# sloptells — all tells (generated 2026-07-02)

A style linter's taboo list, not a detector. Measured; see https://sloptells.com/methodology

## active

### “rather than” (lexical, ×4.7 vs humans)
The 2026 flagship. Models reach for “rather than” where humans write “instead of”, “not”, or just rephrase. Five times the human rate across every register we measure, and rare in acclaimed pre-AI prose — which makes it one of the safest tells to act on.

### “not just X” (lexical, ×8.0 vs humans)
The soft residue of the famous “it's not X, it's Y” construct. The hard form has been trained away (see its entry), but the reflex survives as “not just”: elevating whatever is being described by denying a smaller frame first.
Note: Also appears in good human prose — weighted down accordingly.

### “That said, …” (lexical, ×12 vs humans)
The pivot connective models use to perform balance. Twelve times the human rate in matched comparisons; the single strongest frame in our first experiment.

### “(this) feels like” (construct, ×6.3 vs humans)
Hedging with vibes. Where a human states an opinion or a fact, models report an impression: “this feels like a step change”. Roughly 10× the human rate.

### “the real question/problem is…” (construct, ×10 vs humans)
Reframing-as-insight: declaring what the *real* issue is. “The real question to ask yourself is…”, “the real issue may not be X as much as Y”. Seven times the human rate.

### “one of those X” (construct, ×12 vs humans)
“This is one of those cases where…” — the false-familiarity gesture, filing the topic into a recognizable genre of situations. 26× the human rate.

### “the actual X” (lexical, ×4.3 vs humans)
Emphatic “actual/actually” to signal getting-to-the-substance: “the actual problem”, “what's actually happening”. Six times the human rate.

### “genuinely” (lexical, ×25 vs humans)
Sincerity as an intensifier: “genuinely impressive”, “genuinely curious”. 48 occurrences vs 1 in matched human text — near-categorical.

### “worth noting/mentioning” (construct, ×11 vs humans)
The throat-clearing qualifier before a fact: “it's worth noting that…”. Models use “worth” itself at 13× the human rate.

### “curious what others think” (discourse, ×26 vs humans)
The engagement-bait close, and “curious” as a social lubricant in general: “I'm curious how it handles…”, “genuinely curious whether…”. 8× the human rate; widely complained about on social platforms.

### “in practice” (lexical, ×14 vs humans)
Part of the emerging plain-spoken-confidence cluster: sounding grounded and empirical without citing anything. 12× the human rate and rising across model generations.

### hedge-adverb pileup (lexical, ×1.9 vs humans)
“usually”, “mostly”, “typically”, “generally”, “slightly” — each is normal English; the density is the tell. Models qualify nearly every claim with one. Any single match means nothing; several per paragraph reads as AI.
Note: Density tell — scored very low per match by design.

### “especially” (lexical, ×2.6 vs humans)
The add-a-nuance suffix: “…, especially if you're X”. Nearly 3× the human rate in advice registers.

### “a few things/considerations” (lexical, ×1.9 vs humans)
The enumerate-everything opener: “A few things to keep in mind…”. Precedes the bullet pileup. 3× the human rate.
Note: Common in human writing too — low weight.

### “culinary” (lexical, ×17 vs humans)
Humans who cook say “cooking”. Models say “culinary”: “a strong culinary culture”, “culinary applications”. 26 occurrences vs 1 in matched human answers — register-elevation in one word.

### therapist-mode vocabulary (discourse, ×5.4 vs humans)
Advice answers that sound like a counseling session: “boundaries”, “frame it as”, “in the moment”, “a safe space”, “meet them where they are”. Strongly overrepresented in parenting-register output.

### “what (actually) matters” (construct, ×7.5 vs humans)
Gravitas by decree: “what matters is…”, “it matters because…”. 6× the human rate and rising across model generations.

### bold everywhere (formatting, only in AI)
**Bolding** key phrases in contexts where no human would: forum answers, comments, casual replies. Humans in our matched corpora essentially never bold; models do it constantly (90×+ in advice registers).

### bullets for a conversation (formatting, ×8.6 vs humans)
Answering a casual question with a formatted mini-wiki: bullet lists, sections, a summary. In the cooking register models used bullets at 134× the human rate. A list is innocent on its own; the tell is a list where a human would write three sentences.
Note: Scored per item; only fires meaningfully when text is short relative to its list count.

### typographic quotes (formatting, ×18 vs humans)
Models emit “curly” quotes; humans typing into a web form produce "straight" ones. 7–26× the human rate depending on register. Partly a platform artifact, not a style choice — but as a signal it's nearly free.
Note: Weak on its own (word processors also produce curly quotes) — meaningful in combination.

### sentences that march in formation (cadence)
The robot monotone is not short sentences — it's uniform ones. AI sentence-length variance is consistently below human (coefficient of variation ~0.52–0.57 vs ~0.64–0.67) in every register we measure.
Note: Listed but not yet auto-scored by the checker (length-gated metric).

### suspiciously clean punctuation (formatting)
An inverse tell: the *absence* of messy human punctuation. Current models use exclamation marks at ~2% of the human rate, ellipses at ~10%, question marks at half. Nothing dangles, nothing trails off. Too clean to be human.
Note: Listed but not yet auto-scored (requires enough text to measure absence).

### emoji as structure (formatting, only in AI)
🔹 Bullet emojis, ✅/❌ contrast lists, the 🎉 after an announcement, 🚀 anywhere. Present in ~half of AI LinkedIn posts, absent from ordinary human prose. The single most reliable social-slop marker we measure.

