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

The Olang’ Trap

A note on attribution before this entry begins

The argument begins with Marcus Olang’, not the Dictionary. Olang’ published I’m Kenyan. I Don’t Write Like ChatGPT. ChatGPT Writes Like Me. on 8 July 2025.¹ The Dictionary’s contribution is to name the recurring failure mode and connect it to the AI Writing cluster.

The reader should read the original. Olang’s voice—quietly furious, very funny, and written in the formal register he is defending—does not survive complete in a summary. The Dictionary also preserves the apostrophe because that is how he presents his name.

In one sentence

The Olang’ Trap is the failure mode in which formal human prose—especially prose shaped by post-colonial English education—is mistaken for machine writing because readers or detectors treat one narrow contemporary register as the measure of authentic human voice.

The argument

Olang’ begins with an injury from his working life. After spending days on a proposal, he was asked to rewrite it with “a more human touch” because it sounded like ChatGPT.

He traces his register to the Kenya Certificate of Primary Education and its composition paper. Students were trained to begin strongly, demonstrate vocabulary, and build an essay with an explicit structure. These were not machine habits. They were the marks by which an education system, shaped by British rule and its aftermath, recognized command of English.

Then comes Olang’s punchline: language models learned to sound authoritative from enormous collections of books, academic papers, reference works, legal writing, and other structured prose. In his formulation, the machine “ended up sounding like a KCPE graduate who scored an ‘A’ in English Composition.” The world then looked at the human result of that education and called it artificial.

This is a historical and rhetorical argument, not a measured reconstruction of a particular model’s training corpus. Its force is that the traits now treated as suspicious—formality, structure, grammatical precision, and an authoritative cadence—were taught as marks of educated English long before ChatGPT. Stylistic suspicion collapses that human history into a machine label.

What the evidence supports

The trap does not mean every suspicion of AI assistance is wrong. It means that prose style alone cannot establish authorship, and that errors can be distributed unequally.

A 2023 study tested seven then-current detectors on essays by native and non-native English writers. The detectors frequently classified the latter as AI-generated.² That is direct evidence of bias in a class of detectors and a particular test set. It is not proof that every current product fails in the same way or at the same rate.

There is also a related labour history. Reporting has documented poorly paid data-labelling and content-moderation work in Kenya and other countries. That history matters to the political economy of AI. It does not establish that the same workers supplied the prose or evaluative preferences responsible for a model’s formal register. It belongs beside Olang’s argument, not inside its causal proof.

What this means for operators

For teachers. Never make an authorship or misconduct decision from a detector score alone. Use the student’s prior work, the course process, drafts where available, and a conversation about the text. A formal, structured essay from a student whose first language is not American English is not evidence of machine authorship.

For hiring managers and editors. “Too formal” and “too structured” are aesthetic judgments, not provenance tests. Asking for a “more human” rewrite may be asking a writer to abandon a register the reader does not share.

For writers caught in the trap. The machine’s imitation does not make the human register less human. Naming the trap will not remove its cost, but it gives the writer and the institution a more accurate account of what may be happening.

The detection economy

AI-detection vendors sell confidence in a classification that remains difficult. Their incentives reward a legible score; the consequences of a false accusation are borne elsewhere. The appropriate response is not to invert the score or declare detection useless. It is to refuse the false precision of treating a probabilistic, version-dependent signal as an authorship verdict.

The Olang’ Trap is therefore a case for procedural humility. If the operator cannot see what a tool filters out, the operator must lower the weight placed on what it reports.

The Olang' Trap

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