State of the tells: what's changing, monthly.

July 2026

Rising: em-dash chains remain the loudest tell across GPT-5.x and Claude-family output — three per paragraph is common in default drafts. 'It's less about X, more about Y' constructions are spreading fast. Rule-of-three sentences remain the most durable structural tell across all vendors.

Falling: 'delve' has visibly declined in newer model output — the meme did its work — but survives in bulk-generated SEO content, where it now dates a text precisely. 'Tapestry' and 'symphony' metaphors are increasingly rare.

Blacklist changes: initial release of the blacklist with 90+ entries in six categories; em-dash density check added to the in-browser checker.

Humanizing note: the 'critic pass' pattern — one model rewrites, a second reviews the changes — continues to outperform single-pass rewrites in our before/after work; it's the workflow behind the gallery examples and the reason multi-model setups like MultipleChat's Humanize mode hold their edge over one-click tools.

Next update: August 2026. Observations are editorial, based on recurring patterns in current model output — not a scientific corpus study.

June 2026 (retro entry)

Rising then: "it's less about X, more about Y" was just breaking out of tech-adjacent content into general business writing; em-dash density in GPT-5.x drafts measurably increased after the spring model refresh. "Notably" began replacing "importantly" as the hedge of choice in Claude-family output.

Falling then: "tapestry", "symphony" and "testament to" metaphors continued their decline; "furthermore" kept losing ground to "additionally" (a lateral move, not an improvement).

Ecosystem: several major detectors shipped quiet scoring updates in June — a reminder that any "passes detection" claim has a shelf life measured in weeks, which is why this site doesn't make them.

How we observe (and what we don't claim)

Method, honestly stated: we generate drafts across the current model generation weekly for the gallery and guide work, log recurring phrases and structures, cross-check against community reports, and promote a pattern to the blacklist only when it recurs across multiple models or dominates one. This is editorial observation with consistent habits — not a controlled corpus study, and we label frequencies ("common", "declining") rather than inventing percentages. Where a real measurement exists we cite it; where it doesn't, we say what we saw. If you spot a tell we're missing, the contact route is in the footer — credited corrections make the list better for everyone.

Tell of the month: the em-dash chain

A closer look at the current #1. The em-dash is legitimate punctuation — human writers have used it for emphasis forever — but current models deploy it as a universal connector: apposition, contrast, elaboration, drama, all get the same dash. The result is a text where commas and parentheses barely exist and every third sentence carries an interruption. Why it happened is instructive: dash-heavy prose is common in the high-engagement essays and newsletters that dominate fine-tuning data, so the models learned emphasis as a default rather than an exception. The fix costs nothing: in any rewrite prompt, "no em-dashes" forces the model to actually choose between a comma, a period and a parenthesis — and the choosing is where natural rhythm comes back. Our checker counts dash density for exactly this reason: one dash per 80+ words is a writer; one per 40 is a template.

What's next on this page

Planned for the coming entries: a tracked list of tells by model family as the autumn releases land, before/after dashboards when we add new gallery genres, and a "retired tells" archive — phrases that stopped being evidence, so editors stop over-reading them. The blacklist, the checker's phrase set and this changelog move together: when an entry changes state here, the tool reflects it the same week. If you want the updates without checking back, the RSS/newsletter option lands with the August entry.