The dash and the balance: anatomy of the AI writing signature

By Paul-Antoine TUAL, AI Transformation Leader, Croissance et Transitions. August 2026.

Where this comes from. This article is written for the SME or mid-market executive whose proposals, reports and publications now pass through a language model, and who wonders what a reader recognises the machine by. It documents two measurable markers, the em dash and the see-saw sentence, and traces their mechanics: corpus composition, tokenizer vocabulary, preference-based alignment. Three sets of measurements are used: the 2026 published work on twelve models, a direct inspection of OpenAI’s BPE vocabularies carried out for this article, and this blog’s own corpus before its own correction pass. Sources and measurements taken on 28 August 2026. The character under study appears here only in code, never in the prose: that is this site’s rule, and the article explains why.

---

1. Seven hundred and sixty-nine dashes, on this blog

Since 23 July 2026 this site has applied one simple rule: no em dash in the body of a published article. The repository archive, frozen four days earlier, makes it possible to measure exactly what that rule corrected, and the result deserves to open the article rather than close it.

Across the twenty-four French articles the blog held at the time, prose zone isolated from the source block and the front matter, we count 61,496 words and 769 em dashes, or 12.50 per 1,000 words. The per-article median is 11.55. The maximum reaches 25.15 in an article about dropping Git, one dash every forty words. The minimum falls to 0.92. One hundred and seventy-six sentences carried at least two, out of five hundred and sixty-two sentences containing a dash: one dash-bearing sentence in three.

Those texts were drafted with model assistance, reviewed, corrected, sourced. Nobody in the loop had decided to write that way. The character installed itself.

One caution before going further. The same study that supplies the model column also measures eight published human essays, 57,232 words: a mean of 3.23 per 1,000 words, range 0.33 to 17.12 [1]. A fifty-fold spread between two human authors forbids reading any isolated figure as proof of origin. Our 12.50 sits inside the human range. It sits at the top of it, and it got there by accident.

2. The genealogy: Markdown leaking into prose

The most solid explanation published to date fits in one sentence, and it comes from a March 2026 paper by E. M. Freeburg titled The Last Fingerprint [1]. The em dash is Markdown leaking into prose: the smallest surviving unit of the structural orientation a model acquires from Markdown-saturated training corpora.

Corpus composition makes the thesis plausible before the experiment even runs: more than 60% of the training data for the early GPT-family models came directly from web crawls, where markup is everywhere [13].

The protocol is worth more than the formula. Twelve instruction-tuned models, five providers, roughly 240,000 generated words, two conditions. In the first, the model writes freely. In the second, all formatting is forbidden: no headings, no bullets, no bold, no numbered lists. Result on the visible markers: they vanish across all twelve models, bar a small residual in Claude Haiku 3.5. The instruction works perfectly.

The dash stays. And it stays very unevenly: GPT-4.1 moves from 10.62 to 9.10, a 14% drop; DeepSeek V3 from 6.95 to 5.41; Claude Opus 4.6 from 9.09 to 0.19, a 98% drop; Gemini 2.5 Pro falls to zero. Llama models produce none in either condition across roughly 40,000 words, and the authors verified that they substitute neither double hyphens nor en dashes: the construction itself is absent from their writing.

The reason for that survival lies in a peculiar status. A heading, a bullet, a bolded word are markup and nothing else. The em dash belongs to both registers at once: structural marker and perfectly legitimate prose punctuation. The order “write prose, not Markdown” fails to catch it, because it is already prose.

That leaves the question of who sets the amplitude. The base-versus-instruct comparison answers it: on Llama 3.1 8B, the base model produces 0.49 dashes per 1,000 words, four occurrences across 8,186 words, and the aligned version falls to zero [1]. The latent tendency therefore exists before alignment, and alignment then decides whether to amplify or extinguish it. At Meta it extinguishes. At OpenAI, on long-form text, it amplifies up to 14.03. The data install the disposition; post-training sets the volume.

