The Punctuation Staircase Is Not an AI Fingerprint

I read a lot of Chinese written by Claude, most of it in a terminal, and for months I kept seeing the same shape. The commas, enumeration commas and colons on consecutive lines seemed to line up into diagonals, each one or two characters to the side of the one above, like a staircase running down the screen. I started treating it as a tell. When I saw the staircase, I assumed a model had written the text.

Chinese makes this kind of shape easy to see. Every Han character and every full-width punctuation mark occupies the same square cell, so a block of Chinese text is laid out on graph paper, and the position of each mark is a coordinate. English words vary in length and their punctuation lands anywhere. The question I wanted answered was who puts the marks on those coordinates: the model, the training, or nobody.

To count staircases I needed a rule. A staircase here is three or more punctuation marks on consecutive rendered rows, each one 1 to 3 characters left or right of the mark above it, all stepping in the same direction. Unless stated otherwise the numbers below use lines 40 characters wide, close to what my terminal shows; the conclusions hold at 30 and 50.

The data

The machine side has three parts. The first is 400 replies sampled at random from my own Claude Code sessions, out of 1,554 that are mostly Chinese and longer than 300 characters. The second is 347 replies from seven chat models (Claude Opus 4.6, Sonnet 4.6 and Sonnet 5.5, GPT-4o, GPT-5.1, Gemini 3.1 Pro and DeepSeek V4 Pro), each asked the same 50 everyday questions in Chinese through the raw API with no system prompt. The third is 400 ChatGPT answers from December 2022, taken from the HC3-Chinese corpus (Guo et al., 2023).

The human side was all written before ChatGPT was released. HC3-Chinese supplies 400 answers from a psychology Q&A site and 400 Baidu Baike encyclopedia entries. The fourth human source is Ruan Yifeng, whose Chinese writing on programming I have read for years: nine issues of his weekly newsletter from 2018 to 2021 and five chapters of his ES6 book.

First guess: the rhythm is too even

My first explanation was that the model writes clauses of nearly equal length, so the marks advance by a similar amount on every row and drift diagonally. That predicts a narrower spread of clause lengths in machine text. It is not there. The coefficient of variation of clause length (standard deviation over mean) is 0.69 in my Claude Code sample and 0.64 across the seven API models, against 0.59 to 0.71 for the three human sources. The model's clauses are as uneven as people's.

A sharper test removes rhythm altogether. If I shuffle the clauses inside every paragraph, so that clause lengths stay the same but their order becomes random, any rhythm the writer had is destroyed. Within paragraphs, the staircase count does not move: 4.3 per 100 rows before shuffling and 4.6 after for Claude Code, 13.8 both times for the psychology answers. Whatever produces the staircase, it is not the order in which clauses are written.

Second guess, and a mistake

Next I counted whole replies instead of paragraphs, and compared my Claude Code sample with the only human writing I had collected at that point, Ruan Yifeng's. Claude came out far ahead: 7.3 staircases per 100 rows that contain punctuation, against 2.8. Adding blank lines between Claude's lines removed most of the gap, so I concluded that the staircase was a layout effect: models stack many short lines with no blank line between them, and the marks on stacked lines can chain.

That comparison was wrong, because one careful writer is not a baseline. Ruan Yifeng leaves a blank line after nearly every paragraph (62% of his lines are blank) and uses punctuation sparingly. When I added the HC3 human answers, the ranking reversed. Answers on the psychology Q&A site, written by ordinary people years before any chatbot, have 15.3 staircases per 100 rows. My Claude Code sample has 5.1 and the seven models through the API have 2.3.

The chance baseline

The reversal suggested the count tracks something duller than authorship. To check, I built a baseline for every document: keep the number of marks on each row and the length of the row, but place the marks at random positions along it, then count staircases again, ten draws per document.

Every corpus lands at or near its own chance baseline. For my Claude Code sample, the seven API models, Baidu Baike and Ruan Yifeng, observed and random differ by 1% to 3%. The two corpora that sit above chance are both paragraph prose: the psychology answers (14% above) and ChatGPT's answers from 2022 (25% above). The modern models' replies, the text I suspected, are the ones that sit exactly at chance. Most of the count is whatever punctuation density and line breaks produce when the marks fall anywhere. Writers who put nearly three marks on every row of a long, unbroken paragraph get many staircases, whoever they are, and the psychology answers carry the most marks per row in the data. The pattern holds at 30 and 50 characters per line, where the ratios range from 0.92 to 1.25.

The width slider in the demo shows the other half of the argument. Staircases appear and disappear as the line width changes, because the line breaks belong to the window, not to the text. The model generates a sequence of characters and never sees where they will wrap, so neither the model nor its training could aim at the shape.

What does differ

The count does not separate human from machine text, but the kind of staircase does, and I think that explains what I was noticing.

In human text the staircases sit inside paragraphs: only 3% to 38% of them cross from one source line into the next. In the seven models' replies, 91% do. They form across list items, and 61% of them sit in the left third of the line, against 35% to 43% for the human sources. They are made by a specific format: 68% of the models' lines are list items, and 41% of lines open with a short label followed by a colon (运动：…, 兴趣培养：…), against 4% to 6% in the human sources. When two adjacent list items both open with a label, the labels have the same length 32% of the time and differ by one to three characters 48% of the time. Equal labels put the colons in a vertical column; labels that differ slightly put them on a diagonal.

Format is also the only thing an intervention removes. Inserting a blank line after every line cuts the models' staircases from 6.1 to 0.4 per 100 punctuated rows, measured against the same text with all blank lines removed, while the psychology answers only drop from 17.9 to 11.9, because their staircases live inside paragraphs that a blank line does not split.

So the answer to "who did it" has three parts, and none of them is aiming at a staircase. Preference training makes models write lists: reward models and human raters favor lists, bold text and similar formatting, and a small share of biased preference data is enough to plant the bias, which best-of-n sampling and online DPO then amplify (Zhang et al., ACL 2025). Chinese has an old habit of opening parallel items with short, similar labels, familiar from official documents and slides, and the models have learned it. The character grid does the rest, and as the chance baseline shows, the geometry is accidental.

Why I saw it in machine text

This part I have not measured. A staircase inside a dense paragraph is one diagonal among many marks, a texture the eye does not separate out. A staircase down the left edge of a stacked list, where short lines leave white space around each mark, stands alone. I also read far more model output than casual human Chinese, and I read it already expecting a model. Any of these could account for my impression, and only a perception study, with the same text shown in list and paragraph layouts, could tell them apart.

What the data does support is narrower. The punctuation staircase is not a fingerprint of Claude, or of language models. Human Chinese produces it at least as often, and in every corpus the count stays within a quarter of what random placement predicts, with the modern models closest to chance. What the machine replies carry is a layout (stacked list items opening with labels) that puts the staircase where I happened to be looking.

Limits

The staircase rule is mine, and I have not checked it against what readers mark by hand. My Claude Code sample comes from my own sessions and is shaped by my instructions file, which pushes toward compact reports; the API sample, one reply per question at default settings, is the cleaner machine baseline. The human corpora differ from the machine ones in genre as well as authorship: encyclopedia entries and Q&A answers are not chat replies. None of this changes the chance result, which compares each corpus only with itself.