Short-input tuning
Calibrated for 100–500 word inputs, where generic essay detectors lose signal.
Content-type · Email
AI email detector, tuned for short-form correspondence.
Emails are short, which is hard for generic detectors. This one is calibrated for the length and for the specific corporate-GPT closing patterns that survive into sent mail.
- Short-input tuning (starts working at 100 words)
- Common closer detection
- Free, 5,000 characters, no account
About the Email detector
Emails are a different detection problem. They're short, genre-constrained (greeting, body, closer), and often semi-templated by the writer long before any AI was involved. Generic detectors that need 300+ words to produce a reliable score simply do not work on most emails. This detector is calibrated for the length.
What this detector looks for
- Closer tics. "I hope this email finds you well," "Please let me know if you have any questions," "Looking forward to your response." These survive into sent AI-assisted email at remarkable rates. Weighted heavily because they're high-signal in short inputs.
- Unnaturally smooth transitions. Emails tend to be drafted quickly, so the transitions between sentences are often abrupt in genuine human writing. AI-drafted emails smooth them.
- Over-formal greeting. "Dear Mr. [Lastname]" when the context (reply chain, shared project) would normally call for first-name greeting.
- Paragraph structure. AI-drafted emails default to 3-paragraph structure (intro, ask, close). Genuine human emails vary more, often no greeting, often one long paragraph.
Known failure modes
- Sales and template-heavy email workflows. People who write a lot of semi-templated outreach produce prose that scores higher than average. Not a verdict.
- Non-native English senders. Same failure mode documented for all AI detectors. Weighting in emails is particularly brittle because of the short length.
- Under 100 words. Below 100 words we flag low confidence rather than produce a score. Two-sentence emails cannot be classified reliably.
What this is and isn't useful for
Useful: checking whether long-form business correspondence is AI-drafted; spotting cold-outreach sequences; verifying that a hiring-committee reference letter is hand-written. Not useful: deciding whether a two-sentence reply is human; adjudicating workplace conflicts based on detector scores.
As with every detector on the site: the score is a signal, not a verdict. Use it to start a conversation, not to end one.
Closer-pattern detection
High-signal business-email closers are weighted appropriately instead of being treated as generic formality.
Confidence floor
Under 100 words, we report low confidence instead of fabricating precision.
Short-form detection
Short messages are the hardest, here's what we deliver.
0.86
AUC on email
Lower than essay due to short length, but useful.
40–300 words
Tuned range
The typical work-email length where signal is reliable.
<1s
Median latency
Fast enough to score email drafts in-line.
Low confidence
Marked clearly
We surface uncertainty explicitly under 80 words.
Frequently asked questions
Why not just use the essay detector?
Generic essay detectors need more signal than short emails provide. The email detector trades peak precision for working at short lengths, the right trade for this content type.
Can I check a whole thread?
Paste it in. The detector scores the combined text. If you want per-message scoring, run each message separately.
What about reference letters?
Reference letters are a useful case, they're typically 300+ words and formal enough that generic detectors misfire. The email detector's short-input tuning helps, but for >500 word letters the essay detector is also appropriate.
Is this private?
Yes. Pasted emails are stored 30 days for re-checks, then deleted, and never used to train the detector. See /privacy for the full policy.
Ready to check email?
Free up to 3,000 characters. No account required for a single check.
Run the detector