Various LLM Smells: How AI-Written Text and Sites Give Themselves Away

A developer who started a mathematics blog late last year decided to use large language models to polish his writing. The result, he admits, read noticeably better than his own prose: richer vocabulary, more interesting sentence structures, and nothing that looked artificial to him at the time. About three months later he began seeing the same sentence structures everywhere — across the entire internet. That observation became a short essay, “Various LLM smells”, which ended up on the front page of Hacker News.

What gives AI-assisted writing away

The essay catalogues the patterns he found in his own drafts. One is an excess of punchlines: standalone lines engineered to land, such as “Humans trust symmetry because it feels like intelligence made visible”, or “Symmetry becomes a trap”. Another is runs of consecutive short sentences — “Then AlphaEvolve arrived. It had no preference for symmetry. No aesthetic prior. No instinct to preserve harmony.” A third is the “X is the Y of Z” construction, his example being “Cringe is the visible signature of moving along a gradient you chose.”

He also flags the negation formula that promises a twist: solutions that “do not merely satisfy the constraint but satisfy the aesthetic instincts”. Each device is unremarkable on its own; repeated at scale, he argues, they become a signature that readers can now recognise from a single paragraph.

The same tells on websites

The second half of the list moves from prose to design. AI-generated sites, he writes, converge on the same font — JetBrains Mono — and on the same layout habit of numbered “step” blocks followed by bullet lists. The same buttons recur, the same card components recur, and so does a small badge with a blinking status dot. None of these choices is bad in itself; the point is that they cluster, and the clusters are visible from a glance at a screenshot.

Why it matters

The author is explicit that this is not an argument against using AI for creative work. His framing is narrower: models leave artefacts, and those artefacts are now recognisable across unrelated tasks and unrelated sites. For editors, designers and readers, the practical value is a checklist — punchline density, short-sentence runs, formulaic antithesis, JetBrains Mono, step cards — that can be applied in seconds, whether or not the author admits to using a model.