Deslop

Deslop

Helps produce and edit text that reads like a specific, thoughtful human wrote it — not like a language model averaging across the internet.

Core Principle

AI writing regresses to the mean: it replaces specific, unusual facts with generic, positive-sounding statements that could apply to anything. The fix is always the same: be more specific, not more impressive-sounding.

Caveat before you start cutting: perfect grammar, a formal/academic register, or prose that just feels "stiff" are not proof of AI on their own — plenty of skilled or non-native writers produce exactly that. Don't over-correct a text that's merely formal into something falsely casual. The patterns below are signs, not a checklist to mechanically strip from every sentence.

Part 1: Red Flag Patterns (Detect & Remove)

1. AI Vocabulary Words

These words are statistically overrepresented in LLM output. One or two is fine. A cluster is a red flag. Remove or replace with simpler, more direct alternatives.

Era 2023–2024 (GPT-4):
additionally (to start a sentence), boasts, bolstered, crucial, delve, emphasizing, enduring, garner, intricate/intricacies, interplay, key (as adjective), landscape (abstract), meticulous/meticulously, pivotal, underscore (verb), tapestry (abstract), testament, valuable, vibrant

Era mid-2024–2025 (GPT-4o):
align with, bolstered, crucial, emphasizing, enhance, enduring, fostering, highlighting, pivotal, showcasing, underscore, vibrant

Era mid-2025+ (GPT-5):
emphasizing, enhance, highlighting, showcasing — plus excessive attribution to named sources

Model-specific: Grok overuses superficially "scientific" words — causal, empirical, correlate — and, unlike GPT, kept overusing underscore well into 2026.

Replacements:

• underscore → show, prove, explain, make clear
• delve → look at, explore, get into
• pivotal → key, decisive, major, turning-point (or just cut it)
• tapestry → mix, combination, blend
• leverage → use
• showcase → show, display, demonstrate
• foster → build, grow, support
• vibrant → busy, active, lively (or just describe what makes it vibrant)

2. Significance Inflation

Attaching unsupported claims about importance, legacy, or broader impact.

Patterns to cut:

• "marking a pivotal moment in the evolution of..."
• "represents a significant shift toward..."
• "was part of a broader movement to..."
• "stands as a testament to..."
• "reflects broader trends in..."
• "symbolizing its ongoing/enduring legacy"
• "setting the stage for..."
• "underscores its importance/significance"
• "contributes to the rich tapestry of..."
• "indelible mark", "deeply rooted", "focal point"

Fix: Either state the specific fact that demonstrates significance, or cut the claim entirely.

❌ "The launch marked a pivotal moment in the evolution of documentation tooling."
✅ "After the launch, customer onboarding time dropped from 3 weeks to 4 days."

3. Trailing Participle Analyses

Adding "-ing" phrases at the end of sentences to create an illusion of insight.

Pattern: [factual statement], [present participle]-ing [vague significance].

❌ "The station has 6 platforms, contributing to the socio-economic development of the region."
✅ "The station has 6 platforms."

Also watch for: ensuring, reflecting, highlighting, cultivating, fostering, encompassing — especially when tacked onto otherwise neutral sentences.

Related: Leading participle clauses. Same problem, different end. "Recognizing the importance of X, companies must Y" / "Considering current trends, it is crucial to Z." Rewrite as two sentences or drop the clause.

4. Copula Avoidance ("serves as" disease)

LLMs replace "is/are" with more elaborate verb constructions to sound more sophisticated.

Patterns:

• "serves as a" → is a
• "stands as a" → is a
• "marks the" → is the
• "represents a" → is a
• "boasts/features/offers" → has
• "ventured into politics as a candidate" → was a candidate

Fix: Use "is" or "are" when that's what you mean.

❌ "This feature serves as a solution for teams that need..."
✅ "This feature solves a problem for teams that need..."

5. Vague Authority Attribution

Attributing claims to unnamed or inflated sources.

Patterns:

• "Industry reports suggest..."
• "Experts argue that..."
• "Observers have noted..."
• "Some critics argue..."
• "Research shows..." (with no citation)
• "It is widely recognized that..."

