The Olang’ Trap

Essay This entry carries an argument or interpretive position, not just a neutral definition.

A note on attribution before this entry begins

The argument in this entry is Marcus Olang’s, not the Dictionary’s. The Dictionary’s contribution is naming the argument as a structural failure mode of the AI-detection economy and cross-linking it to the AI Writing cluster. Olang’ published I’m Kenyan. I Don’t Write Like ChatGPT. ChatGPT Writes Like Me. on his Substack on 8 July 2025¹. The piece went viral, was discussed in Hacker News in December 2025, and was cited by Jason Koebler in Your AI Use Is Breaking My Brain (404 Media, 11 May 2026) — the proximate source through which the argument arrived in this Dictionary’s morning brief.

The reader is encouraged to read the original piece in full. The summary that follows is faithful but compressed, and Olang’s voice — quietly furious, very funny, full of the rhetorical structures he is defending — does not survive the compression.

A note on the spelling: the apostrophe in Olang’ is the conventional orthographic mark in Dholuo for the glottal stop. The Dictionary preserves it throughout, in deference to the writer’s name as he writes it. The Olang’ Trap, accordingly, is named with the apostrophe.

The trap, in one sentence

The Olang’ Trap is the structural failure mode of the AI-detection reflex, in which the human writers most reliably misclassified as AI are those whose native English register most closely matches the formal-empire register that AI training corpora were built on — a register that AI then learned to produce, partially through labour from the same populations whose writing it now displaces from human-author status.

The trap is recursive. It is also unjust in a specific and namable way. The cost of the AI-detection reflex falls hardest on the populations whose writing taught the AI to sound the way it does. Olang’ names this with the bitterness it deserves.

Olang’s argument, in his own structure

The argument has four moves, each of which the original piece makes more vividly than this summary can.

Move one: the experience. A real moment from Olang’s working life:

“I received a reply to a proposal I had laboured over for days. ‘This is a really solid base, but could you do a rewrite with a more human touch? It sounds a little like it was written by ChatGPT.’”

The piece’s emotional spine is the experience of being told, in a professional context, that one’s own native register sounds inauthentic, and that the request is to rewrite it to sound more human — to deliberately produce text in a register that the requester considers more genuinely human, which is to say, more American, more informal, more peppered with the conversational errors that have become the new fingerprint of authenticity. The trap is not a technical curiosity. It costs Olang’ work.

Move two: the inheritance. Olang’ traces his own register to a specific institution: the Kenya Certificate of Primary Education, the KCPE, and its English Composition paper. The KCPE composition was a forty-minute, high-stakes test on which much of a Kenyan student’s life trajectory could pivot. Three commandments structured the preparation:

“The first commandment? Thou shalt begin with a proverb or a powerful opening statement… The second? Thou shalt demonstrate a wide vocabulary. You didn’t just ‘walk’; you ‘strode purposefully’, ‘trudged wearily’, or ‘ambled nonchalantly’… The third, and perhaps most important commandment, was that of structure. An essay had to be a perfect edifice. The introduction was the foundation, the body was the walls, and the conclusion was the roof.”

These commandments were not arbitrary. They were the residue of a colonial education system that had, over a century, taught English as a tool of power. The KCPE preserved that system into the post-independence era, where Olang’ notes the language “became the official language, the language of opportunity, the new marker of class and sophistication.” Mastery of formal English was, in Olang’s framing, “a signal. It was proof that you were educated, that you were civilised, that you were ready to take your place in the order of things.”

Move three: the punchline. This is the load-bearing move of the piece, and the line deserves its full weight on the page:

“And right there is the punchline to this long, historical joke. An ‘AI’, a large language model, is trained on a vast corpus of text that is overwhelmingly formal. It learns from books published over the last two centuries. It learns from academic papers, from encyclopaedias, from legal documents, from the entire archive of structured human knowledge. It learns to associate intelligence and authority with grammatical precision and logical structure.”

“The machine, in its quest to sound authoritative, ended up sounding like a KCPE graduate who scored an ‘A’ in English Composition. It accidentally replicated the linguistic ghost of the British Empire.”

The construction is itself a piece of formal English in the register Olang’ is defending, and the irony is deliberate. The AI sounds like Olang’ because the AI was trained on the corpus that taught Olang’ how to write, which is the corpus that the British Empire left behind across its former territories. The detection reflex is therefore not detecting machine writing; it is detecting the register of formal English that was both colonially imposed and machine-replicated. These are not the same thing. The reflex collapses them.

Move four: the inversion. Olang’ closes by inverting the AI-detection frame entirely. The world is now looking “through its new and profoundly flawed technological lens, [at] the result of our very human, very analogue training and call[ing] it artificial.” The detection is broken not because it has false positives in some statistical sense, but because the category itself — human-written-versus-AI-written — is not separable in the way the detection reflex assumes. The AI register and the colonial-formal-English register are the same register. A reflex that calls one of them inauthentic is, structurally, also calling the other one inauthentic. Most users of the reflex do not realise this is what they are doing.

