AI detectors for resumes, how recruiters spot ChatGPT, and how to write past them.

Token entropy on bullet-dense copy

Standard token-entropy scoring measures how predictable each word is given the words before it. Resumes are inherently low-entropy because the genre forces a small set of verbs (led, drove, owned, scaled, reduced) and a small set of nouns (team, pipeline, revenue, latency). A detector that does not know it is reading a resume will score that low entropy as machine output. Resume-mode scoring re-baselines against a CV corpus, so what counts as "expected" entropy shifts to match the genre.

Burstiness on parallel bullet construction

Burstiness measures how much the predictability varies across the document. Essay prose is bursty because long sentences sit next to short ones. Resume bullets are deliberately parallel: same verb position, same length range, same metric placement. A naive detector reads that uniformity as a smoothness signal. A calibrated detector treats parallel construction as a genre marker, not a fingerprint of generation.

Semantic uniformity across sections

Semantic uniformity flags documents where every paragraph orbits the same concept space. On a resume that is partly the point: every bullet is about your work. The signal that actually matters is whether the semantic content is specific (Snowflake, 14 microservices, Q3 2025) or generic (cross-functional collaboration, results-oriented leadership). Resume mode weights specificity, not topical breadth.

Action-verb-to-content-word ratio

Resume copy uses more verbs per word than narrative prose, and LLMs over-rotate that ratio because they were trained on generic resume samples. A detector that knows the genre expects the high verb ratio but flags it when the action verbs cluster around a narrow vocabulary (drove, leveraged, spearheaded, optimized) without specific context. TextSight resume mode reads the verb ratio against the noun specificity, not against an absolute target.

STEM resumes: narrow vocabulary, hard metrics

A senior backend engineer's resume reaches for the same 60-word vocabulary on every bullet: Python, Go, Kafka, Postgres, AWS, throughput, latency, p99, SLA, RPS. The structural payoff is clarity for the hiring manager. The detector cost is that the document looks lexically uniform compared to the average resume in a generic training set. TextSight resume mode re-baselines that vocabulary expectation against a 1,000-CV reference set heavy on software, data, and biotech roles, so the narrow vocabulary stops reading as a generation signal.

Recent-grad resumes: templated structure, parallel bullets

Career-services workshops teach a specific resume structure: name and contact, education, projects, technical skills, leadership. Bullets are coached into parallel form: action verb, deliverable, quantified outcome. That structure is good for human readers and bad for naive detectors, because the parallelism reads as machine-output. The calibration fix is the same as STEM: teach the model that parallel bullet construction is a genre marker, not a fingerprint of generation.

How TextSight handles it

Resume mode swaps in a CV-tuned classifier head, applies a section-aware tokenizer that treats bullets as their own unit, and re-baselines burstiness expectations against a corpus of 1,000 human-written resumes spanning STEM, business, design, and recent-grad cohorts. On our 100-resume benchmark, the STEM-resume false positive rate sits at 5.6 percent, against 18 to 24 percent for competitors that run essay-tuned scoring on the same documents.

Heavily hand-edited LLM output

If a candidate runs a ChatGPT draft through three or four passes of human editing (replacing generic verbs with specific ones, adding real tool names, swapping in real metrics from their actual work), the document becomes structurally indistinguishable from a hand-written resume. That is fine: the editing is the human contribution. Detection lives in the gap between unedited LLM output and a resume that reflects real specific work; we cannot, and should not, flag a document where the candidate has done that editing.

Voice-matched rewrites

The newer generation of LLM rewriters (TextSight's own Maximum mode included) train on the candidate's older writing samples to imitate voice. When the rewrite preserves cadence, sentence-length distribution, and vocabulary cluster, the structural signals we read collapse toward the human baseline. We catch some voice-matched rewrites; we miss the well-tuned ones. This is a known limit of statistical detection.

Resumes that were never AI to begin with

The honest scope statement that matters most: we are not in the business of guessing intent. If a recent grad writes 30 templated bullets across 12 internship descriptions, the document looks AI-likely to a naive detector and reads human to a calibrated one. We tune resume mode to weight specificity over structural smoothness so the templated-but-human case is treated correctly, but the trade is that the well-edited AI resume gets the same benefit. We err on the side of fewer false accusations.