Proof of Person
Pangram, Dead Internet Theory, and the new, regressive labor of being human online
Last week, Substack did something extremely funny. In partnership with Pangram, the current least-worst player in the burgeoning space of AI-powered AI writing detectors, they gave every human user the ability to scan any new 100+ word post for non-human writing. Writers can also pre-scan drafts before posting, and have the option to switch AI detection off altogether. However, if they do that, any readers will get a notice that the author has taken the easy way out—a refusal that arguably concedes more ground than a clean scan gains. Already, people are losing their minds.
Jay Springett described the update as the moment Substack entered its “bloodsport era,” and I’m hard-pressed to disagree. Just as 17th-century Harvard was not immune from the simple human joys of Witch Hunt hysteria, so too, I imagine, is a website full of current and former gifted kids not immune from the joys of casting doubt on peers in a time of change. From here on out, the doomsayers suggest, every new piece will be subjected to a rigorous slop-interrogation that is far from false-positive free. Freddie deBoer has already offered a worrisome trial run of this thesis, feeding Pangram a 300-word excerpt from one of his own essays (verdict: 100% AI-written, high-confidence), followed by the full essay (verdict: 100% human, high-confidence)—and his “brittle” diagnosis is hard to improve on. Anyone who fails the vibe check will be branded with the scarlet letter of slop-raker, or worse, sloptimist. Or something like that. Meanwhile, others — like Substack CEO Chris Best — have heralded the update as a push for transparency in a muddied-and-getting-muddier Web 2 media ecosystem. For a platform that already leads the rapidly dwindling pack of text-focused internet platforms with an enviable ratio of human signal to machine noise, I can see the arguments either way—but for better or worse, I doubt the new era we’re debating will last long.
In 2022, when the earliest Stable Diffusion image models went live for public consumption, it was exceedingly easy to tell when an image was generated. Now, it is increasingly not—and as much as I pride myself on being a close reader and intentional writer of natural human language, I can feel my relationship to text starting to change in the same way. If even the most conservative medium-term projections about the capabilities of AI end up paying out — and I’m not entirely convinced they won’t, hype aside — expecting tools like Pangram to continue to measure a text’s humanity in any truly empirical and systematized way could be a tall order. This is not to say that the anxieties inherent to our current time of growing pains are unfounded: the prospect that we are only a few years away from a world where it is honest-to-god impossible to tell whether a machine made anything you see online is not a reassuring one, and I don’t trust anyone who says otherwise.
Unfortunately, a good chunk of this piece will come back to the recent statements and actions of one Samuel Harris Altman—starting with a post he made last September on X, The Everything App.
While this post’s replies are a predictable flood of the Tim Robinson hot dog guy photo, the basic picture is worth sitting with. Here we have the CEO of OpenAI, the company arguably most responsible for the generative AI boom that has filled the internet with slop, claiming to be discovering that the internet is full of slop. This observation is not new.
Patriots in control
On January 5th, 2021, a user with the handle IlluminatiPirate logged onto a forum called Agora Road’s Macintosh Cafe and made a post. The title: “Dead Internet Theory: Most of the Internet is Fake.” The eponymous theory didn’t start here — Pirate begins with a disclaimer that it originated on the imageboard Wizardchan — but the post is widely agreed to be the cleanest initial synthesis of what it’s all about. In brief: since the mid-2010s, the internet has been quietly hollowed out and refilled. Now, in the 2020s, every single post, meme, trend, reply, like, comment, view, and subscription we see is the work of an immense network of bots. Critically, none of this is accidental, either: “The U.S. government,” the post explains, “is engaging in an artificial intelligence powered gaslighting of the entire world population.” This is followed by a clip from Metal Gear Solid 2.
Many people, myself included, have since absorbed Pirate’s argument with all the detached amusement that goes along with any sufficiently vivid conspiracy theory. The initial leap from “most internet engagement is inauthentic and some metrics are inflated” to “The Patriots from Metal Gear Solid are real, and they have trapped you in a digital Torment Nexus” is fun in the same way that the fantastical maps made by some members of the flat earth community are fun. When I first encountered the theory, a few lines of its most grandiose context were enough to earn it a spot on that same mental shelf. Now, a few years later, I can’t help wondering if that was an oversimplification.
What sets Dead Internet Theory apart from many comparable theories is the way the world keeps moving toward it—not in the loose, abstract way that theories about the Kennedy Assassination eroded public trust in institutions, but in a way that leans more on charts full of “hard” numbers and “raw” data. If we accept Pirate’s assertion that all machine-made content is “dead,” the internet is becoming measurably less alive every year. As early as 2024, a full half of all web traffic came from bots—but the theory didn’t need those numbers on its side to start drawing real human traffic.