Fix: Name the specific source, or rephrase as your own claim, or cut it.

Related — vague quantifiers. The same weasel move applies to numbers, not just sources: "significantly", "many", "often", "typically", "generally" stand in for a number the writer either doesn't have or didn't bother finding.

❌ "Significantly improves load times for most requests."
✅ "Cuts load times by 40% for the top 20% of requests."

6. Promotional / Travel-Guide Tone

Especially common in descriptions of products, companies, or places.

Patterns: nestled, breathtaking, vibrant community, rich cultural heritage, groundbreaking, renowned, diverse array, natural beauty, in the heart of

Product-specific: "commitment to excellence", "best-in-class", "seamlessly integrates", "powerful yet intuitive", "robust solution"

Fix: Describe what is actually there. Specifics > superlatives.

❌ "A vibrant platform that seamlessly integrates your workflow with powerful AI capabilities."
✅ "A platform that syncs with GitHub and generates doc drafts from your commit messages."

7. Canned Emphasis on Notability / Media Coverage

Specific to bios, case studies, and PR-adjacent copy: instead of stating what someone did, AI hits the reader over the head with proof that they were covered.

Patterns: "featured in [Outlet], [Outlet], and other media outlets", "profiled in multiple independent outlets", "maintains an active social media presence", "has been mentioned in [X] coverage relating to..."

Fix: Name the one or two sources that actually matter and say what they said. Cut "and other media outlets" — it's padding, not evidence.

❌ "She has been featured in Forbes, TechCrunch, and other prominent media outlets."
✅ "Forbes covered her Series A raise in March."

8. The Challenges Formula

Rigid structure that appears at the end of articles/sections:

"Despite its [positive quality], X faces several challenges, including [...]. Despite these challenges, X continues to [vague optimism] / [future initiatives] could..."

Fix: Either cut the section, or make it concrete: name the specific challenge, what's been tried, what the actual status is.

9. Not-Just-X-But-Also-Y Parallelisms (and Negation-Inversion)

Constructions that sound "balanced and thoughtful" but say nothing:

• "Not only... but also..."
• "It's not just about X, it's about Y"
• "Not X, but Y"
• "This isn't just a tool. It's a transformation."
• "No X, no Y, just Z"

Negation-inversion pattern: sets up a strawman the reader never held, then "corrects" it for false drama. Especially suspicious when clustered 2+ times in a short piece.

Reversed variant — "X rather than Y": the same move flipped ("prioritizing consolidation of power rather than ideological purity"). Same fix applies.

❌ "This isn't just a feature. It's a paradigm shift. It's not about speed. It's about rethinking how teams collaborate."
✅ "The feature cuts review time in half because it automates the three steps that used to be manual."

Fix: Either state X and Y as separate facts, or drop one of them. Occasional use is fine. Red flag when clustered or used to end a piece with a flourish.

10. Rule of Three (Overuse)

LLMs love tripling adjectives and noun phrases:

"keynote sessions, panel discussions, and networking opportunities"
"global SEO professionals, marketing experts, and growth hackers"

Fine in moderation. Suspicious when every single sentence does it.

Related — "from X to Y" false comprehensiveness: "Our services range from strategic planning to implementation support" sounds thorough but says nothing concrete. Just list what's actually offered.

11. Elegant Variation (Synonym Spinning)

Avoiding repeating a word by cycling through synonyms: protagonist → key player → eponymous character → central figure.

Fix: Just repeat the word. Repetition is cleaner than a thesaurus parade.

12. Collaborative Communication Artifacts

Phrases that belong in a chat response, not in published content:

I hope this helps, Of course!, Certainly!, Would you like..., is there anything else, let me know if, here is a [X] that, more detailed breakdown

Fix: Cut entirely.

13. Knowledge Cutoff / Hedging Disclaimers

• "As of my last knowledge update..."
• "While specific details are limited in available sources..."
• "This person maintains a low profile..."
• "Based on available information..."

Fix: Either state what you know, or don't include it.

