AI tell · grade B
Does AI overuse colons? Yes, and it's the same tell as the bold-header list
The colon gets less attention than the em dash, but in one model's output it did more of the work. In our logs, a single model, poolside/laguna-xs, opened sentences and short paragraphs with a bold word or phrase and a colon, 35 times across roughly 3,500 words, about one every hundred words. This is the same tell underslop already tracks as bold-header-list, scored B, not a separate one. The broader habit, turning any answer into headers, bullets, and colon lead-ins even in casual conversation, showed up in all 12 of the 12 models we logged. Laguna-xs is that habit's most concentrated single-model case. One gap is worth stating plainly: underslop's own pattern-match for this tell looks for a bulleted line. A colon label sitting alone mid-paragraph, the form most of the laguna-xs evidence takes, does not always trip that specific check, even though it is the same habit the check exists to catch.
What it is
A sentence or short paragraph that opens with a short label, usually bolded, followed by a colon and then the content: '**Here's what I notice:** you're already halfway to an answer.' It is the inline form of the same reflex underslop's checklist tracks as bold-header-list, turning free-form answers into report-shaped scaffolding, headers, bullets, colon lead-ins, even when the topic is casual or emotional. That reflex is universal across our sample, 12 of 12 models. What is narrow is the density in one model's output: poolside/laguna-xs leaned on the inline bold-colon version of it far harder than any other model we logged, mostly without a bullet marker in front of the line.
How often it shows up
This is measured at two resolutions inside the same checklist entry. Broadly, the scaffolding reflex, headers, bullets, and colon lead-ins together, showed up in all 12 of the 12 models we logged (research/reports/ai-tells-empirical.json, entry emp-002). Narrowly, one of those models, poolside/laguna-xs, used the bold colon lead-in specifically 35 times across about 3,452 words in a five-conversation sample, a rate of roughly 10 per 1,000 words (entry emp-025), well beyond what the other 11 models logged. We have not measured plain, non-bold colon usage or semicolons at all, so we make no claim about either.
Why readers flag it
A colon sets up an expectation: what follows should complete or explain what came before. When a document opens every third sentence the same way, the colon stops doing that job and starts functioning as a verbal tic, a label standing in for a transition rather than real emphasis. underslop's pattern-match for this tell currently looks for a bulleted line, '- **Header:** …'. A colon lead-in sitting loose in running prose, the way most of the laguna-xs evidence does, can slip past that specific check even though it is the same habit the check exists to catch.
The fix
Read the sentence with the colon replaced by a period or a comma. If nothing is lost, replace it. Keep the colon where it introduces a genuine list or a definition, somewhere a person would naturally reach for one instead of a plain transition. Do this by eye for inline labels. Do not assume underslop caught every instance on its own, since the form outside a bulleted line is the one most likely to slip through the pattern-match.
Examples
Here's what I notice: you're already halfway to an answer. Red flags to watch for: skipping meals, working through the weekend, and ignoring calls from friends.
You are closer to the answer than you think. The red flags to watch for are the ones that tighten a spiral: skipping meals, working through the weekend, ignoring calls from friends.
Great question! The bare minimum: skip the extras and focus on one habit. The real goal: consistency, not intensity.
The bare minimum means stripping away the extras and building one habit. Consistency matters more than intensity.
Questions
Does ChatGPT overuse colons?
One model in our sample did, heavily: 35 bold colon lead-ins across roughly 3,500 words, about one every hundred words (emp-025). Every model we logged did some version of the broader habit, turning an answer into headers, bullets, and colon lead-ins even in casual exchanges (emp-002, 12 of 12 models). What's unusual about laguna-xs is the density of the inline, non-bulleted form specifically.
What punctuation does AI overuse?
The em dash is the famous one, and our own logs back that up, roughly one every 80 to 90 words when a model leans on it. The colon lead-in is the less-discussed case: the same bold-header-list tell, at its most concentrated in a single model rather than spread evenly across all of them.
Does AI use semicolons a lot?
We have not measured this. Our logs were not built to track semicolon frequency, so we are not going to claim a pattern we have not counted.
Is the colon lead-in the same as the bold-header list?
Yes. It's the same tell id in underslop's checklist, bold-header-list, scored B. The bulleted version, a vertical list of '- **Label:** …' lines, is the common form: 12 of 12 models did some version of it. The inline version, a bold label mid-paragraph with no bullet in front of it, is rarer, and comes from a single model in our logs at unusual density. Different form, same reflex, same tell.
underslop marks this tell and the rest of the consensus set in any draft you paste, using the same open lint that scores it here. The free tier caps at 500 words a run, and your text is held for 0 seconds.