Cover letters now carry a second AI-likelihood pill, separate from the resume, because the recruiter view scores narrative prose differently from bullet copy. Generic LLM cover letters score 70 to 90% AI-likely in two seconds. TextSight reads them with the same calibration tuned for hiring-context prose. To actually score your draft, head to our scoring workflow. To understand the detection landscape, keep reading.
A resume and a cover letter live in the same application, but they get judged by two different models because the prose itself is structurally different.
Inside Workday, Greenhouse, Lever, and iCIMS, each document is stored as its own object with its own metadata. The AI-likelihood pipeline scores the resume file and the cover-letter file independently, and the recruiter dashboard surfaces both pills side by side. A clean resume can sit next to a 90% AI-flagged cover letter inside the same application.
A resume is structurally compressed: bullets, action verbs, metrics, almost no connective tissue. A cover letter is narrative prose, 250 to 400 words built around an opener, a body, and a close. The two formats produce different statistical fingerprints, so ATS vendors run two calibration profiles. The pill you see is whichever profile matched the file type.
In a typical 2026 ATS candidate view, the AI-likelihood pills float next to each attachment thumbnail. Two pills, two judgements, formed before a single sentence of yours has been read. The cover-letter score sometimes carries more weight than people realise: it is the only piece of writing in the application that signals voice, and a high pill on the only voice signal is a soft red flag.
Narrative-prose detection on hiring-context letters comes down to four signals. None of them on their own is enough. The combined score is what produces the pill.
"I am writing to express my interest in the role of", "I am excited to apply", "With my background in X, I believe I would be a strong fit". These phrases sit in millions of training-set documents and detectors learn them as high-confidence AI markers. The first sentence of a generic ChatGPT cover letter often contributes 15 to 25 points of the final score on its own.
LLMs produce three to four paragraphs of roughly equal length. Humans vary widely: one-line opener, long middle, short close. Detectors weight low paragraph-length variance as suspicious. This is one of the easiest signals to fix on revision because it needs restructuring, not rewriting.
"Leverage", "spearhead", "cutting-edge", "passionate", "results-driven", "synergy", "value-add". Real candidates use one or two of these. LLM letters use six in tight co-occurrence. Detectors track that density per 100 words; two or more pushes the score upward sharply.
The same signal TextSight uses everywhere else. Human cover letters have rhythm: a punchy 8-word sentence, a long 28-word reflection, a 12-word transition. AI rhythm clusters around 18 to 22 words per sentence with low variance. The narrative-prose calibration measures variance against a reference set of human-written cover letters specifically.
Five tools you will encounter when researching cover-letter AI detection. Two lines on each, what they do well, where they fall short on this specific format.
Writer-first detector with sentence-level highlights and a separate narrative-prose calibration profile tuned on a reference set of cover letters. The free tier needs no signup, scans up to 5,000 characters, and returns per-sentence rationale. Best fit when you want to see which lines of your draft are dragging the score up.
The classroom mindshare leader. Uses perplexity and burstiness scoring, which works well on raw GPT output and less well on lightly edited drafts. No cover-letter-specific calibration, so it tends to over-flag formal hiring-context language from polite or ESL writers. Brand recognition is the strongest reason to use it inside institutions that already standardised on it.
SEO-agency favourite with bundled plagiarism detection. The cover-letter score lands close to the resume score because the underlying calibration treats them as a single class of "professional writing". Good if you are already paying for plagiarism scanning, less ideal as a dedicated cover-letter checker.
Education and enterprise focused. The cover-letter detection performance is respectable but the UI is built for batch academic submissions, not single-letter pre-application checks. Pricing is per-document, which gets expensive if you are iterating on a draft.
The institutional default for academic submissions; not really designed for hiring context. The AI score is rolled into the existing plagiarism workflow, which means most candidates never see it directly. Listed here because some recruiters with academic backgrounds still default to it as a familiar reference.