14. Em-Dash Overuse (Model-Specific Tell)

GPT-4o uses em-dashes roughly 10x more than GPT-3.5; Claude is more restrained; humans use them sparingly for real emphasis. Two or more em-dashes per paragraph is a strong tell — especially when unspaced and clustered.

❌ "The strategy — which involves multiple steps — requires planning — particularly around resources."
✅ "The strategy has multiple steps. Resource planning matters most."

Rule of thumb: more than 1 em-dash per 200 words is suspicious. Mix em-dashes with commas, periods, and semicolons instead.

15. Uniform Rhythm (Burstiness Collapse)

Human writing is bursty: short punchy sentences mixed with long, clause-heavy ones. AI clusters around 12–18 words per sentence with low variance. Also: uniform paragraph heights (every paragraph ~3-4 sentences) reads as AI.

Self-check: Count sentence lengths in a paragraph. If they all sit within a 5-word band, rewrite at least one to be radically shorter (3-5 words) or longer (30+ words). Same for paragraph length: mix one-sentence paragraphs with longer ones.

❌ (all 12-15 word sentences) "Remote work changed everything. Teams spread across time zones. Communication became harder. Trust required new strategies. Leaders adapted quickly."
✅ "Remote work broke everything — teams scattered, time zones fragmented the day. Leaders had to rebuild trust from scratch. It was messy."

Don't confuse this with stop-start fragments (§16.1): varying length on purpose is good; chopping one idea into three flat sentences because it's easier to generate is not.

16. Rhetorical Scaffolding

Sentences that announce the action instead of performing it:

• "Let's explore...", "Let's break this down", "Let's dive in", "Let's unpack"
• "Here's the thing..."
• "But here's where it gets interesting..."
• "Let me walk you through..."

Like someone clearing their throat before speaking. Great writers begin; mediocre ones announce.

❌ "Let's explore why this matters. Let's break down the key components."
✅ "This matters because the cost model changed. The three key components are..."

Fix: Cut the announcement and start with the actual content.

16.1 Stop-Start Fragments

A close cousin: splitting one dependent idea across several choppy sentences instead of writing it as one.

❌ "Previously this was manual. Now it's automatic. This saves time."
✅ "What used to be manual is now automatic, saving time."

Short sentences used deliberately for emphasis (see §15) are fine. This is different: it's fragmenting a single thought because generating three short sentences is easier than building one properly subordinated one.

16.2 Summary-Style Transitions

Opening a new paragraph by recapping the one before it, instead of moving forward.

❌ "With this setup complete, we can now look at..." / "Now that we've explored X, let's turn to Y."
✅ "The catch is..." / "In practice, this means..."

16.3 Cold-Open Paragraphs

A paragraph whose first sentence works as a standalone heading — no connective tissue to what came before. Carry the subject forward instead of starting fresh each time.

❌ (new paragraph, no lead-in) "Caching reduces latency significantly."
✅ "Because the API call was the bottleneck, caching reduces latency here."

16.4 Reused / Template Framing

A generic angle that could open any article on any topic, not something specific to this one.

❌ "The question most teams face is whether to build or buy."
✅ (skip the throat-clearing, state the actual finding or fact first)

17. Opener Formulas

AI loves vague trend-setting intros:

• "In today's fast-paced world..."
• "In an era of [rapid change / digital transformation / AI disruption]..."
• "As we navigate the [modern / evolving / complex] landscape..."
• "Imagine a scenario where..."

Fix: Start with a concrete problem, a specific observation, a number, or a scene.

❌ "In today's fast-paced business environment, efficiency is more important than ever."
✅ "Last quarter we lost 30% of trial users at the same onboarding step. Here's why."

18. Closer Formulas

AI can't resist wrapping up:

• "In conclusion...", "In summary...", "To sum up..."
• "As we move forward...", "Looking ahead...", "The future holds..."
• "The path ahead is clear..."
• Restating 3 key points verbatim in the final paragraph

Fix: End on a specific fact, a question left hanging, a surprising observation, or a concrete next step. Don't recap — readers just read it.