### headers on a social post (formatting, only in AI)
A markdown title on a LinkedIn post — “# One Year In: What Nobody Tells You” — as if a 150-word post needed an H1. Half of the punchy-prompt posts in our eval set open this way.

### “the privilege of” (lexical, only in AI)
“I've had the privilege of working alongside…” — gratitude in its corporate dress uniform. Zero occurrences across 556 human comments and the entire canon.

### “X taught me Y” (construct, only in AI)
Everything is a teacher: the failed project, the difficult customer, the canceled flight. “Three decades taught me something:”. In 10% of AI LinkedIn posts, zero in human baselines.

### “here's what actually happened” (construct, ×16 vs humans)
The reveal-promising colon: “Here's what surprised me most:”, “Here's what nobody tells you:”. Setup theater before ordinary content.

### “The lesson? …” (construct, only in AI)
The one-word-question pivot: “The lesson? Always follow up.” “The real wins?” Rhetorical self-interview as paragraph glue.

### “Not because X. Because Y.” (construct)
The fragment-pair reveal: “Not because I'm done. Because I'm finally starting.” Rare per-post but essentially exclusive to slop — zero hits across all human corpora including the canon.

### “showing up” (lexical, ×4.4 vs humans)
“It's early mornings, long drives, and showing up — again and again.” Consistency-as-virtue vocabulary; 15× the human-baseline rate.
Note: Also lives in good human writing (weighted down accordingly).

### the mic-drop closer (cadence)
Ending on a five-words-or-fewer punch: “Let's see what happens.” “They were wrong.” “Watch this space.” Present in ~half of AI LinkedIn posts vs 8% of human comments.
Note: Low weight by design — plenty of good writers land a short closer; it counts in combination.


## saturated

### “thrilled/excited/humbled to announce” (discourse, only in AI)
The announcement voice: “Grateful and excited to share that…”, “Humbled to announce…”. The oldest LinkedIn cliché family — humans invented it, models industrialized it. As a slop linter we flag it whoever wrote it.

### journeys and chapters (discourse, only in AI)
“What an incredible journey”, “excited for the next chapter”, “here's to the ride”. Life as a memoir with quarterly milestones. In 12.5% of AI LinkedIn posts; zero in matched human prose.

### “And honestly? …” (construct, ×0.6 vs humans)
Fake candor as a rhetorical fragment: “And honestly? That's rare.” Everyone has learned to spot it; the pattern persists in social-media-register output.

### “here's the kicker/thing” (construct, only in AI)
Announcing the good part instead of delivering it: “But here's the thing…”, “Here's the kicker…”. Payoff-promising phrases that rarely precede a payoff.

### “You're not alone/imagining it” (construct)
Unsolicited therapeutic reassurance in non-therapeutic contexts — a business article that suddenly comforts you: “You're not imagining it. You're not broken.”

### “no fluff” (lexical)
The phrase “no fluff” is intrinsically fluff. Also: “no jargon”, “no buzzwords”. A self-defeating seriousness disclaimer that humans read as AI (or as trying too hard, which lands the same).

### signposting overuse (discourse, ×0.3 vs humans)
“Firstly… Secondly… In addition… The key point is…” — guiding the reader through a text too short to need a map. Borrowed from academic writing, deployed everywhere.

### “it's not X, it's Y” (construct, ×4.9 vs humans)
Contrast framing — the most attested construct tell of the era: “It's not chaos, it's clarity.” Reports of its death were premature: our first small experiment suggested it was fading, but the larger multi-register corpus shows it at ~4.5× the human rate and *rising* across model generations. A good example of why we measure instead of assuming.


## fading

### “Great question!” (discourse, only in AI)
Sycophantic openers, measurably dying: 9× less frequent in current models than their 2024/25 predecessors. The clearest fingerprint of the anti-sycophancy training wave.

### the 2024 hedging cluster (lexical)
“might”, “could”, “consider”, “perhaps”, “potential” — the tentative, option-listing voice of the 2024 model generation. Current models use “might” at one-fifth the rate of their predecessors. Replaced by the plain-spoken confidence cluster (see “in practice”, hedge-adverbs).
Note: Words too common to pattern-match usefully; tracked as corpus rates, not by the checker.


## stale

### the em dash (formatting, ×26 vs humans)
The internet's favorite tell — “the ChatGPT dash” — and our headline negative result: acclaimed pre-AI human prose uses em dashes as much or more than current models do (0.67 vs 0.56 per 1k chars in our measurements). Accuse someone of being AI over an em dash and you will flag more good writers than models.
Note: Kept public as history and as a caution: single formatting marks are terrible evidence. Density and co-occurrence are what count.


## retired

### staccato mic-drop sentences (cadence)
“Short sentences. Back to back. For emphasis.” Widely attested as a tell — and refuted by our measurements: current models produce runs of short sentences *less* often than humans in every register we tested. The robot monotone is uniformity, not shortness (see “sentences that march in formation”).
Note: Negative result, published on purpose.

### “delve” (lexical)
The original. The word that launched a thousand detector threads in 2023–24. Absent from every current model we measure. Rest in peace.
Note: Zero occurrences in our current-model corpora. Seeing it in fresh text suggests an old model, a cheap model, or a human joke.

### “tapestry” (lexical)
“A rich tapestry of…” — 2023-era purple vocabulary, trained away alongside “delve”. Not observed in current models.

### “a testament to” (construct, only in AI)
Everything was a testament to something in 2023–24. Not observed in current models.