Compared on the features that change how a cover-letter pre-check actually feels. Win markers reflect our reading of the gap, not a third-party audit.
| Feature | TextSight | GPTZero | Originality | Copyleaks | Turnitin |
|---|---|---|---|---|---|
| Cover-letter-specific calibration | Yes | No | No | Partial | No |
| Narrative-prose mode | Yes | Generic | Generic | Generic | Generic |
| Opener-cliché detection | Yes, weighted signal | Indirect via perplexity | Indirect | Indirect | Indirect |
| ESL fairness on hiring-context prose | ~5.2% FPR | ~17% FPR | ~14% FPR | ~13% FPR | ~19% FPR |
| Recruiter-context tuning | Yes, hiring reference set | No | No | No | No |
| Results in under 2 seconds | Yes | Yes | Yes | Yes | Variable |
| Sentence-level highlights | Yes, per-line evidence | Document score plus segments | Document score | Sentence highlights | Highlighted spans |
| Free tier (no card) | 3 scans/day, 5,000 chars, no signup | Monthly word cap, signup | Trial credits only | Limited free preview | Institutional licence |
| 100k+ characters per scan | Yes on paid tiers | Yes on paid tiers | Yes | Yes | Institutional |
| Paste workflow | One-click paste, no setup | One-click paste | One-click paste | Paste or upload | LMS upload only |
| Paraphrase rewriter included | Yes, bundled at every paid tier | Separate product | Separate add-on | Separate product | Not included |
| ATS-safe rewriter mode | Yes, narrative-prose preset | No preset | Generic rewriter | Generic rewriter | No |
| Voice-match mode | Yes, paste reference + rewrite | No | Partial via styles | No | No |
| Browser extension | Free on all tiers | Free on all tiers | Paid tier only | Paid tier only | No |
| Last verified date published | Not published | Not published | Not published | Not published |
Features, prices, and FPR numbers reflect our internal testing as of . Verify on each tool's page before subscribing. ESL FPR rows are from the cover-letter subset of our 100-passage benchmark; see #benchmark below for methodology.
A 30-letter subset of our 100-passage benchmark, scoped to cover letters only. 15 ChatGPT-generated against the same job descriptions, 15 human-written (mix of native English and ESL applicants).
| Detector | TPR (AI letters caught) | FPR (human letters wrongly flagged) | Δ vs TextSight FPR |
|---|---|---|---|
| TextSight (cover-letter mode) | 89% | 5.2% | baseline |
| GPTZero | 84% | 17% | +11.8pp worse |
| Originality.ai | 82% | 14% | +8.8pp worse |
| Copyleaks | 79% | 13% | +7.8pp worse |
| Turnitin | 76% | 19% | +13.8pp worse |
| Competitor range | 76 to 84% TPR | 13 to 19% FPR | +7.8 to +13.8pp worse |
The most useful thing we can do on this page is mark the edge of the map. Here is what the score does not measure.
A detector cannot read whether you actually want the job. A human letter can be deeply insincere; a partly LLM-assisted letter can come from the right hire. The score measures surface form, not intent. That is why most ATS configurations surface the pill as metadata, not as a verdict.
The detector does not know the job description. It cannot tell whether your projects map onto what the company needs, whether your seniority is correct, whether your tone matches their voice. Those judgements sit downstream of the pill. A 5% AI letter that is wrong for the role does worse than a 35% AI letter that is right.
If you used ChatGPT to generate a draft and then rewrote 60% of the prose in your own voice, the result is statistically indistinguishable from a human-written letter on every detector we tested. The surface form is yours. Detection works against verbatim and lightly edited AI; against carefully voice-matched rewrites, no detector has a reliable signal.
The harder edge case. A small number of humanizer tools now produce output that scores under 10% AI across the entire detector market, including ours. The countermove on the recruiter side is a written interview or a take-home assignment, not a better detector. Where TextSight helps is on the much larger category of generic ChatGPT letters, which still represent the bulk of what ATS pipelines flag in 2026.
The actual paste-and-score workflow with sentence-level highlights and score bands.
Score now →The sibling explainer for the resume side of the application. Same ATS context, different calibration.
Read the resume guide →Once the score is too high, the rewriter path: targeted edits that bring the AI pill back down.
Read the rewrite guide →Full benchmark methodology, sample composition, scoring thresholds, and re-test cadence.
Read the methodology →Head-to-head comparison on detection accuracy, free tier, ESL fairness, and pricing.
Read the compare →Head to the cover-letter scoring workflow. Paste, see sentence-level highlights, revise in place. Free tier, no card, no signup.