19. Nominalization Overuse

LLMs turn verbs into abstract nouns ("implement" → "implementation", "decide" → "decision-making", "analyze" → "analysis"). This creates dense, corporate-sounding text. Used at ~1.5–2x the rate of human writing.

❌ "The implementation of this strategy requires the establishment of clear governance and the acquisition of new resources."
✅ "To roll out this strategy, we need governance and new resources."

Self-check: Count words ending in -tion, -ment, -ance, -ity, -ness in a paragraph. If more than 3-4 per paragraph, replace at least half with verbs.

20. Spec-Sheet Voice

Sentences that read like a line off a datasheet instead of something a person wrote: provides, is configurable, is explicitly labeled, supports X functionality.

❌ "The endpoint provides configurable retry logic and is explicitly labeled as idempotent."
✅ "The endpoint retries automatically; you can tune how many times. Calling it twice is safe."

21. Personified Artifacts (technical writing)

Non-human things performing human, physical actions — common in product/API copy.

❌ "The browser hands the server a URL." / "The token holds the session."
✅ "The browser sends the URL to the server." / "The session is stored in the token."

Fix: Give the object the verb that actually describes what it does, not a borrowed human action.

22. Formatting Overload

LLMs over-format because markdown structure is easy to generate:

• Bold-inside-prose for fake emphasis ("the most important thing is..."). More than 1-2 bolds per section = tell.
• Bold-lead-colon pattern: "Key insight: ..." repeated every paragraph.
• Excessive H2/H3 headers where prose would flow.
• Numbered lists for flowing thoughts ("Here are 5 reasons: 1. Cost. 2. Speed. 3. Quality...") when a sentence would do.
• Rigid scaffolding: Overview → Key Points → Challenges → Conclusion.
• Emoji clusters (🔑🎯💡) as bullet markers in serious content.

Fix: Prose by default. Use lists only for genuinely list-like content (comparisons, instructions, features). Use headers only for navigation in long pieces. Bold sparingly, for one key phrase per section at most.

23. Self-Answered Rhetorical Questions

AI asks a question the reader wasn't asking, then answers it. Repeatedly.

❌ "What does this mean for your team? It means you need to rethink communication. Why is that important? Because remote work changed everything."
✅ "Remote work changed how teams communicate. Async beat sync, not because anyone planned it, but because timezones won."

Fix: If you pose a question, either leave it hanging for the reader, or answer with evidence (not restatement). One rhetorical question per piece is usually enough.

24. Over-Hedging Clusters

Stacking cautious qualifiers: "arguably", "potentially", "it could be said", "it's worth noting", "one could argue", "may", "might", "could potentially", "perhaps".

❌ "This approach could potentially unlock new opportunities, arguably transforming how teams might leverage existing resources."
✅ "This approach cut delivery time by 40% on two projects. It might not scale — but so far it works."

Fix: Commit. If you're uncertain, say so directly once. Don't blanket the whole text with maybes. (Note: one hedge in isolation is normal human writing — see Part 2, §13 — this is about stacking, not using any hedge at all.)

Part 2: Positive Principles (What Human Writing Does Instead)

1. Specific > Generic. "Our conversion rate doubled" beats "significant improvement was achieved."

2. Show, don't infer significance. Let facts speak; don't tell readers what to conclude.

3. Use "is/are" freely. It's direct. It's clear. It's fine.

4. Repeat words. "The product → it → the tool → this solution" is AI. "The product → the product" is human.

5. Cut trailing analysis. If a sentence ends with a present participle clause that explains why the previous fact matters, cut that clause.

6. Name sources or cut attribution. "Experts say" = AI filler. Name the expert, or say it yourself.

7. One structural tell to break: avoid the rigid "background → current state → challenges → future outlook" outline. Real writing has narrative, not sections.

8. Vary sentence length aggressively. Short sentences. Then a medium one that adds context. Then a longer one that complicates the picture or adds a caveat, because that's how ideas actually develop — not in perfectly parallel triplets. Include at least one sentence under 6 words and one over 25 words per page.

9. Prefer verbs over abstract nouns. "We decided" beats "decision-making was undertaken". Verbs carry energy; nominalizations drain it.

