AI broke the internet’s honor system. Blockchain is one of the repair crews.

A confession, before we start

I follow a lot of people at the intersection of crypto, AI, and cybersecurity. Podcasts, newsletters, the whole feed. And for the past couple of years, those voices have been making the case, over and over, that AI and blockchain belong together: blockchain will verify what AI creates, pay for what AI does, prove who’s human and who isn’t. I heard it so often, from people whose thinking I respect, that somewhere along the way it hardened into a belief. Not “blockchain could help AI.” More like: AI must have blockchain to do any of the good stuff.

Then two things occurred to me. The first is that I live in a bubble. If you’re not marinating in the same feeds I am, you’ve probably never heard this pitch at all. “AI needs blockchain” isn’t a thing most people believe, because it isn’t a thing most people have ever heard. And if you did hear it, it probably sounded like two buzzwords stapled together on a pitch deck.

The second thing is more embarrassing: the belief is just wrong, and the proof was right in front of me. I researched and outlined this article with Claude. Before that, I played around with ChatGPT and Perplexity. None of them has a blockchain anywhere in sight, and neither does Gemini, Midjourney, the AI that recommends your next show, or the one that reads medical scans. No blockchain in the training, none in the running, none in the delivery. AI doesn’t need blockchain any more than your car needs a boat. The entire modern AI industry, including the tools I use every day, is a standing, functioning counterexample to the belief I’d absorbed.

So why write this article at all? Not to parrot what I’d been hearing. The voices I follow weren’t hallucinating the benefits. They were just overselling the verb. The way they talked, blockchain should be part of the foundation AI is built on. It isn’t. AI got built, spectacularly, without it. The better picture is the one in this article’s title. AI is busy breaking things all over the internet (quietly, at scale, right now) that our current systems have only limited remedies for. And blockchain, whatever you think of the coins and the hype, is one of the repair crews pulling up to the job site.

Not the only crew on the job, to be clear, but one of them. And here’s the thing about repair crews: nobody calls one for a building that’s fine. But before we can talk about what blockchain fixes, you have to understand what AI broke, which is a challenge, because the things that broke were invisible when they worked.

The honor system we didn’t know we were using

The internet was built on a handful of assumptions so basic that nobody bothered to write them down. They were just... true. Or true enough, often enough, that everything built on top of them held together.

Assumption one: recordings reflect reality. A photo showed something that happened. A video showed someone doing what the video shows. A voice on the phone belonged to the person it sounded like. Fakes existed, sure, but making a convincing one took skill, money, and time. So the default assumption held.

Assumption two: accounts belong to people. Behind every user account, review, comment, and follower was, roughly speaking, a human being. Bots existed, but they were clumsy enough that platforms could mostly filter them, and we could mostly spot them.

Assumption three: content has an origin. An article was written by someone. A photograph was taken somewhere. Work could be traced, at least in principle, back to a person who made it and could be held responsible for it.

Assumption four: money moves when a human decides to move it. Every payment system on earth, from cards to banks to apps, is designed around a person with a legal identity authorizing a transaction.

What that looks like in practice

Most people have heard one or two of these stories and filed them under weird news. They’re not weird news. They’re previews.

Seeing is no longer believing. In January 2024, an employee at the British engineering firm Arup joined a video call with several colleagues, including his company’s chief financial officer. Everyone on the call looked and sounded right. Following instructions from the meeting, he made fifteen transfers totaling roughly HK$200 million, about US$25.6 million. Every other person on that call was an AI-generated fake. Read that again. Not a hacked account sending a suspicious email. A live video call full of convincing digital puppets.

That same month, thousands of New Hampshire voters got a robocall in President Biden’s cloned voice telling them not to vote in the primary. The political consultant behind it was later fined $6 million by the FCC. The cost of producing the fake voice? Reportedly a few hundred dollars and a few minutes.

Scale it down and it’s already in your neighborhood. Security researchers at McAfee showed back in 2023 that three seconds of audio was enough to clone a voice convincingly. Three seconds. Like your voicemail greeting, or a clip from social media. The FTC issued its first consumer alert about AI voice-cloning “family emergency” scams that same year, and Americans reported $2.7 billion in losses to imposter scams in 2023 alone. The identity-verification firm Sumsub measured a tenfold global increase in detected deepfake fraud attempts between 2022 and 2023, and Deloitte’s Center for Financial Services projects that generative AI could push fraud losses in the United States to $40 billion by 2027.

And there’s a second-order problem that’s arguably worse than the fakes themselves: the liar’s dividend. Two law professors, Bobby Chesney and Danielle Citron, named it in 2019, before the tools even got good. Once everyone knows convincing fakes exist, real evidence becomes deniable. Caught on tape? “It’s a deepfake.” The technology doesn’t just let liars fabricate proof. It lets the guilty dismiss authentic proof. Courts, journalists, and insurers are already fielding exactly this defense.