Why this is more than a footnote in the AI Writing cluster

The Dictionary’s parent entry on AI Writing names the Imperial College Number (35% of new websites in 2026 are AI-generated) as the empirical anchor for the something has changed intuition. The Olang’ Trap is the structural complication that anchor needs. Yes, more AI-generated text is in circulation. And the AI-detection reflex that arose in response is a vector for harm that distributes unequally and predictably. The harm falls hardest on:

Writers who learned English through formal post-colonial education systems. Kenyans, as in Olang’s case, but also Nigerians, South Africans, Indians, Pakistanis, Filipinos, Singaporeans, and many others trained in registers descended from British or American formal-educational English.
Writers whose native register is academic. Junior academics whose prose has been shaped by graduate training, and whose careful structured sentences now read as AI-shaped to readers more accustomed to informal social-media cadence.
Writers who are not native English speakers but who learned formal English well. The signal that used to indicate this person worked hard and learned the register now reads as this person used a machine. The reversal is structurally exact, and the cost falls on the population that did the harder work.

It is worth being precise about what the Dictionary is not claiming. The Olang’ Trap does not mean the AI-detection reflex is always wrong. It is sometimes right, sometimes useful, and there are real cases of AI-generated commodity prose that the reflex correctly identifies. The Olang’ Trap is the failure mode of the reflex, not its inversion. What the trap names is that the reflex has a predictable bias, and the bias falls on the human writers who taught the AI to sound the way it does.

There is a further, harder layer to this argument that Olang’ himself does not develop in detail but that is worth naming here. Much of the labour of training, testing, and moderating the major language models in the 2020s has been performed by low-paid workers in countries including Kenya, Nigeria, the Philippines, and India — workers who in many cases passed through the same formal English education systems Olang’ describes. The AI register was therefore not only trained on the corpus that taught these workers how to write; it was judged against their evaluations of what good writing should look like. The labour was extracted twice: once when their writing entered the training corpus, and again when their judgement was used to shape the model’s output. The AI-detection reflex now treats the output of that labour as a marker of inauthenticity. The people whose work taught the machine how to sound human are now told they sound like machines. This is the Olang’ Trap at its full structural depth.

What this means for operators

The Dictionary’s editorial position, in AI Writing, is that the AI-detection reflex should be shifted: instead of asking is this AI?, ask would I want to be reading this if I knew? The Olang’ Trap makes this shift more urgent, not less. Three concrete operational consequences:

For teachers. A formal, structured, error-free essay from a student whose first language is not American English may not be AI-generated. The AI-detection software in widespread use in 2026 — including the tools sold by Pangram Labs and competitors — has documented bias against non-native and post-colonial English. The Dictionary’s recommendation: never act on an AI-detection score alone. Always combine with classroom observation, prior writing samples, and a conversation with the student. The cost of a false accusation falls heavily on the population the reflex is structurally biased against.

For hiring managers and editors. A proposal, application, or piece of writing that reads as too formal or too structured is not, by that fact alone, AI-generated. It may be the work of a writer trained in a register the reader does not share. Asking for a rewrite with a more human touch — Olang’s own painful example — is in some cases a request that the writer abandon their native register in favour of one the reader recognises. This is a form of cultural pressure that did not exist in 2022 and that, in 2026, is being applied without most operators noticing.

For writers in the affected populations. The most uncomfortable consequence: the trap is real, the cost is yours to bear, and the Dictionary cannot make it go away. What the Dictionary can do is name the trap clearly, so that a writer who encounters it can recognise what is happening and articulate it. The register you have been trained into is not less human because a machine has learned to imitate it. The machine learned from you. You did not learn from the machine. This is a load-bearing sentence for an operator caught in the trap.

A closing observation about detection economies

A small industry has grown up around AI detection. Pangram Labs, Originality.ai, GPTZero, Turnitin’s AI-detection add-ons, and dozens of competitors charge money to perform a classification that, as the Olang’ Trap demonstrates, has a structural bias. The detection economy is not neutral. It has commercial incentives to appear accurate even when its accuracy is unequally distributed, and it has commercial incentives to not advertise the populations its tools systematically misclassify.

The Dictionary does not recommend the use of these tools without an honest reckoning with the Olang’ Trap. The reckoning is hard, because the tools’ false-positive rates against post-colonial English are not always published; researchers have to construct test sets themselves to find them. But the operator who is honest about what they do not know is better positioned than the operator who treats the score as authoritative. The Olang’ Trap is, in part, an argument for that honesty.

The Olang' Trap

A working reference for the language of agentic AI. By Matthew D. Langenkamp / 雷邁德.