Building an AI Detector for Your Inbox: Techniques That Actually Work in 2026

Why Inbox-Side AI Detection Matters Now

Roughly six out of ten pieces of unsolicited outbound email arriving in a professional inbox in 2026 are AI-drafted. Some are useful; most are not. The signal-to-noise problem has degraded past the point where a human can efficiently triage by reading the first sentence.

This piece is about the other side of the Presend product surface — not helping you write better outbound, but helping you filter and prioritize the flood of AI-drafted inbound. It covers the detection techniques that actually work in production, why the well-known third-party detectors underperform on short-form email, and how to build a lightweight filter for your own inbox.

Why Off-the-Shelf Detectors Struggle on Email

Tools like GPTZero, Originality, and Copyleaks were built primarily for essays and long-form articles. Their published F1 scores in the 0.85-0.95 range apply to inputs of 500+ words. On the 80-200 word block that constitutes a typical business email, F1 drops to 0.55-0.70. Two structural reasons:

• Not enough signal. Stylometric fingerprints (perplexity, burstiness, sentence-length variance) require a corpus. A single 100-word email has insufficient text to compute a stable estimate.
• Domain shift. Business email has its own genre conventions (formal openers, ceremonial signoffs, bullet lists) that overlap heavily with AI output. A high perplexity score on a formal business email does not reliably discriminate AI from a careful human writer trained in corporate style.

The practical implication: you cannot just plug a general-purpose AI detector into your inbox and expect it to work. You need email-specific techniques.

Technique 1: The Phrase Signature Match

The most effective single technique is a curated phrase library. Certain phrases occur in raw LLM outputs at rates 15-50x their base rate in human corporate email:

• "I hope this email finds you well" — 41x
• "In today's fast-paced world" — 38x
• "Should you require any further information" — 27x
• "Please don't hesitate to reach out" — 22x
• "I wanted to take a moment to" — 19x

A weighted phrase count over the email body produces a "phrase signature score" that correlates highly with AI authorship. Simple, fast, and hard to fool without deliberate effort.

Trade-off. Detectable and gameable. An LLM prompt of "avoid clichés" reduces the phrase signature dramatically. Effective against unedited raw AI, weak against carefully edited AI.

Technique 2: Structural Fingerprinting

Even carefully edited AI often retains structural fingerprints:

• Paragraph-length variance. AI writes metronomically. Compute the coefficient of variation across paragraph lengths; values below 0.20 are suspicious.
• Sentence-length distribution. Real business writing has a bimodal distribution (short punches + long arguments). AI is unimodal around 20 words. A Kolmogorov-Smirnov test against a reference distribution catches this.
• Tricolon density. Rule-of-three constructions per 100 words. AI averages 2.1; human averages 0.6.
• Transition word ratio. "Furthermore," "moreover," "additionally," "in conclusion" per 100 words. AI averages 1.8; human averages 0.4.

Combine these as features in a small classifier (logistic regression is often enough) and you get a robust signal that survives even careful editing.

Technique 3: Contraction Rate

A remarkably durable signal. LLMs in formal register default to expanded forms ("do not," "cannot," "I am"). Native business writers use contractions at a rate of roughly 0.8 per 100 words even in formal contexts. AI averages 0.15 per 100 words.

Contraction rate alone is a weak feature but it is nearly impossible to fake reliably. Even prompts of "use contractions" produce forced, unnatural contraction placement that a native speaker can flag.

Technique 4: Reference Specificity Analysis

Real emails from a human who engaged with the recipient's work will reference specific details. Real emails from an AI will reference the recipient's role or company but rarely go deeper.

A specificity score:

• Count proper nouns.
• Count numbers.
• Count references to specific dates or times.
• Count references to specific pieces of the recipient's public work (posts, papers, talks, products).

Score below a threshold across a well-written email of decent length is a strong AI signal. An LLM can be prompted to include specifics, but it will hallucinate them; a downstream verification pass can catch hallucinations.

Technique 5: Behavioral Signals

The above are content-based. Behavioral signals add another dimension:

• Response latency to your reply. AI-outreach pipelines often auto-follow-up at machine speed. If a first reply to your question arrives in 18 seconds, that is a signal.
• Uniformity across senders. Cold outreach with slightly-different-but-structurally-identical patterns arriving from multiple senders in a week suggests a common AI pipeline.
• Sender domain patterns. Recently registered domains, generic-looking domains, and domains with no historical email history score higher for AI outbound.
• Recipient list uniformity. You are one of many recipients whose emails share a common substring pattern (bulk send).

Behavioral signals are lower precision individually but stack well with content signals.

Assembling the Filter

A practical filter:

1. Rule out obvious human email first. Threads with prior human replies, contacts in your address book with historical thread depth, and senders inside your organization — all get a pass.

2. Score inbound cold email. Compute phrase signature, structural, contraction, specificity, and behavioral scores.

3. Combine into a tier. Very high AI likelihood → dedicated "AI outreach" folder. Ambiguous → priority inbox with an "AI-drafted?" tag. Low → main inbox.

4. Learn from your responses. When you reply to a message, weight the classifier toward "worth reading." When you archive without reading, weight the classifier away. This tunes the filter to your tastes.

A version of this filter runs in the Presend inbox layer today. It reduces cold-inbox reading load by 40-70% for typical users.

What About Detection Bypass Tools?

A cottage industry has emerged around "humanizing" LLM output — tools that take AI drafts and rewrite them to defeat detectors. Do they work?

Partially. They reliably degrade the phrase signature signal. They partially degrade the structural signal. They rarely fix the specificity gap. They cannot compensate for behavioral signals at all.

The upshot: an AI outreach that has been through a humanizer still scores as ~70% likely AI on a well-designed multi-signal detector, down from ~95% likely raw. Not zero, and not below the threshold that would categorize it as human.

The Ethics Layer

Two important caveats:

Detection is a signal, not a verdict. Some of the best cold outreach in 2026 uses AI to draft and a substantial human rewrite. Automatically deprioritizing all AI-scored email will lose real signal. Detection should route attention, not eliminate access.

Explainability matters. If your filter blocks a message and the sender is a real human who happened to write formally, you owe them a mechanism to be seen. A "flag for review" surface that lets senders explicitly signal "I am a real person, please read" — verified through some lightweight proof of humanity — is part of a mature detector deployment.

Building Your Own: A Weekend Project

If you want to build this yourself, a minimal implementation:

1. Corpus. Grab 200 emails you have marked as high-value and 200 you have deleted immediately. Label them.

2. Feature extractor. Write Python code that computes: phrase signature (from a starter list of 40 phrases), sentence-length CV, paragraph-length CV, tricolon count, contraction rate, specificity score.

3. Classifier. Train a logistic regression on the labeled data. Expect ROC-AUC around 0.80-0.85.

4. Deploy. Gmail add-on or Cloudflare Worker that scores inbound messages and applies a label.

5. Iterate. Every archive/reply action feeds back as a training signal. Retrain weekly.

This is a real product surface, and building it once teaches more about email intelligence than any amount of reading.

The Bigger Picture

Inbox-side AI detection is the natural counterweight to outbox-side AI drafting. As the drafting side becomes more sophisticated, the detection side has to be too. The equilibrium in 2027 will likely look like this: most cold outreach is drafted with AI assistance, most recipients use some form of AI-aware filtering, and the messages that break through are the ones with genuine specificity and stance — the ones a human bothered to write.

That is a healthier equilibrium than the current one, where undifferentiated AI slop clogs every inbox and gets low response rates. Detection makes drafting better because it raises the bar on what gets read.

Add Presend to Chrome — Free →