unslop

Unslop

Prevents and removes the patterns that make readers recognize text as AI-generated – in German and English. This file follows its own rules; the tells it quotes are data, not violations.

Core principle: three levels, one universal rule

AI slop lives on three levels. Swapping words alone achieves little: structural features by themselves identify AI text with over 93 % accuracy, and professional surface rewriting barely shifts detection (StoryScope 2026). Word lists also age with every model generation. So always work through all three levels, weighted toward 2 and 3:

1. Word level – vocabulary, stock phrases, punctuation. The weakest but most visible signal. Lists: references/word-lists.md.
2. Sentence and rhythm level – uniform sentence lengths, parataxis chains, triads, negative parallelisms, load-bearing metaphors, hedge stacking, copula avoidance.
3. Structure level – the spelled-out takeaway, uniform paragraphs, template scaffolding, vague instead of named references, narrative arcs that explain and resolve everything, shape convergence across texts. The most durable signal.

The universal rule: swap the vague claim for the checkable fact. That means a figure, a name, a date, a price, or the mechanism behind the claim. „Deutlich schneller“ needs a sourced time comparison; „von 4 Klicks auf 1“ needs sourced click counts and does not itself prove a shorter duration. But when the concrete fact is missing, mark the gap; never invent a number, anecdote, or source. A made-up detail damages trust more than an honest gap.

The short version

When attention is scarce, use these ten checks:

1. Swap vague claims for sourced, checkable facts; keep relative measures such as „halbiert“ at their original strength. Flag missing evidence rather than inventing it.
2. Pick the register first (profile table below): the same rule can be right in one register and wrong in the next.
3. State the core point once, except in summaries required by the register; every sentence adds information; no formulaic opener or generic closer.
4. Vary sentence construction without staccato; avoid parataxis chains and decorative triads, keeping groups whose meaning has three parts.
5. At most one rhetorical negative parallelism per piece, usually none; corrective instructions, corrected figures, and authenticity statements stay.
6. Check word-list hits in context; apply tier thresholds and protected uses. Keep the punctuation budget (em dash: rare in prose, zero on slides).
7. Prose: actor and verb early, the everyday word. Slides: sober nominal style.
8. Headings and slide titles name the content, usually with noun labels; deliberate action titles carry checkable claims.
9. Name things (sources, tools, dates); name feelings instead of performing them.
10. Never add facts, persona, or staccato the source did not have.