Within a year of Pirate’s original post, the Atlantic fielded a piece titled “Maybe You Missed It, but the Internet ‘Died’ Five Years Ago”—an exemplary case of a mainstream outlet inspecting a bit of possibly radioactive low digital culture to see if there are any kernels of truth inside. Then, to paraphrase Adam Curtis, something strange happened. While many members of the Atlantic crowd seemed to find the theory just as compelling as its original audience, most of them were far more interested in the corpse than the alleged murderer. A subsequent surge of academic engagement wasted no time excising any lingering conspiratorial elements with surgical precision. Now, five years out from that first formative post, the version most people invoke is curiously devoid of Pirate’s original man-behind-the-curtain.
To be clear, I don’t think the internet is dead — either in IlluminatiPirate’s sense or the current more anodyne definition — and I don’t think it will ever be 100% devoid of human intent. If I’m wrong about that, I will have several much larger problems to contend with. But I can’t deny that the descriptive half of the theory, the part that argues that we are all, statistically, increasingly alone in here, has become the weather. As Dead Internet Theory got propagated over the last few years, it got a curious promotion: not from false to true, but from conspiracy to metaphorical description. The map was deemed accurate, but the dragons in the corners got papered over.
In IlluminatiPirate’s original articulation of the theory, some kind of architect was needed; surely, a catastrophe this big couldn’t have happened on its own. This is a completely understandable and human impulse. Still, as loath as I am to take the less fun side here, it’s probably wrong—not because I don’t think that there are architects capable of this kind of psy-op, or even that they haven’t been trying to spur things along, but because they’re definitely not the only factor in play. In much the same way that the world was overtaken by global capitalism over the last four centuries, most of the actual heavy lifting that has made the internet feel so dead has happened diffusely, at scale, through millions of ordinary commercial decisions. Like Alex Jones with his much-vaunted “gay frogs,” Dead Internet Theory’s only shortcoming was to underestimate the sheer entropic potential of the market. No one wanted it to happen; it was simply cheaper than any alternative. This is why I’ve made the difficult decision to take off my parapolitics hat and treat Dead Internet Theory as the domesticated metaphor it has become. For now.
Once we accept the idea that the internet is at least metaphorically dead, we’re faced with another equally existential question: how do you spot another person when you find one? How do I prove that to you, reading this, right now, in this sentence—and how can you reliably check? So far, the answers seem to be self-selecting into competing camps—one top-down, and the other bottom-up. Surveying this same territory for Big Think earlier this year, Andrey Mir proposed three — digital IDs, shows of effort, embraced flaws — and warned that the effort route inevitably curdles into performative “labor signaling,” but in my view, his last two categories have already collapsed into a single bottom-up act. Neither of these response categories seems to work, but they fail in usefully different ways.
Who up pondering they orb
The answer you’re most likely to see discussed in The News is verification: delegating the whole problem of certifying humans over machines — or vice versa — to a trusted institution or platform. A few competing ways of making this work already exist, many of which feel like holdovers from the previous version of this problem: how to distinguish Special Internet People from Regulars.
Up until Twitter became X, the standard of internet verification was the blue check—a symbol used to assign a given user the role of “verified entity.” Before the takeover, the symbol was reserved for journalists, celebrities, brands, publications, and other accounts in high-trust and high-influence positions that others might seek to emulate for nefarious reasons. Then, after the rebrand, new management locked it behind a subscription—leaving the door open for anyone with $8 of monthly disposable income to ascend to verified entity-hood. While this can be understood as a failure of custodial duty, it’s also an instructive example of how we view verification itself. If institutions see it as a safeguard, platforms see it as an inherently scarce commodity to be monetized. This means that, as a response to the big, scary question of “who’s there,” verification is only as good as the organizations doing it—which are increasingly the ones who have noticed that proof can be a product.
The fully realized version of verification-as-commodity can be found in World, formerly Worldcoin. The company’s flagship device is a chrome sphere about the size of a bowling ball, called the orb, which is used to scan your iris. Once World confirms that your iris does not belong to anyone else in its ledger, it issues you a cryptographic credential called a World ID. This, per the company’s materials, is a kind of digital passport that you can present to assert that a human being was, at least once, physically present behind whatever account it’s being used to back up. For World, proof of personhood “can be thought of as the first and most fundamental building block in establishing digital identity”—and while the company spent several years relegated to the questionable jungle of Crypto Guys, it is slowly beginning to migrate into more serious circles.
