The uncomfortable truth about email length

A typical business email runs 50 to 150 words — and below roughly 150 words, no statistical detector on the market can produce a reliable verdict. Cobalynx included. The signal simply isn't there: statistical detection works by accumulating small regularities across many sentences, and a five-sentence email doesn't contain enough of them. Tools that hand you a confident percentage for a two-line reply are guessing with a straight face. We show “too short” instead, because a fake verdict is worse than no verdict.

What works at short length: context, not style

For a single short email, behavioral and contextual signals beat any scan:

  • Shared-context check. Does the message reference anything only your real correspondent would know — the meeting you both attended, the nickname, the running joke? Generic AI output is context-free.
  • Template smell. Perfectly balanced paragraphs, a formulaic opener (“I hope this email finds you well”), a closing that restates the opener — the shape of text produced without a specific reader in mind.
  • Generic urgency. Pressure to act fast on something strangely unspecific is a phishing signal regardless of who — or what — wrote it.
  • Vocabulary tells, held loosely. Words like “delve” or “moreover” circulate on giveaway lists, but humans use them too. One word proves nothing; the aggregate pattern across a long text is what a statistical scan actually measures.

When a real scan becomes meaningful

Length changes everything. A long email, a multi-message thread pasted together, a newsletter, or a batch of messages from the same sender can cross the ~150-word floor where a calibrated scan starts to mean something. Then you get what short emails can't give you: a probability with a published error rate, and an honest “inconclusive” when the evidence is thin. Batching matters for pattern spotting too — five suspiciously similar outreach emails pasted as one text reveal the sameness a single message hides. The same batch method is the core of checking marketplace reviews and seller messages, where single texts are just as short.

The security angle: polish is no longer a trust signal

For years, broken English was the folk test for scam email. AI ended that: phishing messages are now routinely fluent, warm, and typo-free. This cuts in an unexpected direction — “was this email written by AI?” matters less for safety than “is this email asking me to do something I should verify?” Check the sender's actual address, verify unusual requests through a channel you already trust, and treat fluency as exactly zero evidence of legitimacy. An AI-flagged email isn't necessarily malicious either; your colleagues use AI assistance every day for perfectly honest mail.

If you're the sender wondering who can tell

Mostly, nobody can — from a single message. Human intuition about AI text measures near chance in published tests, and short messages are below every detector's floor. Longer texts and repeated patterns across many messages are more detectable. If a recruiter or client might scan your longer documents — cover letters, proposals — the honest questions become disclosure and quality, not detectability; both sides of that table are covered in the cover-letter guide, and the workplace-accusation scenario, if it ever lands on you, in the false-accusation playbook.

What Cobalynx can and can't do here

A 60-word email is below any detector's reliable floor, including ours — we'll tell you that instead of inventing a verdict, and the checklist above is the honest alternative at that length. For longer emails and batched threads we return a calibrated probability with published error rates, computed by a documented method — and even then, trust signals (context, specifics, verified sender) matter more than style ever will. More email and workplace questions are in the FAQ.

Common questions

Can I check a 60-word email?

Not meaningfully — 60 words is below any detector’s reliable floor, and Cobalynx will say “too short” rather than invent a verdict. Batch the whole thread, or several messages from the same sender, into one paste to reach the roughly 150 words a statistical check needs.

What words give away ChatGPT in an email?

Lists like “delve,” “I hope this email finds you well,” and “I trust this message reaches you” circulate for a reason, but any single word is weak evidence — humans use them all. Word choice only becomes meaningful in statistical aggregate across a whole text, which is precisely what a calibrated scan measures (see how our detector works).

Are AI-written emails a security risk?

The real shift is that well-written no longer means safe: AI-polished phishing removed the broken-English tell people relied on for years. Judge emails by verified sender, expected context, and what they ask you to do — never by fluency.

Can people tell if I use AI to write my emails?

Rarely from a single message — style intuition performs near chance in published tests, and short texts are below every detector’s floor. Longer messages and repeated patterns across many emails are far more detectable. The more durable answer is disclosure norms at your workplace, not stealth.