⏱️ Lectura: 10 min
78% of developers stop reading a piece of text as soon as they suspect an AI wrote it, and 71% avoid that author forever, according to a survey of 668 developers cited by Bryan Cantrill, CTO of Oxide Computer.
📑 En este artículo
Cantrill published The Revolt of the Reader on September 5, 2026, putting Pangram 4, Pangram Labs’ AI-generated text detector, at the center of his argument, with an accuracy he describes as surprisingly high.
TL;DR
- 78% of developers stop reading if they detect AI, and 71% avoid the author afterward.
- 98% of respondents prefer their own imperfect writing over text polished by AI.
- Bryan Cantrill published “The Revolt of the Reader” on September 5, 2026, on his blog.
- Pangram Labs released Pangram 3 in late 2025 and Pangram 4 just over a month ago.
- Cantrill compares AI text detection to the war on spam in the 2000s.
- Oxide Computer’s RFD 576 states that using AI to write voids the reader-author social contract.
What happened
Bryan Cantrill has been writing long essays on dtrace.org for over a decade, almost always about systems engineering and work culture at Oxide Computer. The one published on September 5, 2026 breaks that pattern: instead of talking about hardware or Rust, he explains why he stopped trusting other people’s writing.
His starting point is personal. Cantrill says he notices obvious signals in text substantially assisted by an LLM: empty transition phrases, self-congratulatory statements (“and that’s why this approach matters”), and a uniformity of rhythm that a trained reader spots within seconds. That suspicion used to be subjective. Now, he says, there’s a tool that confirms it: Pangram 4.
Cantrill himself had anticipated much of this argument in RFD 576, an internal Oxide Computer design document: writing with substantial AI assistance, without the reader knowing, voids the social contract between author and reader. That contract, he says, only obligates the reader to work at understanding a sentence if the author also worked at writing it.
Pangram Labs released Pangram 3 in late 2025. Cantrill adopted it as a daily tool and reports a very low false-positive rate: when the model flagged a text as mostly AI-written, the suspicion was confirmed almost every time. Just over a month ago the company released Pangram 4, which Cantrill describes as a step-function leap over its predecessor, with an even lower false-positive rate and fewer false negatives as well.
Context: the war on spam repeats itself
Cantrill draws a direct parallel with unwanted email. In the early 2000s many feared spam would end email as a useful tool: inboxes were becoming unusable. The turning point came late in that decade, when spam filtering improved enough to break the spammer’s economics: sending mass email stopped being profitable if nobody saw it.
That shift brought an important side effect: being labeled as spam started carrying a very high cost. Today a company that lands on email blacklists can lose its domain and its reputation within days. Cantrill argues something similar could happen to undisclosed AI-assisted writing: as detectors like Pangram 4 become reliable, being identified as the author of AI-generated text starts to carry the same reputational cost as being marked as spam.
The difference is that the filter isn’t applied by a mail server: it’s applied by the reader themselves, in real time, deciding whether to keep reading or close the tab.
How Pangram 4 and other AI text detectors work
Pangram Labs doesn’t publish the exact details of its classifier, but the family of techniques behind this kind of tool is well known. A typical detector combines three signals:
- Perplexity: how predictable each next word is according to a reference language model. LLM-generated text tends to systematically pick the most probable word, producing a lower and more stable perplexity than human prose.
- Burstiness: the variation in sentence length and complexity. A human alternates long and short sentences irregularly; an LLM tends toward a more uniform rhythm.
- Stylistic fingerprint: repetition of connectors and filler phrases typical of an LLM (“it’s worth noting,” “it’s important to point out,” “in summary”), highly symmetrical paragraph structure, and self-explanatory transitions.
Using these signals, the classifier is trained on millions of examples labeled as human or AI-generated, and learns to separate both distributions. The result is a probability score, not a binary certainty: that’s why Cantrill insists on looking at the model’s false-positive and false-negative rates, not just its existence.
The following diagram summarizes the typical flow of this type of classifier:
flowchart TD
A["Unsigned text"] --> B["Signal extraction: perplexity, burstiness, style"]
B --> C["Classifier trained on human and LLM text"]
C --> D{"Probability score"}
D -->|"high"| E["Flagged as AI-generated"]
D -->|"low"| F["Flagged as human"]None of these signals is foolproof on its own. What makes a classifier like Pangram 4 strong is combining them and training them against a huge, up-to-date database of real examples, something a homemade heuristic can’t match.
How to test it: DIY heuristics before publishing
You don’t need to pay for Pangram 4 to run a first check on a draft. A simple burstiness heuristic, calculated from the variance in sentence length, already works as an early warning. It doesn’t replace a trained classifier, but it helps catch suspiciously uniform paragraphs before publishing.