World was co-founded in 2019 by Sam Altman. Six years before he started tweeting about Dead Internet Theory, he was already laying the groundwork to sell people a solution to the problem his other company might create. Some people might call this a cynical business strategy, but I think the more honest — and worse — reading is more like consistency of vision. The problem and the solution both ladder back to the same investment thesis: being human online used to be free, and now it’s not.
It brings me no pleasure to acknowledge that this is exactly what has just happened with Substack, too. Pangram is an orb for prose—detection-as-a-service and personhood-as-a-service are the same product, rented from a platform, which is why I don’t see the fears that the tool will spark any number of witch hunts as a bug in the partnership. This cohort of anxious, agitated human writers and readers isn’t the byproduct of Pangram, but its most highly coveted market.
Clear surveillance objections aside — I’m not the first person to discuss these guys — I think the orb has a much larger conceptual problem. On paper, World ID exists to defeat Sybil attacks by proving that a distinct human body enrolled. It cannot prove — and, to their credit, does not claim to prove — that the body in question writes its own posts. An orb-scanned human with Claude or Grok can still pass every check in World’s toolbox; the company treats humanity as a binary, quantitative login credential rather than a quality. When I want to know whether the sentence I’ve just read was cooked up in a meat computer or a datacenter, my first reaction is generally not to demand to see some biometrics; I want to know what the sentence cost the person who wrote it.
Beyond badges and scanners, there’s another move in the top-down toolbox worth noting: the possibility of adding more tools. Last year, the Australian creative agency Cocogun debuted a piece of punctuation called the am dash—billed as “unmistakably human, unusable by AI,” a symbol of “real pondering, genuine daydreaming, and true editorial wordsmithery,” and shipped inside two typefaces called Times New Human and Areal. The name is Descartes by way of advertising copywriter humor — as in, “I think, therefore I am” dash — and the implementation is even better: the am dash has no Unicode point; the only way to type the certified-human mark is to install a proprietary font and let a ligature insert it. All of this is charming and completely dead-on-arrival. After all, the only thing that signals AI writing more reliably than the em-dash, at least to me, is the last batch of symbols we invented to assert emotion in an age of automation: emojis. A signal handed down from above arrives already-legible, and in the age of LLMs, legible is another word for dead. The crowd seems to know this instinctively, which is why most bottom-up answers to the problem look a lot less like the DMV.
Come—and take it
Somewhere around mid-2025, it became common knowledge — a phrase doing enormous unexamined work we don’t have time for here — that the em-dash is the unmistakable signature of Claude, Grok and company. Many human writers were quick to point out that this is the natural result of LLMs being trained on millions of human-written books. They are, simply, the most flexible and generous form of punctuation, which is presumably part of why they show up so much in stereotypical LLM text: the machines learned to write from us, and they picked up our best habits first. However, plenty of other writers have succumbed to peer pressure—deciding that every em-dash now costs them credibility they can’t spare.
Over the course of about 18 months, a 400-year-old punctuation mark has been branded a mark of the beast. Teachers call it out in homework, copywriters report clients rejecting copy over it, and people who don’t read or write have latched onto it as an immediate and immutable tell that Thing Bad. None of this has gone unnoticed by the major AI players, either. Last November, two days after GPT-5.1 shipped, Sam Altman announced a “Small-but-happy win”—tell ChatGPT not to use em-dashes in your custom instructions and “it finally does what it’s supposed to do!” When I started writing this piece, I really did not plan to have this guy holding the knife behind every pain point, but here we are.
The underlying pattern behind the em-dash war is perhaps the most unsettling part of the whole equation I’m trying to unpack here. The most prevalent bottom-up response to the question of humanity in the age of AI involves abandoning our best habits because the machines also have them. Some human writers are deliberately writing in flatter, plainer, and less precisely jointed ways to be more legible as human. Others are putting typos in their resumes to scan as human to AI resume-reading tools. The internet may not be comprehensively dead, but a world where a polished sentence is the enemy and lapses in spelling and grammar are armor bears little resemblance to the one I was promised in middle school English class.
I’m no less implicated here than anyone else. At a few points throughout the last year, my love of em-dashes has prompted me to do something that, from any kind of sensible writing standpoint, is kind of insane: upon writing a sentence and looking back at it, I will feel compelled to redo it so that it scans less like AI. I don’t seem to be alone here, either. Of course, we’ve been leaving grammar out of sentences for as long as we’ve had digital sentences to exchange—I only add a period to a text message as a modifier for whatever I’m trying to say, or if I want to avoid double-texting. But those little roughnesses are used to convey tone for their intended reader. The new adjustments I’m describing are addressed to another reader, the one standing over your shoulder and checking your work.