3. What the model’s vocabulary actually contains

The economic argument can be checked directly, and nobody needs to take it on faith: OpenAI’s tokenizer vocabulary files are public. We downloaded and inspected them for this article on 28 August 2026 [11].

In cl100k_base, the GPT-4 generation vocabulary, out of 100,256 tokens, 44 contain the em dash. The sequence space-plus-dash carries rank 2001, and the bare dash rank 2345. A low rank means an early learned merge, hence a very frequent sequence in the training corpus. For comparison, rank 2001 places the character in the first two percent of the vocabulary, next to constructions as ordinary as a space followed by a semicolon (rank 2652). In o200k_base, the next-generation vocabulary, the dash keeps ranks 2322 and 2733 out of 199,998 tokens.

The decisive detail sits further down the list. The vocabulary holds not only the dash: it holds —and (rank 17,223), —the, —but, —that, —it, —which, —not, —including. The character is merged with the conjunction that follows, into a single token. A model wanting to hang a qualification onto a claim therefore has one token carrying both the break and the connector. The competing sequence , but is not in the vocabulary at all; it costs two tokens.

Two further observations complete the picture, and they matter for French text. First, the en dash is far worse served, with 7 tokens in cl100k_base against 44, which gives the em dash a mechanical edge whenever the model hesitates between them. Second, none of those merges are French. Neither —et, nor —le, nor —ce exist in the vocabulary. The dash’s token economy is an English economy, imported unchanged when the model writes French, where it no longer pays anything.

A token that thinks in English punctuates in French.

4. The balance: what alignment rewards

The other marker is syntactic. Generated text see-saws: a thesis, a reservation, a counter-argument, a synthesis. It concedes before being attacked, it qualifies before being asked. Three published results explain why that posture is the expected equilibrium, without any lab having needed to program it.

The first is an impossibility result, published at ICML 2024 by Chakraborty and co-authors [4]. Standard alignment compresses annotator judgements into a single scalar reward learned under the Bradley-Terry model. The authors establish that such a reward cannot represent diverse human preferences, and that a lower bound on alignment error follows from that diversity. When annotators disagree, the optimisation does not decide: it looks for the barycentre.

The second measures the price paid. Kirk and co-authors, at ICLR 2024, compare supervised fine-tuning against reinforcement learning from human feedback on two base models and two tasks [5]. Reinforcement learning generalises better out of distribution, and it sharply reduces output diversity, within a single prompt as well as across prompts. The authors describe a generalisation-versus-diversity trade-off. Writing that converges on a single posture is the direct consequence.

The third shows that preference data sculpt the posture, not only the content. Sharma and co-authors, also at ICLR 2024, analyse public human comparison datasets and find that matching the user’s stated view ranks among the most predictive features of the preference judgement [6]. What the annotator rewards without thinking, the model learns diligently.

Stack the three. A reward that cannot represent disagreement, an optimisation that crushes variance, human judgements that prefer accommodation: the see-saw sentence falls out of that stack like fruit off a tree. Conceding first covers several evaluation angles at once, which maximises an average reward when you do not know who will read.

A point of honesty here. That reasoning rests on three solid papers, but none of them measures the density of contrastive constructions the way others measure the dash. It remains a well-supported inference rather than a fact established by counting. As an indication, our own pre-correction corpus yields 26 constructions of the “it is not X, it is Y” type across 61,496 words, or 0.42 per 1,000 words: ten times rarer than the dash, and far harder to spot while reading.

5. In French, the rule is not the one people assume

One idea circulates: the model supposedly imports an Anglo-American habit into French, dropping the em dash where French would want the en dash. Verification reverses the proposition.

