AI-likeness detection: what the rubric actually catches

The 'sounds AI' problem isn't only about emails that AI wrote. Human-written emails increasingly read as AI because the conventions of corporate writing have converged with the patterns AI tools default to.

Our AI-likeness detector flags structural tells, not source. Here's what it catches and why.

The structural tells

Five patterns the detector flags consistently:

Aggressive symmetry. Lists with three bullets that all start the same way, or paragraphs with parallel sentence structure across multiple sentences. Humans use parallelism intentionally; AI uses it by default. Sample flag:

We help you ship faster, scale smarter, and grow bigger. We support teams of every size, in every industry, in every country.

This reads as AI because the symmetry is too clean. Human writing has more irregularity.

Hedge-laden openers. Phrases like 'I hope this email finds you well,' 'I trust this finds you,' 'I just wanted to reach out.' Generic conversational placeholders that humans use when they don't know what to say. AI defaults to them because they're statistically common in training data.

Verb-noun nominalizations. 'Implementation,' 'utilization,' 'optimization' used as nouns where verbs would carry the meaning more directly. The pattern is corporate writing convention; it also reads as AI because AI defaults to nominalized forms.

Empty intensifiers. 'Robust,' 'leverage,' 'delve,' 'comprehensive,' 'cutting-edge,' 'unprecedented.' Words that signal seriousness without adding meaning. The product literally penalizes these in scoring because they correlate with low-conversion writing.

Over-hedged claims. 'May possibly,' 'could potentially,' 'tends to typically.' Stacked hedges that drain conviction. Used once, hedging is calibrated honesty. Used reflexively, it reads as both AI-flavored and untrustworthy.

What the detector does and doesn't do

Our detector is a Haiku pass running on top of a local heuristic. The heuristic catches the easy patterns (specific word usage, structural symmetry); the Haiku pass catches the subtler tells (hedging density, sentence-rhythm uniformity).

What it doesn't claim: that flagged emails were written by AI. The detector flags structural patterns that read as AI; the source could be a human, an AI, or a human editing an AI draft. The structural patterns are the problem regardless of source.

What you get: inline annotations on the flagged phrases, with a one-tap rewrite that breaks the pattern. The rewrites are typically minor — replacing 'leverage' with 'use,' restructuring a too-symmetric list, dropping a hedge.

Why this matters

Two reasons.

One: Reader pattern recognition has improved. Recipients have learned to spot AI-flavored writing and now discount it. An email that reads as AI converts worse than one that reads as human, independent of any other variable.

Two: Even when humans write the email, the structural patterns are correlated with worse writing. The detector catches conventions of corporate marketing prose that have always been weaker than the alternatives; they just happen to also be the patterns AI defaults to.

A useful exercise

Run your last 5 sends through the detector. The likely result: 1-2 flags per email, mostly on hedges and intensifiers. Each flag is easy to fix — replace 'leverage' with 'use,' drop 'just,' restructure one parallel list.

The cumulative effect: emails that sound more like a person wrote them, which is what you wanted anyway.

The hardest pattern to catch

The hardest tell, which the detector still misses sometimes: structural completeness.

AI-written emails tend to be too complete. Every paragraph touches the topic; nothing is left for the reader to infer. Real human writing often leaves things unsaid — the reader's mind fills in the gap, which creates a sense of connection.

When you write an email, ask: 'is there anything I can leave out and trust the reader to get?' If yes, leave it out. The space is doing work.

What to do

Three takeaways:

1. Run your sends through AI-likeness detection (ours or a similar tool). Easy structural improvements per email.
2. Manually check for completeness. Anywhere your draft says everything, try saying less.
3. Notice the conventions you've inherited. If you reflexively start emails with 'I hope this finds you well,' you're using a phrase because it's safe. Replace it with nothing or with something specific.

For the AI-likeness detector in action, try our editor — anonymous, no signup. And our scoring guide has the full rubric.