43 humanization rules for AI-generated cold email and LinkedIn copy

The rules don't fix it. The frameworks underneath do. The rules just keep the copy from announcing itself as AI on the way to landing.

Most humanization advice optimizes against detection - banned words, sentence-rhythm checks, length constraints. That misses the actual problem. AI copy fails for three structural reasons. The rules below patch the symptoms; reading these three first patches the cause.

Why AI cold copy fails before you fix the words

Failure mode 1

AI produces the statistical mean.

The average of every framework, every blog, every corporate content marketing piece it has been trained on. Smooth. Structurally predictable. Inoffensive. Nobody replies to the mean.

Failure mode 2

AI describes states. Humans describe movement.

Every AI message describes a state - "our product does X" - without implying movement from one state to another. Human readers, even reading a 50-word email, are wired for narrative: there was a before, now there's an after.

Failure mode 3

AI is structurally trained to hedge.

"Could be argued." "Potentially." "May help you." Hedging is the linguistic signature of a system that can't be held accountable for its claims. Humans notice this even when they can't name it.

Tier 1: Banned words (kill on sight)

These words spiked 200-6700% in AI-generated text (FSU / COLING 2025). A single one can trigger inbox suspicion.

delve, delves, delved, delving
tapestry
groundbreaking
realm
seamlessly
underscore, underscores, underscoring
showcasing, showcases
robust
pivotal
utilize - use "use"
leverage as a verb - use "use"
holistic
revolutionary
transformative
nuanced
multifaceted
paradigm
bespoke
intricate, intricacies
paramount

Tier 2: Structural AI tells

The contrastive reframe is the single most recognizable AI writing pattern. These ten patch the structural tells the model defaults to.

Never write It's not X, it's Y - state what something IS directly.
Never write The problem isn't X, it's Y - same pattern, different surface.
Never use Not just X but also Y - two short sentences instead.
Never start a sentence with a present participle ("Building on this…", "Thinking about…") - start with the subject.
Never use three bullet points of equal length - vary lengths, make one a fragment.
Never use perfect parallel structure in lists - break parallelism intentionally.
Never use rhetorical question into answer ("So what's the play? The play is…") - just say the play.
Never use closing call to reflection ("Something to think about.") - end on a specific.
Never use the corrective framing Not just X but Y in outbound - it judges the recipient's present.
Never lead with capability lists - lead with what changed for the reader, not what you can do.

Tier 3: Sentence rhythm - force burstiness

AI text has abnormally low burstiness (CMU / PNAS 2025). Sentences cluster around the same length. Variation is the unlock.

Never write 3 consecutive sentences within 3 words of each other in length. Mix short (under 8 words) with medium (15-20 words). Use fragments when natural.
The short sentence is the release after the build, not the default. A piece in all-short sentences drones at a different pitch.
Always use contractions. It's, we've, don't, can't, won't, they're.
Mix register within a single message. One formal observation + one casual aside.
Use occasional self-correction ("Actually, scratch that - the real issue is…") sparingly.

Tier 4: Hook + opening rules

First 210 characters of LinkedIn (the "…see more" cutoff). First 100 characters of email (Gmail Gemini reads this for deprioritization).

Never open with "I hope this email finds you well." Instant skip.
Never open with "I noticed that [Company] is…" or "I came across your profile…" - both are AI tells the recipient sees 12x a day.
Front-load value in the first 100 characters. Gmail Gemini reads this for deprioritization. Burying the value buries the message.
One specific detail per opener that only the sender would know about this prospect. Generic personalization (their company name, their title) doesn't count.
No questions in the opener of a cold email or connection request. Questions trigger filter flags.

Tier 5: CTA rules

One ask, one shot.

One ask per message. The second CTA dilutes the first - cut it.
The ask must be specific. "What's your experience with Clay at 5k rows?" beats "Thoughts?" by 3-5x in reply rates.
No sign-off beyond first name in cold email. "Best regards / Looking forward / Warmly / Cheers" all read like business letters from 2014.

The Lavender data: hard numbers on what actually correlates with replies

The Lavender cold-email analysis surfaces what moves reply rates independent of the words you pick. Use these as constraints on every draft.

Under 75 words+83% replies
3rd to 5th grade reading level+17% replies
Casual tone+23% replies
Informative tone−26% replies
First opened on mobile80% of cold emails

The frameworks behind the rules

The 43 rules patch surface-level AI tells. These three frameworks explain why the rules work - and what to do when the rules don't cover the situation in front of you.

Schwartz 5 levels of awareness

Cold email almost always lands in Level 1 (Unaware) or Level 2 (Problem-Aware). Lead with the prospect's world, never your product. The fatal mistake is pitching to someone who hasn't named the problem yet. At Level 1 the minimum move is to feel recognized. At Level 4 or 5 the minimum move is a specific next step.

Sugarman The slippery slide

The only job of any line of copy is to make the reader read the next line. Every sentence earns the next. Plant open loops. Front-load the most arresting truth you have. If a sentence could be cut and the reader would still continue, it should be cut.

Provost Sentence rhythm

"Five-word sentences are fine. But several together become monotonous." Vary length. Short, then long, then short again. The short sentence is the release. The long one is the build. Music, not drone.

How to use this

Short-form copy (cold email, LinkedIn DMs, comments)

Paste the rules directly into the generation prompt - the model follows them while writing. Layer them: persona, then constraints (Tier 1 + Tier 5), then structural rules (Tier 3), then output format, then 3 real examples. That order matters: persona anchors the register before the model generates anything; constraints override the model's RLHF-trained instinct to use "impressive" vocabulary.

Long-form copy (newsletters, blog posts, sequences)

Use the rules as an edit pass after the draft, not during. Embedding rules in the generation prompt for long-form produces compliant text that reads dead - the model optimizes individual sentences instead of the whole piece. Write first. Check rules second.

For batch generation at scale

Run every message through a second AI column with this checklist as a YES/NO filter. Any YES (a rule was broken) goes to manual review before export. Sample 10% of every batch by hand even with automated checks - no automation catches register problems.