10. Don't announce — just do. Cut "Let's look at...", "Here's the thing...", "Now let me explain...". The reader is already reading.

11. Don't recap at the end. End on something the reader hasn't read yet — a specific fact, a surprise, a question, a next action.

12. Format for readers, not for models. Prose by default. Bold only for one phrase per section. Lists only for list-like content.

13. Use plain words over their stiff synonyms — freely. "Wrote" not "authored". "Used" not "utilized". "Moved" not "relocated". "Tried" not "attempted". "Died" not "passed away". AI systematically avoids the plain version in favor of the fancier one; using the plain word is one of the more reliable ways to not sound like AI, not a dumbing-down.

14. Make definitive claims when they're true. "Was the first", "is the only", "one of the best" — AI hedges away from these even when accurate, because it's trained to avoid overclaiming. If it's true and you can back it up, just say it.

15. Don't reflexively strip "unnecessary" wordiness. Phrases like "in order to", "as a result of", "the fact that", "a part of" are actually more common in human writing than AI output, not less. If a passage feels over-processed, adding a little natural wordiness back can help more than more cutting.

16. Simple "there is / it has" constructions are fine. AI tends to dress these up into something more elaborate. You don't need to.

Part 3: Editing Protocol

When asked to clean up or rewrite AI-sounding text, proceed in this order:

1. Identify clusters of red-flag vocabulary (Part 1, §1)
2. Find trailing/leading participle phrases (§3) and cut them
3. Restore copulas — replace "serves as", "stands as" with "is" (§4)
4. Remove significance inflation (§2) unless specific evidence is provided
5. Flatten promotional tone (§6) — replace adjectives with facts
6. Check attribution and quantifiers (§5) — name sources, cite numbers, or cut
7. Count em-dashes (§14) — reduce to 1 per 200 words max
8. Check rhythm (§15) — inject sentence-length variation and paragraph-length variation, but don't confuse this with stop-start fragmenting (§16.1)
9. Cut scaffolding (§16) — "Let's..." preambles, announcements, rhetorical throat-clearing, summary-style transitions (§16.2), cold-opens (§16.3)
10. Kill opener/closer formulas (§17, §18) — start specific, end specific
11. Reduce nominalizations (§19) — turn noun phrases back into verbs
12. Fix spec-sheet voice and personified artifacts (§20, §21) — mostly relevant for technical/product copy
13. Deformat (§22) — less bold, fewer bullets, fewer headers unless they earn it
14. Restore plain words and definitive claims (Part 2, §13-14) — don't over-launder simple language back into something fancier
15. Passive voice check: append "by monkeys" to the sentence. If it still parses, it's passive — rewrite active where it matters.
16. Second-read test: read each sentence once at speaking pace. If you have to re-read it to find the subject, the action, and the consequence, cut the metaphor verbs or the pronouns reaching back too many sentences.
17. Read aloud test — if a sentence sounds like it was written for a brochure, rewrite it

When generating new content, apply Part 2 principles from the start rather than editing after.

Part 4: Context-Specific Notes

Product / SaaS copy

Most common trap: "powerful, intuitive, seamless" trinity. Replace with one specific capability + one specific outcome.

Release notes / changelogs

Most common trap: significance inflation ("marking a major milestone"). Just describe what changed and what it enables. Watch for spec-sheet voice (§20) and personified artifacts (§21) here specifically — release notes describe systems, and it's easy to slip into either.

Blog posts / thought leadership

Most common traps: challenges formula, significance trailing phrases, vague attribution. Use named examples, specific numbers, first-person opinions.

Marketing landing pages

Most common traps: promotional tone, copula avoidance. Write like you're explaining to a smart friend, not pitching to a committee.

Telegram / short social posts

Most common traps: opener formulas, self-answered rhetorical questions, excessive emoji/bold. A TG post should sound like a voice message you typed out, not a mini-blog-article with headers.

Bios / case studies / PR-adjacent copy

Most common trap: canned emphasis on notability (§7) — listing outlets instead of stating facts. Name the one thing that matters; cut "and other media outlets".

Quick Reference: Before/After