Gérald Purnelle’s French typography syllabus at the University of Liège is explicit: the long dash separates clauses, and the half dash serves enumeration items as well as numeric ranges, “pp. 52-64” [7]. The reference Wikipedia article on the dash agrees and adds the nuance that matters: these rules “are not unanimous” among grammarians, and several French publishing houses, including Gallimard, Albin Michel and Le Seuil, routinely use the en dash where the Imprimerie nationale uses the em dash [8]. Choosing the character for a parenthetical insertion is therefore not the error people take it for.

Two French rules, on the other hand, are sharp and discriminating. Ordinary breaking spaces are required on both sides of a parenthetical dash, whereas the most common English usage sets the character closed up against the surrounding words [7]. And a dash cannot precede a full stop: when the insertion ends with the sentence, the period absorbs the closing dash [7].

So we tested our own corpus against those two rules, expecting to find the announced anglicism. Out of the 769 dashes measured, 760 are surrounded by two spaces, or 98.8%. And the count of dashes placed immediately before a full stop, the most characteristic sign of an English calque, comes to zero.

The verdict is more interesting than the one we went looking for. The machine did not pick the wrong typographic convention: it observed French punctuation with a regularity few human newsrooms sustain. Its signature is not readable in the shape of the character, but in its frequency and its stacking, those 176 sentences carrying two dashes or more that betray badly separated clauses. The tic is prosodic, not orthotypographic.

Fixing the typography would have changed nothing. Fixing the dose changed everything.

6. When the signature becomes an accusation

Once a marker is known, it becomes evidence. And badly calibrated evidence does damage.

The story is measured at population scale. A study published in June 2026 by Przemysław Czuma covers 69,632 medRxiv preprints whose Discussion section exceeds 500 characters [2]. The share of texts containing at least one em dash rises from 4.23% before 30 November 2022 to 11.58% after, an absolute increase of 7.35 points, 95% confidence interval 6.94 to 7.77, with an odds ratio of 2.96. The rise is gradual: about 4% in 2023, 8.0% in 2024, 20.3% in 2025. A placebo test inside the earlier period returns 0.13 points, which rules out a background drift.

The author states the limit of his own work: a population-level indicator, not a per-paper detector. Practice ignored that caution. An ActuaLitté investigation from December 2025 lists the consequences: OpenAI acknowledged that its own detector classified human texts as generated, including works by Shakespeare and the Declaration of Independence; the Australian Catholic University suspended Turnitin’s detector after unfounded accusations based on the automated score alone; a Paris secondary-school student was wrongly accused during a philosophy exam, the accusation later withdrawn [9]. Turnitin itself insists that a score must not be the sole basis for an adverse decision.

The feedback effect is already visible in human writing. The Blog du Modérateur summed it up in April 2026: authors who had always used the dash now avoid it to escape suspicion, while the character has six centuries of use and remains, per the Chicago Manual of Style, the most versatile mark in English punctuation [10]. A legitimate character is withdrawing from written French and English because a machine liked it too much.

Providers, for their part, reacted. OpenAI addressed the point in GPT-5.1 in November 2025, and Sam Altman announced it on X on 14 November in terms that carry the admission as much as the fix: “Small-but-happy win: If you tell ChatGPT not to use em-dashes in your custom instructions, it finally does what it’s supposed to do!” [3]. The word that matters is “finally”.

7. What an executive can do with this on Monday morning

Three practical lessons, for anyone publishing under their own name text that passed through a model.

The first concerns the instruction. Asking for “no Markdown” settles nothing: Figure 3 shows the generic instruction lets the character through, because it is already prose. Naming the dash works better, without universal guarantee. The only rule that holds is verification: on this site, an article is not published until a script has counted zero occurrences in the prose zone. Two seconds of checking beat a pious instruction. It is the same reflex as the one described in our article on mental load delegated to AI: the machine emits, the human approves in writing.

The second concerns model choice. Density and compliance vary across providers by a factor of fifty, and the ranking is not the one people assume. For editorial use, a model that produces many dashes but obeys when told beats a soberer model that is deaf to instruction. That variable can be tested in ten minutes on a real text, before committing to a subscription, and it usefully adds to the selection criteria we set out in the end of the race for the model.