The crowd may not be people. CAPTCHAs, those “click all the traffic lights” puzzles, were the internet’s bouncer. They’ve been beaten. A 2023 academic study found bots solving CAPTCHAs with 85 to 100 percent accuracy, which is better than the humans they were designed to admit, who only managed 50 to 85 percent. Meanwhile, Imperva’s annual Bad Bot Report found that automated traffic crossed the halfway mark in 2024, hitting 51 percent of all web traffic. For the first time in a decade, the majority of internet traffic wasn’t human. Sit with that for a second. Modern AI writes fluent product reviews, holds coherent arguments in comment sections, and can run thousands of social media accounts that build credible-looking histories over months. The next time you see a political opinion “gaining traction” online, or a product with ten thousand glowing reviews, there is no longer any reliable way (for you or for the platform) to know how many humans were involved. Astroturfing, the business of manufacturing fake grassroots support, used to require a call center. Now it requires a laptop.

Origins have dissolved. AI models were trained by scraping much of the internet, largely without asking: articles, photographs, art, code. That’s now the subject of major litigation, most famously The New York Times’ December 2023 suit against OpenAI and Microsoft. But you don’t need to care about copyright law to care about the downstream effect. The internet is filling with AI-generated text and images that carry no reliable marker of what they are, which creates a strange loop researchers have already documented. A study published in Nature in 2024 showed that when AI models are trained on the output of previous AI models, quality degrades generation over generation, like a photocopy of a photocopy. The researchers called it “model collapse.” So the industry now has a genuine interest in something it never needed before: a way to tell, at scale, which content came from a human and which came from a machine. Right now, no such system covers the open internet.

Software is starting to spend money. The newest AI systems don’t just answer questions. They act. They book, purchase, subscribe, and pay for services, increasingly without a human approving each step. Our entire financial system assumes that’s impossible. An AI agent can’t get a credit card. It can’t pass a bank’s know-your-customer check. There is no box on any banking form for “autonomous software program.” Yet the agents are here, and they need to transact.

Four assumptions, four breaks, and an internet that gets harder to trust by the day. Blockchain was the first crypto-era buzzword it felt fine to ignore; AI is the one you can’t. So why does a technology most people associate with cryptocurrency speculation keep coming up as part of the fix?

What blockchain actually brings to this job site

Strip away the coins and the culture, and a blockchain is one thing: a public record book that nobody owns and nobody can quietly edit. Anyone can write an entry (following the rules), everyone can read it, and once an entry is in, it’s in. Permanently, with a timestamp, and any tampering shows. (If you want the deeper mechanics, I’ve covered them in my earlier explainers on what a blockchain is and how one works in practice.)

That specific combination of no owner, no edits, and public timestamps turns out to map neatly onto the broken assumptions.

Repairing “seeing is believing”: the fingerprint registry. Every digital file can be reduced to a unique fingerprint called a hash, a short string of characters that changes completely if even one pixel of the file changes. The repair works like this. At the moment a photo is taken or a piece of content is verified, its fingerprint gets written to a blockchain, along with a timestamp and the identity of whoever recorded it. The blockchain doesn’t store your photo, just its fingerprint. Later, anyone handed that file can recompute the fingerprint and check the record. Is this on the record? Since when? Says who? If the file was altered anywhere along the way, the fingerprints won’t match.

This is no longer theoretical. Umanitek, a company focused on detecting harmful and illegal content online, recently upgraded its Guardian Agent to do exactly this for deepfake detection. When Guardian analyzes a suspicious piece of content, it doesn’t just render a verdict of “authentic” or “AI-manipulated.” It issues a certificate of that finding and anchors the certificate in a decentralized knowledge graph, which is a shared, blockchain-anchored record that no single company controls. In plain terms: the analysis gets stapled to a public, tamper-proof bulletin board at the moment it’s made. Why does that matter? Because a screenshot or an internal company log can be edited, backdated, or simply disbelieved. A certificate cryptographically linked to an immutable public record can be independently checked by a platform, a regulator, or a court, without anyone having to take Umanitek’s word for it. That last part is the whole game. In a dispute over whether a video is fake, “trust our lab” is an assertion. A verifiable, tamper-evident record is evidence. It’s also the start of an answer to the liar’s dividend: “that’s a deepfake” gets a lot harder to claim about a file whose fingerprint has been sitting on a public record since the moment it was captured.

And that’s the general principle. A registry of “what’s real” controlled by one corporation just relocates the trust problem. A registry that no single party can edit or delete, not even the company that wrote the entry, is the point.

