scheduling from participants, which limits how large it can grow.

Multi-signal reputation. Human Passport (formerly Gitcoin Passport) aggregates many weak signals, a GitHub account, a wallet, a phone number, a government credential checked with zero-knowledge proofs, into a single “humanity score.” No single point of failure, and quite flexible, though a determined attacker can assemble enough stamps to farm a score.

Video and court-based registries. Proof of Human combines a video submission with community vouching and an on-chain dispute court to maintain a public list of verified humans. Robust and decentralized, but slow, and the public registry raises its own privacy questions.

Biometric networks. This is the most visible family, and the most debated. It uses a biometric check to establish uniqueness at scale, on the reasoning that bodies are the one thing an attacker cannot mass-produce. It reaches the largest numbers but draws the most scrutiny, because anything touching biometrics raises real and reasonable privacy concerns.

None of these has won. Each buys uniqueness with a different currency: your social connections, your time, your accumulated accounts, your body. That trade-off is the whole game, and it is worth judging any proof of human system by which currency it asks for and what it does with it.

World, as one concrete example

The most prominent project in the biometric family is World, and it is worth walking through as a concrete case, precisely because it shows the trade-offs the whole field is wrestling with.

It helps to be accurate about what World is, because it gets described carelessly a lot. World is not an identity company, and it is not a system that collects your biometric data onto a central server to hold. It operates a proof of human network: a way for a person to prove they are a real, unique human, anonymously, without handing personal information to the apps they use. As of its third anniversary it reported nearly 18 million verified people across more than 160 countries, which makes it one of the largest live attempts at this.

The lighter-weight piece worth understanding, because it maps directly onto the “is a real person here” question, is a tool called Selfie Check. It uses a selfie to confirm that a real, unique human is present, without working out who they are.

The selfie stays on the person’s own device, the liveness check runs locally, and an app that uses it receives only an anonymous signal that a genuine, singular person passed, plus a sense of whether that person has been seen before. It leans on a privacy technique called anonymized multi-party computation, which lets it flag a duplicate person without keeping a central pile of faces to match against.

The point of including it here is not that it is the answer. It is one implementation, in one family, making one set of trade-offs. It reaches large scale and keeps the person anonymous, and in exchange it relies on a biometric signal that some people would simply rather not use.

That is exactly the kind of trade every approach in the previous section is making, just with a different currency. The honest way to see World is as one serious attempt among several, not as the definition of the category.

The honest limits nobody should skip

Proof of human is a genuinely promising idea, and it is not a solved one. Anyone selling it as a finished fix is overselling it. A few limits are worth holding onto.

Uniqueness is not virtue. Confirming that a real, unique person is behind an account says nothing about whether they are honest. A verified human can still post lies, run a scam, or be paid to vote a certain way. Proof of human raises the cost of running a thousand fake accounts. It does not make the one real account trustworthy.

Inclusion cuts both ways. Any system that becomes a gate to online life has to work for everyone, including people without smartphones, documents, stable internet, or a social network to vouch for them. A proof of human that quietly excludes the already-marginalized is a serious problem, not a footnote, and every approach struggles with the hardest-to-reach edges.

Adoption decides everything. A proof of human standard is worthless until the places you actually use adopt it. Right now the field is fragmented across many competing systems, and fragmentation is its own failure mode. The technology can be excellent and still go nowhere if it stays a patchwork.

Every approach has a real cost. Biometrics raise privacy concerns. Social graphs can be gamed and exclude newcomers. Puzzle ceremonies do not scale comfortably. Reputation scores can be farmed. There is no free option. Picking one means picking which downside you can live with.

Where this leaves us

The internet spent thirty years assuming the entity on the other end was probably a person. That assumption is now false more often than it is true, and most of the systems we rely on were built on top of it.

Proof of human is the most coherent attempt to rebuild that assumption deliberately rather than take it for granted. The core move is quietly elegant: stop trying to detect what is fake, which is a race you lose a little more every year, and start confirming what is real, once, in a way a machine cannot counterfeit and a swarm cannot hide inside.

Whether it works will not be decided by the cleverest cryptography. It will be decided by whether these systems can reach everyone fairly, whether the places we spend our time online agree to use them, and whether people decide the trade any given approach asks for is one they are willing to make. Those are human questions, not technical ones, which is oddly fitting. The whole point, after all, is to make room for the humans again.

What “proof of human” means when bots outnumber people was originally published in The Repo on Medium, where people are continuing the conversation by highlighting and responding to this story.