This phenomenon isn’t unique to written text, either. Instagram is filled with badly-lit photos taken by a new cohort of digital natives for whom competence now scans as synthetic, and the platform itself has converged on this logic from the other end. Meta’s own Adam Mosseri has argued that it will soon be “more practical to [cryptographically] fingerprint real media than fake media” and has suggested that creators lean into “unflattering” images to prove they’re real. This, too, is a form of verification: in a world where it’s getting harder to tell the difference between fake money and real money, you stop stamping the fakes and start stamping the real thing. Unfortunately, there’s no EXIF data for text.
Audience Captcha
Everything happening here has another purpose, beyond any actual signal-from-noise utility, and it’s worth giving that a name: proof-of-person rituals. Costly ethological signals — best example, the kind of biological looksmaxing practiced by peacocks — that a human was here. Every slop machine in the world is built and continues to be sold on the promise of cutting costs—and in response, cost itself can, and has, become proof.
These rituals are not a new and unprecedented reaction to new and unprecedented technology. One need only look back to the Unicode-heavy album titles of mid-2010s Seapunk and Witch House — genres heavily informed by the exact techno-cultural moment when the internet allegedly stopped being alive — to see the same logic brought to bear the better part of a decade before anyone started worrying about LLMs. The niche electronic producers of the 2010s were paying to be illegible before there was a machine to be illegible to—or at least, when the machine was still just good old-fashioned SEO. This also suggests we may not see the last of these rituals when the detection era inevitably ends.
The problem with proof-of-person rituals is that they almost always end up feeding the counterfeiters they’re meant to ward off—something you might already understand if you know anything about CAPTCHA. For two decades, the Completely Automated Public Turing test to tell Computers and Humans Apart has been the internet’s default PoP ritual: forcing human users to prove their humanity hundreds of millions of times every day by copying down cryptic phrases and identifying pictures of stop signs. It’s the common ancestor of this essay’s two camps — simultaneously administered from above and performed as a tiny ritual below — and also constitutes one of the largest human-assisted machine learning programs in history.
The first iteration of CAPTCHA required users to prove their humanity by using their real-human eyes to read and transcribe mangled scraps of text. This labor — routed through reCAPTCHA — digitized Google Books and the New York Times archive and taught non-human readers to read. Text CAPTCHAs continued until Google’s own researchers were cracking even the most distorted bits of text with 99.8% accuracy, at which point the test was deemed to have served its purpose and new techniques were rolled out.
Since then, human CAPTCHA participants have been presented with grids of images — most conspicuously automotive — and asked to identify things like bicycles, crosswalks and stop signs. From this, the machines have learned to do the same—and while the test has yet to be called off, it is rapidly approaching obsolescence once again. Other forms of CAPTCHA also exist, like “I’m not a robot” checkboxes that analyze the imperceptible quivers in a user’s mouse movement to determine their humanity. But these two stand out for the way in which the proof-of-person work they gather has been harvested to build the exact capacity that invalidates those proofs.
Within this logic, it’s clear that the arms race of writing to sound human online is a similar losing game for anyone taking it seriously. The moment that “typos=human” becomes a legible enough trend to train on, asking an LLM to write in a more human style may start to yield text with one or two plausible typos. No matter how we try to shift the meta of stringing words together online, the models will always be one fine-tune behind us and closing fast. Biologists call this a Red Queen’s race — not a race with winners, but one that’s run to stay in place — and if the various evolutions of CAPTCHA are any indication, our competitors get better and faster with every lap. The style rotations, typo armor and de-em-dashed sentences are all public, which means they’re already bound for the next batch of training data. Even if we end up adding new punctuation, the models will absorb it, requiring even more new punctuation, and so on. What we are talking about here is an arms race—and if there’s one thing we should all know by now, it’s that the winner of any AI-related arms race is the idea of AI itself.
The new cost of being human
We now arrive at two ugly truths:
Truth #1: because being human online is no longer free, the act of proving you are human has become labor—and that labor is regressive. Every proof-of-person ritual takes time, skill, and social capital: knowing the current metas of human and AI writing, having an audience that will give you the benefit of the doubt, and being a good enough writer to sound deliberately or charmingly rough instead of just rough. But not everyone is subject to these rituals. If you’re already an A-list novelist, a blogger with a significant following, or the kind of writer who gets tapped by the Atlantic to write about imageboard cranks, you already have numerous forms of verification at your disposal and pay almost nothing to maintain your humanity. Meanwhile, everyone else is stuck paying full price—and that price, like everything else, just keeps getting ratcheted up.