This first example calculates the variance of a single paragraph:
def burstiness(texto):
oraciones = [s.strip() for s in texto.split(".") if s.strip()]
largos = [len(s.split()) for s in oraciones]
promedio = sum(largos) / len(largos)
varianza = sum((l - promedio) ** 2 for l in largos) / len(largos)
return varianza
texto_ejemplo = "Human text tends to vary a lot. Sometimes a sentence is long and strings together several linked ideas. Sometimes not."
print(burstiness(texto_ejemplo))
Low variance (sentences of very similar length) is a warning sign; high variance suggests a more human rhythm. The next example applies the same idea to a whole file, paragraph by paragraph, useful for reviewing a long draft before publishing it:
import statistics
import re
def analizar_parrafos(ruta_archivo):
with open(ruta_archivo, encoding="utf-8") as f:
contenido = f.read()
parrafos = [p for p in contenido.split("\n\n") if p.strip()]
for i, parrafo in enumerate(parrafos, start=1):
oraciones = re.split(r"(?<=[.!?])\s+", parrafo.strip())
largos = [len(o.split()) for o in oraciones if o]
if len(largos) < 2:
continue
varianza = statistics.pvariance(largos)
alerta = "suspicious" if varianza < 4 else "ok"
print(f"Paragraph {i}: variance={varianza:.1f} -> {alerta}")
analizar_parrafos("borrador.md")
To confirm the signal works, run the script on a text you know you wrote yourself and on a long, unedited chatbot response: if the second one’s variance is consistently lower, the heuristic is capturing the signal you’re looking for. It’s the same check any team runs before trusting a model: test it against cases with known origin.
⚠️ Heads up: no detector is foolproof. Pangram Labs acknowledges a higher-than-desired false-negative rate: a text can pass the filter without being entirely human. Don’t use a single detector, whether your own or a commercial one, as definitive proof against an author.
Impact and analysis
The most uncomfortable data point in Cantrill’s essay isn’t about the detectors, but about the readers. 98% of surveyed developers said they preferred their own text, errors included, over one polished by AI. That contradicts the most common argument for using an LLM to write: that it improves quality. For the reader, authenticity outweighs polish.
The following table compares the approaches currently available for detecting or preventing undisclosed AI-generated text:
| Approach | How it works | Advantage | Limitation |
|---|---|---|---|
| Proprietary classifier (Pangram 4) | Model trained on millions of human and LLM examples | Very low false-positive rate, according to Cantrill | Code and training data aren’t public |
| Burstiness heuristic (DIY) | Measures variance in sentence length | Free, runs locally, no external dependencies | Easy to dodge by varying rhythm by hand |
| Generator watermarking | The model marks its own output as it generates it | Near-binary detection if the generator cooperates | Useless if the author edits the text or uses an unmarked model |
| Expert human review | A frequent reader recognizes stylistic tics | Doesn’t depend on any external tool | Doesn’t scale beyond a few texts per day |
The core trade-off is honest, and Cantrill acknowledges it: detectors like Pangram 4 minimize false positives at the cost of letting some AI-generated text slip through (false negatives). For an editor, this means a low score isn’t a guarantee of human authorship; it’s only insufficient evidence to accuse.
💭 Key point: the parallel with spam isn’t just rhetorical. When the cost of being detected outweighs the benefit of saving time on writing, behavior changes. That’s what happened with email spam in the late 2000s, and it’s Cantrill’s bet for AI-generated text.
What’s next
If Cantrill’s analogy holds, the logical next step is for publishing platforms and technical communities to start running detectors like Pangram 4 routinely before accepting an article, just as any mail server today runs a spam filter before delivering a message to the inbox.
That doesn’t solve the underlying problem: an author can still use AI to research, organize ideas, or check grammar without the final text reading as generated. Cantrill’s point isn’t to ban AI use in the writing process, but for the final result to reflect the author’s own work, even if imperfect.
For communities like Hacker News, where according to the cited data 78% of readers abandon a piece as soon as they detect AI, the pressure is already real: publishing with obvious LLM assistance stops being a neutral decision and starts carrying a measurable reputational cost.
📖 Summary on Telegram: View summary
Try it yourself: run this article’s burstiness function on your latest draft and compare the result against a paragraph you know is 100% yours.
Frequently Asked Questions
What is Pangram 4?
It’s the latest version of Pangram Labs’ AI-generated text detector, released just over a month before Cantrill’s essay on September 5, 2026, with an accuracy Cantrill describes as a notable leap over Pangram 3.
Is Pangram Labs free?
Available research doesn’t detail public pricing; Pangram Labs offers its detector as a service aimed at publishers and platforms, so it’s best to check its official site directly for current plans.
Why do 78% of developers abandon AI-written text?
According to the survey cited by Cantrill, readers don’t reject imperfection, they reject the lack of authenticity: they prefer their own text with errors over one artificially polished.
Can an AI text detector be fooled?
Yes, manually editing sentence rhythm or rewriting passages reduces the signal these classifiers look for, although Cantrill reports that Pangram 4 keeps a low false-negative rate even so.
What is Oxide Computer’s RFD 576?
It’s an internal design document at Oxide Computer, the company where Cantrill is co-founder and CTO, where he argues that using AI to write voids the social contract between author and reader.
What’s the connection between this and email spam?
Cantrill compares the maturation of AI text detectors to the improvement of spam filtering in the late 2000s: in both cases, once detection becomes reliable, the reputational cost of being flagged outweighs the benefit of the practice.
References
- The Revolt of the Reader, by Bryan Cantrill: the original essay on readers’ revolt against AI-generated text.
- Pangram Labs: the company behind Pangram 3 and Pangram 4, the AI-generated text detectors mentioned in the essay.
- Email spam, on Wikipedia: historical context on the war on spam that Cantrill uses as an analogy.
- Oxide Computer Company: the company where Cantrill is co-founder and CTO, source of the internal RFD 576 document cited in the essay.
📱 Enjoy this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.
Imagen destacada: Foto de Igor Omilaev en Unsplash
0 Comments