The AI roleplay slop lexicon: 95 phrases, with receipts

"Are you tired of ministrations sending shivers down your spine? Do you swallow hard every time their eyes sparkle with mischief and they murmur to you barely above a whisper?" That's the opening of Sukino's Banned Tokens list, the most-circulated slop lexicon in the roleplay community, and it manages to commit five entries from its own list in two sentences. Everyone who chats with AI characters knows these phrases by feel. This page is the feel, written down, with sources.

We compiled it for our own card-authoring pipeline (the scorer that enforces it is open source), cross-checking three independent kinds of evidence so no single taste dictates the list: community ban lists (Sukino's), statistical over-representation measured against human baselines (the Antislop paper, ICLR 2026 submission), and benchmark component weights (EQ-Bench's slop score: 60% word frequency, 25% contrast constructions, 15% slop trigrams). 95 phrases total, ten groups. Data note: Foreverse research, 2026-07.

A scale-setter before the tables, because the numbers are the argument: Antislop measured the character name "Elara" at 85,513x the human baseline in one model's creative output, the trigram "heart hammered ribs" at 1,192x, and one mid-size model producing "eyes never leaving" 102 times across 96 test prompts. These aren't vibes. They're measured fixations.

A. Body reactions

The workhorse group. Twelve printed here; three more from this family are NSFW-register and live with group I below.

B. Voice

C. Eyes and expressions

D. Touch and motion

E. Sentence constructions

The highest-value group, because constructions survive vocabulary swaps. If you only police one thing, police the first row.

F. Stock nouns and similes

G. Atmosphere

H. Endings and outlooks

The fishing-line family. These matter double in character cards because a greeting that ends on one trains the model to end every reply on one.

I. The NSFW cluster

Ten phrases, plus the three body-family ones held back from group A. This is a general-audience page, so we won't print them; if you need them for filtering, they're the NSFW_PHRASES constant in the scorer source. On SFW cards our gate treats any hit from this group as an automatic fail rather than a deduction.

J. Naming slop

Models fixate on character names, and the fixations are the most statistically extreme entries on the whole list.

The fix isn't a longer ban list — models will fixate on new names next year. It's a naming procedure: anchor in a real language culture (Welsh, Polish, Nigerian, Irish surnames…) or coin something, then search the result to confirm you didn't collide with a franchise.

How to use a slop list without ruining your prose

By cluster density, not word bans. The anti-slop projects themselves warn about this: not every use of these words is wrong, but clusters of them are a giveaway. Humans wrote every phrase here first — that is precisely how the models learned them — so a zealous find-and-delete pass produces prose with a different problem: the sanded, evasive texture of text that's afraid of itself. Our house rules quota the constructions ("not X, but Y" at most once per card, emotion cocktails at zero, fishing endings at zero) and treat the vocabulary groups as rewrite triggers rather than deletions.

Also worth knowing: slop fingerprints cluster by model family. The Antislop data shows each model line has its own favorites, so a list tuned on one model's output will under-detect another's. That's the strongest argument for pattern-level rules (group E) over vocabulary-level ones — constructions transfer across models better than words do.

The two numbers most slop lists don't tell you

First: star counts don't filter slop. We scanned the greetings of the 30 top-starred SFW character cards on chub.ai (sampled 2026-07, human-written, community-validated) with a 13-pattern subset of this lexicon and got 20 hits across the sample — "you feel" in 5 cards, "a mixture of" in 4, plus scattered gleaming eyes and mischievous grins. The most popular human cards in the world average 0.07 lexicon hits per greeting. Slop phrases are not an AI marker; slop density is.

Second: flagship models now pass word-list checks in a single polished greeting. In our calibration set — 9 raw greetings from 5 current model families, generated with a plain prompt and zero anti-slop instructions — the average was 0.22 hits, with the two strongest models scoring zero across the board. The separation direction is correct (0.07 human vs 0.22 machine) but the honest reading is that word lists catch weak models and long multi-turn drift, not a frontier model on its best behavior. Where the slop actually resurfaces is turn forty of a chat, when the context fills with the model's own output and the fixations compound.

Method note

Sources: Sukino's Banned Tokens (community ban list, continuously updated), the Antislop paper (arXiv 2510.15061, statistical over-representation vs human baselines), EQ-Bench slop-score components, NousResearch's anti-slop writing guide, and RP community complaint threads for the 2025-26 additions (calculating gaze, predator similes). Human baseline scan: 30 top-starred SFW cards, first messages, July 2026. Machine baseline: 9 single-shot greetings across 5 model families, plain prompt, no style instructions. The scorer implementing all of it — including the red lines and the construction quotas — is score_card_en.py in character-card-skills, runnable with no dependencies. If a phrase on this list false-flags your human prose, file an issue; reader-flagged false positives are how the list stays calibrated. And if you came here wondering about the em dash: it's deliberately absent, because the data says it isn't a tell in English.

We maintain this lexicon because we run character-card authoring on top of it — the full writing guide shows where the quotas slot into the larger procedure, and the cards it produces are chattable in Foreverse or any SillyTavern-compatible frontend.