In 2025, even a perfectly written sentence can be suspicious. The smoother the prose, the more likely someone is to ask: was this written by a machine? We have reached a point where authorship itself is on trial.
An NBER paper puts numbers on a question that feels existential: can we prove our humanity on the page?
A new NBER working paper by Brian Jabarian and Alex Imas, “Artificial Writing and Automated Detection,” brings much-needed empirical clarity to a noisy debate about AI-text detectors. The study evaluates leading systems on a carefully constructed corpus and asks a practical question: under what conditions can these tools reach policy-grade performance without falsely accusing human writers? You can read the paper here: NBER Working Paper No. 34223 (see also the Booth/BFI version (PDF)).
Their methodology is straightforward and rigorous. The authors assemble a 1,992-passage corpus spanning six everyday genres (news, blogs, consumer reviews, novels, restaurant reviews, and résumés). For each human passage they generate AI counterparts with four frontier models (GPT-4.1, Claude Opus 4, Claude Sonnet 4, and Gemini 2.0 Flash), and they also probe adversarial conditions by passing outputs through a “humanizer.” Across this landscape, they benchmark four detectors: Pangram, OriginalityAI, GPTZero, and a RoBERTa-based classifier.
The headline finding is measured but clear. On medium-to-long passages, Pangram uniquely sustains what the authors call policy-grade performance: within-sample, it achieves essentially zero false positives and false negatives at thresholds that are set both endogenously and exogenously. Performance weakens somewhat on very short texts, and humanizers erode accuracy across the board. Commercial tools tend to outperform open-source baselines. In sum: detection can be strong in well-specified settings, but it is not a universal oracle—and short or strategically altered text remains difficult.
What does “written by a human” mean in 2025?
Virtually no one writes without tools anymore. We rely on grammar checkers, search engines, translation assistants—and, increasingly, AI co-writers that suggest tone, structure, phrasing, even metaphors. If you use an AI assistant to rewrite a clumsy sentence, to clarify a thought, or to unblock a moment of confusion, does that make the resulting text less yours?
The answer should be obvious: no, it does not. And yet, the very existence of detectors—and worse, the way their scores are often interpreted—creates an artificial and dangerous line between “genuine human work” and “suspiciously smooth text.”
The risk of illegitimacy
The fear is not merely that a machine will be mistaken for a person. It is that a person, using modern tools responsibly, will be mistaken for a machine. This is not idle paranoia; it is a broader cultural anxiety. Writers, thinkers, and students—by engaging in the most natural use of available tools—may find their work discounted, questioned, or flagged. The very act of being an effective writer now—of using language with fluency, of seeking clarity through assistance—can get you labeled a cheat.
The damage is not theoretical. If institutions accept detector outputs as proof of inauthenticity, our notion of authorship—judgment, originality, intellectual labor—collapses into a statistical black box. A regression replaces reflection. A probability score replaces intention.
There is another dimension to this debate that is rarely acknowledged. Much of the public conversation and virtually all top academic publishing happen in English. For non-native speakers, writing clearly in that language has always required disproportionate time and cognitive effort—often unrelated to the quality of their ideas. AI assistance does not fabricate insight; it reduces linguistic friction. To oppose the responsible use of AI for clarity is, in practice, to defend a status quo in which fluency functions as a gatekeeping device.
Seen from this angle, resistance to AI-assisted writing can resemble a subtle form of protectionism. Native speakers have long enjoyed an invisible advantage in global debates and scholarly publication. When new tools reduce that asymmetry, the instinct to police their use deserves scrutiny. If the goal is equal competition of ideas, then lowering the linguistic barrier strengthens meritocracy rather than undermines it.
Finally, dismissing an argument by claiming “this was written by AI” is not a serious critique. It does not engage with the substance of the ideas; it sidesteps them. If a claim is wrong, explain why. If the reasoning is flawed, expose the flaw. But attributing authorship to a machine is not an argument—it is a way of avoiding one. Intellectual debate should be about evidence, logic, and interpretation, not about stylistic suspicion.
In fact, such accusations often reveal more about the accuser than the accused. When the response to an idea is to question its origin rather than its content, it suggests a lack of counterargument. A healthy intellectual culture demands disagreement at the level of reasoning, not insinuation at the level of process.
Shouldn’t we be rethinking what we value?
What we need is not ever-stricter detectors, but better ways to understand human-machine collaboration—and a deeper respect for intention. The author’s criteria, vision, and ideas matter. A text can be assisted without being artificial. A sentence can be generated without being fraudulent.
It would be odd to tell a painter that a canvas “doesn’t count” because Photoshop touched the initial sketch, or to reject a symphony because it was first composed in Sibelius. Tools have always been part of creation. What matters is what we make—not the imagined purity of the process.
If the future of writing devolves into judging every line by its algorithmic origin rather than its beauty or depth, we turn authorship into paranoia—a kind of literary witch trial in which writers must prove innocence to a machine that cannot read intention.
But there is a hopeful path. We can re-enchant writing not as a proof of humanity, but as an act of meaning. What we say, why we say it, and how it moves others—that is where value lives.
A personal note on writing and AI
Consider a concrete example. I recently published a Substack essay, “Lightness with Conscience,” a reflection on the tension between freedom and responsibility. Those ideas had lived in my head for years. I finally wrote the piece in a single afternoon, prompted by a co-author who asked why I sometimes take public positions that might carry professional costs. Einstein, after all, once suggested that in academia the safest strategy is to keep one’s mouth shut.
I did not write alone. I used ChatGPT-4o to help edit, refine phrasing, and push the draft into clarity—and I acknowledged it. The ideas were mine, and so was the voice—but the process was dramatically faster, perhaps a tenth of the time. The result was, I think, more elegant. More importantly, it existed at all: without these tools, I might never have written it.
The upside of cheap expression
There’s another dimension we seldom acknowledge. By lowering the cost of drafting and sharing, AI allows more ideas to reach the public sphere. This very Substack is a case in point: without an assistant to help with editing and structure, I likely wouldn’t have launched it—or would publish far less frequently. From the standpoint of culture and scholarship, that’s a net gain. Technology, after all, has served the same purpose for millennia: to enhance the way we create.
We cannot un-invent these tools. But we can decide how to use them—and how to judge what others make with them. The Jabarian–Imas study shows that detection can be made policy-grade under specific conditions; it also shows the limits of any one score. Rather than gatekeeping authenticity through opaque classifiers, let us build an ethic that honors ideas, intention, and the courage to write in public, whatever tools we choose.
We can waste years in witch trials, asking machines to certify who is “truly human.” Or we can move past detection and rediscover what really matters: intention, meaning, courage. A sentence is not valuable because a detector gives it a score. It is valuable because it says something worth hearing. In the end, the soul of writing does not live in its origin. It lives in its meaning.
*** Disclaimer: I used ChatGPT-4.0. as an editorial and language-refinement tool. The ideas and arguments are entirely my own, and I take full responsibility for them.
Of ;)
Oh!!! The Hut if that last line. Smooth.