Modern democracy rests on a fragile promise: that strangers can reason together about hard issues across differences in identity and values. For centuries, clear writing was the cultural technology that carried that promise. People who didn’t know or trust one another could meet on shared ground, weighing reasons they could examine for themselves. That promise is now coming under strain from an unexpected direction.

This shift became uncanny when I put a piece of my own writing into GPTZero, one of the AI-detection tools now common on college campuses. After a brief "processing" pause, complete with inspirational quotes from writers like George Orwell, the tool flagged my work as "AI-generated." Its scoring method is a proprietary black box, but it still offered confident explanations for each sentence:

Lacks Creativity: uses a straightforward and simple structure.

Formulaic Flow: transitional phrases maintain a logical progression.

Robotic Formality: presents information in a polished, orderly way, with a focus on clarity.

Writing guides like Strunk and White’s The Elements of Style have long taught that clarity is a virtue because it exposes you. You’re not hiding behind a smokescreen of convoluted language; you’re saying: here’s what I think, here’s why, show me where I’m wrong. It’s the act that makes honest disagreement possible. Today, that act can be misread as concealment rather than exposure.

GPTZero then urged me to buy the premium upgrade, not just for more detection runs, but to boost my "writing skills." A pop-up offered to "humanize" what I wrote. I gave it a shot. When I ran the humanized version back through GPTZero, it came back as "100% human." The writing wasn’t awful. It was just more colloquial. Phrases were clipped in ways that blurred the meaning. The grammatical parallelisms I’d worked to preserve were gone, replaced by looser, chattier constructions. It sounded nothing like me.

Detectors bring out the absurdity of a machine instructing a human how to sound more human, but they’re symptoms of a larger unease. What keeps them alive is a widely shared wish that somewhere in our writing there’s still something a machine can’t fake. It’s easy to want that wish to be true, but even if GPTZero were a perfect detector, it still wouldn’t tell us what we care to know. A person can prompt a model, lightly edit the output, and pass along clever or misguided ideas they don’t understand; nothing in a detection result reveals that. And the question remains: was anyone home? 

In 2010, asking whether a paragraph came from a thinking mind would have sounded like asking whether tables are solid. Today, a careful email might reflect genuine consideration, or it might be something AI produced in seconds. For the first time, producing text is cheaper than reading it. Suspicion is a reasonable response to that inversion.

AI didn’t invent AI-like writing. Corporate copy, marketing fluff, and "inspirational" social media posts often had the same smooth, impersonal tone long before AI existed. That kind of writing was always aimed at producing effects—buy this, feel this, comply. AI just gave us a label. What’s new, and unsettling, is that writing where someone makes themselves vulnerable by being clear, offering their reasoning, inviting scrutiny, now gets treated with the same suspicion.

The reason isn’t that AI is especially clear. Anyone who has tried to write seriously with these models knows that it still takes real effort to get them to produce anything genuinely lucid or precise. Much of AI’s usefulness comes from giving you something to push against, like reading widely and indiscriminately in fast-forward. Some writing seems muddy until the idea snaps into focus; AI often does the opposite, giving you the feeling of clarity until you go looking for the idea and find there isn’t one.

What AI reliably produces are surface features of clear writing like smooth pacing, tidy paragraphs, and predictable transitions. Detectors pick up on those. So do readers. In my classes, when I read new work that sounds careful and composed, students lean back, not in boredom, but in a kind of collective plausibility scan. A pause, then a few sideways glances. The cues that once signaled "a human mind has taken great care" now carry a double meaning. Colleagues across campuses see it too.[1] Polish has become a liability.

When a piece of writing is taken as possibly non-human, readers hesitate, and the suspicion makes it hard to follow an argument on its own terms. Steven Pinker described clarity as a "window onto the world": the language recedes and the idea comes into focus.[2] But for many today, the window flattens into a polished surface. Once the language itself becomes the thing we notice, clarity is no longer doing its job.

Clear writing was never about elegance or flair. Its function was to strip away enough ambiguity so that what someone meant could travel across distance, time, and difference. For a long time, clarity served that function: a 20-year-old student in Philadelphia could make sense of a 60-year-old logician in 19th-century Jena. Clarity didn’t make everything simple, but it signaled that another mind, however distant or different, could grasp the same thought. The practice of public reasoning that it made possible was quietly anti-tribal and anti-authoritarian. You didn’t have to be someone’s cousin or take anyone’s word for it. You could look at the reasons yourself.

As clear-looking writing stops signaling a human mind at work, writers feel pressure to signal their humanness through style. Some adapt by breaking polish, leaning on roughness, quirk, and emotional rawness. Others respond by raising the bar on what counts as human. If AI delivers a flat, interchangeable clarity, then human writing, the thought goes, should push toward richer, more personal forms, with more room for wit, texture, and a distinctive voice. 

Those buffered from immediate suspicion tend to say, "Just write normally. Don’t use AI and you’ll be fine." But writing has always been a technology, something you learn and practice, and "normal" only works when everyone agrees on what it means. With that shared sense gone, the advice stops being comforting and turns into a gamble: normal to whom?

The trouble is that voice, wit, and personal style only work among people who already share your sensibilities, who hear your humor as warm rather than smug, your tone as familiar rather than off-putting. A stranger across distance or difference, someone who doesn’t know you, doesn’t trust you, and doesn’t share your background, has no reason to take those cues the way you intend. Some writers do have voices that travel across time and culture. But we can’t ask every writer—let alone every citizen—to stand out stylistically before anyone will even listen to what they’re saying.