Repairing “accounts are people”: proof of personhood. The emerging idea here is a credential that proves “this account belongs to one unique, verified human” without revealing which human. The math that makes this possible is called a zero-knowledge proof, and the plain-English version goes like this: it lets you prove a statement is true (“I am a real person,” “I am over 18,” “I only have one account”) without handing over any of the underlying personal information. Anchoring these credentials to a blockchain means no government or company controls the master list of who counts as human, and the credentials can’t be quietly revoked or duplicated. Imagine comment sections, reviews, or petitions where every participant is provably a distinct human. Anonymity preserved, bot swarms excluded. Think of it as a digital notary that can stamp “verified: one real human” without ever writing down which human. Fair warning, though: of the four repairs, this one is furthest from ready. The most prominent attempt, World ID (co-founded by OpenAI’s Sam Altman, and which verifies humanness by scanning irises), has been suspended or investigated by regulators in Spain, Portugal, and Kenya over biometric privacy concerns. The problem is real. The winning solution hasn’t been crowned.

Repairing “content has an origin”: receipts for the machines. The same fingerprint-and-timestamp machinery can track what goes into AI, not just what comes out. Datasets, licenses, and payments logged to a shared ledger create an audit trail: this model trained on this data, licensed from these creators, on these terms. Add smart contracts, which you can think of as a vending machine made of code (if condition A happens, then payment B executes automatically, no invoices, no accounts-payable department), and you get royalties that flow to thousands of individual creators every time their work is used, without anyone processing the paperwork.

Repairing “money means a human”: wallets for software. This is, candidly, the strongest case of the four. A blockchain wallet doesn’t care whether its owner has a pulse. It’s just a pair of cryptographic keys: an address anyone can send funds to, and a secret key that authorizes spending. An AI agent can hold one, operate within spending limits written in code, and settle payments in seconds for fractions of a cent, including with other AI agents. Use stablecoins (tokens pegged one-to-one to the dollar) and none of the casino-style volatility applies.

This isn’t speculative, and you can tell by watching where the big payment companies are spending. In October 2024, Stripe made its largest acquisition ever, agreeing to pay roughly $1.1 billion for Bridge, a stablecoin infrastructure company. In April 2025, Visa announced Intelligent Commerce and Mastercard announced Agent Pay, both built specifically to let AI agents transact on their networks. A month later, Coinbase launched x402, an open standard that lets AI agents pay for services automatically over stablecoin rails: one tiny payment per request, sometimes a fraction of a cent. When the payment giants start rebuilding around stablecoin rails, the “it’s all hype” era of those rails is over.

What blockchain can’t fix

If this were a hype piece, it would end there. It’s not, so here are the three things the pitch usually leaves out.

One: a blockchain can’t tell truth from lies. It can only remember. This is called the oracle problem, and it’s the single most important limitation in this whole discussion. A blockchain guarantees a record hasn’t changed since it was written. It cannot guarantee the record was true when it was written. If a scammer registers fake content as “authentic,” the blockchain will faithfully, immutably preserve that lie forever. Garbage in, garbage forever. This applies to every example above, including the good ones. A Guardian certificate proves what Umanitek’s detector concluded and when. It cannot make the detector infallible. If the detection model gets one wrong, the blockchain will preserve that wrong verdict with perfect integrity. Blockchain moves the trust problem to the moment of entry. That’s genuinely useful. It is not the same as solving it.

Two: the biggest real-world authenticity effort isn’t a blockchain. The leading industry standard for content provenance, called C2PA and backed by Adobe, Microsoft, Google, Intel, the BBC, Sony, and major camera manufacturers, works by embedding cryptographic signatures inside files. No blockchain required. It began shipping in a consumer product in late 2023, when Leica released the first camera that signs photos at the moment of capture. Blockchain can strengthen this system, serving as a tamper-proof public index of those signatures so records survive even if a company disappears or a file’s metadata gets stripped. But the fact that the flagship solution chose signatures first should recalibrate anyone claiming blockchain is the answer. It’s a component, not the foundation.

Three: there are other repair crews on this site. Regulation is one. The EU’s AI Act, in force since August 2024, requires AI-generated and manipulated content to be labeled; those transparency and deepfake-labeling rules took full effect on August 2, 2026. Platform-level detection is another. Government digital IDs may solve proof-of-personhood before any blockchain project does. Blockchain’s specific edge is the set of cases where no single company or government should hold the master record. That’s a real category. It’s also a narrower one than the hype suggests.

You won’t see the repair happen

Here’s my actual prediction, and if you’ve read my piece on banks quietly adopting blockchain, it’ll sound familiar: you will never consciously “use blockchain to verify AI content.” There will be no app for it on your home screen.

Instead, one day your phone’s camera will quietly register everything it captures. Your browser will show a small mark on images that carry verified origins, and the absence of that mark will start to mean something. The reviews you read will carry a “verified human” indicator you’ll stop noticing within a month. Your AI assistant will pay for things, and somewhere beneath that transaction, machine-speed payment rails will settle in seconds. No announcements. No tokens to buy. Just an internet that slowly re-learns how to prove things, because the honor system it used to run on is already gone, whether we’ve noticed or not.

AI didn’t need blockchain to be born. But the internet AI is creating, flooded with perfect fakes, tireless bots, and software that spends money, needs something that can keep records no one can forge and verify claims no one has to take on faith. That’s not a slogan. That’s a job description. And it happens to be the one thing blockchain was actually built to do.