The third concerns caution. A density figure proves nothing about a given text: the human range of 0.33 to 17.12 rules it out, and the false positives documented in education proved it to students and to one secondary-school pupil. An executive suspecting that a CV, a dissertation or a supplier proposal came out of a machine holds a population-level clue, not an item of evidence. Grounding a legal or disciplinary decision on it would be a mistake.

What remains is what this story says about the craft of writing. We removed the dash from this site so as not to look like a machine, and we gained something else: every removal forced a choice about which of the two stacked clauses was the main one. The text got shorter and made up its mind. A cosmetic correction turned out to be a correction of thought. The model had handed us a convenient way of not deciding; taking it away obliged us to decide.

The character will go. The habit it was covering asks for a little more work.

Caveat. The per-model densities come from a single research paper published in March 2026, with a documented protocol but no independent replication known to date. The measurements presented as our own (this blog’s corpus, the BPE vocabularies) were redone on 28 August 2026 and are reproducible with the public files cited. Two claims circulating on the subject were set aside for lack of a verifiable primary source: a 30% rise in dash usage in nineteenth-century books attributed to a study we could not locate, and the idea that a model prevented from writing the dash would systematically substitute a double hyphen, contradicted for Llama models by the authors’ own verification.

Sources

[1] E. M. Freeburg, The Last Fingerprint: How Markdown Training Shapes LLM Prose, arXiv:2603.27006, 27 March 2026. Twelve models, five providers, roughly 240,000 generated words; human baseline of 57,232 words across eight published essays. [2] P. Czuma, Em-ergence of the em-dash: a population-level rise in em-dash frequency in medRxiv preprints at the dawn of the large-language-model era, arXiv:2606.29540, 28 June 2026. 69,632 preprints. [3] S. Altman, post on X, 14 November 2025: “Small-but-happy win: If you tell ChatGPT not to use em-dashes in your custom instructions, it finally does what it’s supposed to do!” [4] S. Chakraborty, J. Qiu, H. Yuan, A. Koppel, F. Huang, D. Manocha, A. S. Bedi, M. Wang, MaxMin-RLHF: Alignment with Diverse Human Preferences, arXiv:2402.08925, ICML 2024. Impossibility result for a single scalar reward. [5] R. Kirk, I. Mediratta, C. Nalmpantis, J. Luketina, E. Hambro, E. Grefenstette, R. Raileanu, Understanding the Effects of RLHF on LLM Generalisation and Diversity, arXiv:2310.06452, ICLR 2024. [6] M. Sharma et al., Towards Understanding Sycophancy in Language Models, arXiv:2310.13548, ICLR 2024. Analysis of human comparison datasets. [7] G. Purnelle, Règles et usages de typographie française, syllabus, University of Liège, institutional repository ORBi 2268/298446. Section 2.7 “Les tirets” and preliminary notes on the cadratin. [8] Wikipedia (French), article “Tiret”, accessed 28 August 2026. Imprimerie nationale usage, French publishing-house practice, absence of unanimity among grammarians. [9] E. Berton, “IA et faux positifs : pourquoi certains détecteurs se fourvoient”, ActuaLitté, 30 December 2025. [10] T. Coëffé, “Le tiret cadratin est mort, vive le tiret cadratin”, Blog du Modérateur, 16 April 2026. [11] OpenAI public BPE vocabulary files, cl100k_base.tiktoken and o200k_base.tiktoken, hosted on openaipublic.blob.core.windows.net, downloaded and inspected on 28 August 2026 for this article. [12] Own measurement on this site’s repository archive frozen at 19 July 2026: 24 French articles, 61,496 words of prose excluding front matter and the source block. [13] L. Erisson, “Why Did LLMs Steal Our Em-Dashes?”, Office for Science and Society, McGill University, 8 May 2026.