Clarity offered a way around that problem. You could write in any voice you liked; as long as you were clear, how you sounded wasn’t the price of entry. If someone dismissed an argument simply because they disliked your style, they were being superficial. Personality was optional, not because it didn’t matter, but because it wasn’t guaranteed to travel. Clear writing let you address a stranger without second-guessing whether your style would get in the way, trusting only their ability to follow your reasons. It was the one mode of address that didn’t require the audience to already be on your side.

That was its quiet democratic power. But if clarity now has to be dressed up in style to be recognized as human, the ground shifts. Any stylistic signal shared widely enough to function in public is also something a model can learn to imitate. What’s left are private signals, cues that work only for people who already recognize them, and only for a while.

Before even asking: “Does this argument make sense?,” readers ask: “Do I recognize this voice? Does it sound like one of us?” Those questions favor insiders and screen out strangers. The plain, self-effacing prose that once said don’t look at me, look at the argument no longer earns a hearing, and the circle of who gets to address whom shrinks.

Bias made access uneven. If you spoke the wrong way, you might be judged unfairly for it. But arguments could still get a hearing, and anonymity or impersonal prose offered outsiders a back door. That door is gone. For the true stranger with no network, no name, and nothing that signals one of us, there is no point of entry at all.

When arguments must first pass through cues of belonging to be treated as more than output, public reasoning starts to break down. The practice that lets strangers think through hard issues together was always fragile. But democracy cannot afford to lose this at a moment when trust in public institutions is low and it’s easier than ever to retreat into camps. Replacing shared standards of reasoning with private handshakes—signals only insiders recognize, and only briefly—pulls us in the opposite direction of what a democratic society requires. [3]

The suspicion surrounding clear writing comes from the collapse of an old assurance that prose that looked clear meant someone had actually done the thinking. That assurance can’t be restored by hoping that, as AI-generated text spreads, we’ll get better at sniffing it out. Nor can it be restored through personal disclosures. If I tell you I didn’t use AI, you might believe me because you already know me. A stranger won’t, and practices meant to include strangers can’t rely on pledges. When the world changes in a way that removes something we once took for granted, sharper vigilance and better policies rarely bring it back.

Before modern sanitation, clean water depended on local conditions: the quality of wells and rivers, informal norms about where to dump waste. Most of the time, the system worked. When it didn’t, people got sick, and the failure was blamed on individual habits or carelessness. But once industrialization began releasing massive amounts of waste into the environment—everywhere, out of sight, and hard to trace—those safeguards stopped working. You needed pipes, treatment plants, standards, and public oversight.

AI companies are now releasing a massive amount of intelligence into the environment. It’s not human intelligence, but it performs many of the same tasks, and there’s a lot of it, everywhere. Intelligence is an enormous social good but only when someone is accountable for it. Before AI, intelligence was contained inside people. Now it floats free. Asking us to sort it out based on style is like asking us to smell the water and decide whether it’s safe. We don’t need better noses. We need new public infrastructure that governs how intelligence flows, so authorship stays with people.

Just as sanitation wasn’t left to factories certifying the water supply, the infrastructure for governing intelligence can’t be left to AI companies. But it can’t be handed over to government control either. Thinking, reasoning, and expressing ideas are exactly the capacities we don’t want the government certifying or policing. A system where states decide which writing reflects "real thinking" would be rightly resisted.

The better model is the internet. When your browser connects to a website, neither side knows or trusts the other. And yet information moves back and forth reliably, because both sides follow public rules that verify the connection. That moment of verification is called a handshake—and the name is apt. A handshake evolved to let strangers interact without personal trust: a public signal that certain conditions are met, so exchange can begin. 

We need something similar for cognitive work in the age of AI: a public handshake that replaces today’s private cues. That means shared rules for how work is produced and exchanged, built into the systems people already use. Just as a handshake never guaranteed good faith, an AI handshake wouldn’t guarantee thinking. The handshake would only signal that the work passed through a process where thinking could not easily be outsourced.

What counts as "having done the thinking" would vary by setting. Using AI to translate a document may be perfectly compatible with scientific work; in a language class, it substitutes for the very thinking being assessed. Whatever form it takes, such a system can’t rely on surveillance, any more than sanitation inspects each glass of water or stations an inspector in every kitchen. Good infrastructure is unobtrusive. When it works, it recedes, so you can focus on what matters.

Without such infrastructure, AI suspicion pushes us back toward insider signals, and democracy loses one of its few ways of reaching across divides. You shouldn’t have to signal belonging to any group to be recognized as human. When the measure of humanity shrinks to signs only insiders recognize, who wouldn’t opt for the dignity of being miscounted as one of the robots?

Notes

[1] Mintz, Steven. “Apparently, I’m a Bot.” Inside Higher Ed, 2 Apr. 2025, www.insidehighered.com/opinion/columns/higher-ed-gamma/2025/04/02/apparently-im-bot.

[2] Pinker, Steven. The Sense of Style: The Thinking Person’s Guide to Writing in the 21st Century (New York: Viking, 2014).

[3] Eli Alshanetsky, “When AI Dissolves Trust: Education Can Pioneer New Infrastructure,” Society (2025).