Those who slip through the cracks will be the usual suspects: ESL speakers with technically correct but contextually naïve writing, older people who make the mistake of adhering to conventional grammar, and outsiders and newcomers with no history or record to point toward. More broadly, the people inconvenienced will be the ones who already go overlooked and thus aren’t tested for. There is a word for a system where members of in-groups with established positions of power are authentic-until-proven-fake while everyone else is expected to justify their humanity 24/7—and it’s not “new.”
Of course, a response to this conclusion exists (“who cares?”) and yes, in a healthier, more well-slept society, this could be a non-issue. It just so happens that this specific issue strikes a nerve with any number of people who are freaked out about the underlying technology behind it all because it’s automating them out of their jobs and making their neighbors sick. And again, the target audience of all this proof-of-person infrastructure is not AI-boosted writers trying to sound more human, but justifiably anxious human writers who have suddenly found themselves in an AI-boosted wilderness of mirrors.
Truth #2: Verification from above and signaling from below are not mutually exclusive or separate answers, but competing constitutions—and we will ratify one by default. Verification says humanity is a credential that can be issued — and by the same logic, revoked — from a central authority. Signaling says it’s a texture—something judged contextually by a crowd, and thus unfair in all the ways that crowds can be unfair. The process by which we generally determine “who’s there” is being sorted out right now by the interplay of B2B SaaS providers and paranoid newsfeed drama—but it doesn’t have to be. The most hopeful answer I’ve seen to this question comes from Renée DiResta, who calls proof of personhood “constitutional infrastructure” and argues that the most equitable way out of this mess will start by drafting an answer together, with guardrails. Personally, I have my doubts that we’ll get anything like an open convention, much less a vote.
It’s rituals all the way down
In a recent piece for the Institute of Network Cultures, I argued that 4chan greentext could be a curious form of proof against all this. By prefacing every line of a post with >, the writer tells a certain kind of tapped-in reader any number of things—not least of all that they should give up looking for another person behind the text. There is no “me” behind “>be me” beyond the person reading it while they’re reading it, or the person writing it while they’re writing it; just a set of instructions you’re meant to execute on your own hardware. You can’t counterfeit a self that was never asserted. At the time I thought I was describing a novelty, a form of accidental immunity cooked up by a bunch of NEETs. Now, it feels more like the future. I’m not happy about it, either.
Every form of writing that asserts the existence of an interior is now in the business of defending that assertion at an ever-rising cost against a counterfeiter with the most cartoonish blank check budget in history—and the forms that survive a race like this may not be the ones that best prove interiority, but the ones that never needed it. This brings us back to the question of how much the coming Bloodsport era — on Substack, and everywhere else — should actually be taken seriously.
It’s tempting to conclude that to give these technologies such an existential amount of weight is to assign them more power than they’re worth. Rather than interrogating every piece of writing for any trace of our current interpretation of inhumanity, we could let this new age of uncertainty free us to engage with anything the way the average /r9k/ user engages with greentext. Some would say this is naive. Others would say it’s the healthiest available attitude we can take. Again, I’m torn; while I’d like to say I’m trying to free myself from the checking impulse, I’m still an nth-generation mostly-English guy from the North Shore of Boston. Purity testing is kind of our thing.
I’ll end this piece the only way that I could conceivably end it, and I would apologize for being cute if the trap were not the point: this entire newsletter is a proof-of-person ritual. Every time I write one of these, I am doing a big costly signal, mailed to you on a loose but consistent schedule, whose underlying message is always that I am here, spending my time, energy and attention on one random weird thing or another. If you subscribe, you’re subscribing on that theory—and if you want to take this opportunity to click the three-dots button and check my work, I won’t judge.
In his piece announcing the Pangram partnership, Chris Best concedes that the tool can only estimate whether AI was used, not whether a post reflects genuine human care. A machine text laundered by a human — or a human text laundered by a machine — may avoid the gallows at any future witch trial. In that sense, no matter what the scanner says, the real verification process is all you—weighing the em-dashes, which you may consider suspicious, against the laborious displays of self-awareness, which may also be suspicious, against whatever residue of costliness you think you can feel in any given sentence. All I can really offer is the wager the whole system runs on: that this document was costly in ways that have not been optimized for anything. Hopefully, you can still tell.
Absolute banger. Thank you for writing.
Myles for the win once again.