AI detectors for job applications and recruiters

AIDI-U-04

Updated ·

An AI detector for job applications is judged on one field before any other: the published false positive rate, because a false positive here is a candidate rejected for writing their own cover letter. Two fields decide which cards appear in this view: a false positive figure in rank_fields.published_fpr that comes from a study in the index rather than from the vendor, and a self-serve plan or free tier in free_tier, since a recruiter buys for a team, not for a campus. Of the 72 cards, 23 carry a study-sourced false positive figure, 41 state a self-serve free tier, 46 state at least one of the two, and 15 state both. Fourteen of those 15 cards make up this view, GPTKit, PlagiarismCheck and Fast-DetectGPT having joined on 10 November 2026 with the ninth wave and Binoculars later the same day with the tenth (0.00% at threshold 0.5 in RAID 2024, free open-source code): BrandWell joined with the fourth wave of cards on 9 November 2026 with a 0 of 18 from Weber-Wulff 2023 and a free check inside the Moxby extension, and DetectGPT, which also carries a 0 of 18 and free MIT-licensed code, stays out because no hosted tool exists for a recruiter to run (DetectGPT card). Writer, added on 9 November 2026, carries a study-sourced rate of 1 in 18 documents (5.6%) from Weber-Wulff 2023 and 14% of 42 papers from Walters 2023, but the tool was withdrawn on 22 December 2025 and no free tier exists at the read date, so a recruiter cannot buy it and the card stays out (archived writer.com page, read 9 November 2026). This page is one of five under AI detectors by use case; the school districts view reads the same false positive field for students, the business view adds team seats and retention, and you can start from the full table to see all 72.

  1. Field 1Published false positive rate
  2. Field 2Base and study
  3. Field 3Free tier as stated
  4. Field 4Languages

The false positive figure comes first and the table names the study and the base behind each one, because a rate of 0 in 18 documents and a rate of 0 in 42 are different claims. The false positive rate ranking orders every card on the same field and explains why the figures differ between studies. The free tier is quoted as stated, because screening a hundred applications is a volume question. Minimum input length and the retention policy for candidate text appear in the notes. Field rules are in the methodology.

Which fields decide fit for a recruiter?

AI detectors for job applications and recruiters: the cards in this view and the fields that decide fit
DetectorPublished false positive rateBase and studyFree tier as statedLanguages
Scribbr2%42 human-written first-year composition papers, Walters 2023Unlimited free AI checks, no sign-up required1
Winston AI00 false accusations in 18 human-written documents, Weber-Wulff 20232,000 credits / 14 day trial11
GPTZero0.00%human-written documents at naive threshold 0.5, RAID 2024Free plan covering 10,000 words per month5
Copyleaks00 of 42 human-written papers, Walters 2023up to 25,000 characters per scan30
Sapling00 of 42 human-written papers (31% returned uncertain), Walters 2023up to 2,000 characters per check16
Originality.ai00 of 42 human-written papers (2 returned uncertain), Walters 202310K words (100 credits one time)30
ZeroGPT00 false accusations in 18 human-written documents, Weber-Wulff 202315,000 characters per detection, words allowance 1,250,000 /monthstated as all languages
GPT-2 Output Detector00 false accusations in 10 human samples, Perkins 2024Free: MIT-licensed weights and code run locally; the hosted demo is pausednone stated
GLTR99.3%RAID human-written documents at naive threshold 0.5 (0.05% at 0.95), RAID 2024Free and open source: public demo at demo.gltr.io, no accountnone stated
BrandWell00 false accusations in 18 human-written documents, tested as Content at Scale, Weber-Wulff 2023; 10% of 42 human papers called AI, Walters 2023Free scan of 2,500 characters on the archived page; the Moxby Mod the page now resolves to needs the Moxby Free plan, 10,000 characters per checkEnglish only
GPTKit0 of 114114 human-written English computing submissions, Orenstrakh 2023First 2048 characters free per request for guests1
PlagiarismCheck0 of 1818 human-written documents, 9 machine-translated, Weber-Wulff 2023A short text free to see how it works; 2 pages free on the about page2
Fast-DetectGPT23.2%RAID human-written documents at threshold 0.5, RAID 2024Free and open source (MIT); the hosted demo gives guest trials and free credits on registration
Binoculars0.00%RAID human-written documents at threshold 0.5, RAID 2024Free and open source (BSD 3-Clause); the Hugging Face Space demo publishes no quota

Values as each study or vendor states them, read 8 November 2026; the GPT-2 Output Detector and GLTR rows, two free research tools, were read 9 November 2026.

The false positive figures come from Walters 2023 Table 4, Weber-Wulff 2023 Table 9, Dugan et al. 2024 (RAID) Table 4 and Perkins 2024 Table 10. The two free research tools are described on the GPT-2 output dataset repository and the GLTR repository, read 9 November 2026. Free tiers and language counts come from scribbr.com, gowinston.ai/pricing, the GPTZero blog dated October 25, 2026, copyleaks.com, sapling.ai/pricing, originality.ai/pricing and zerogpt.com/pricing, all read 8 November 2026.

AI detectors for job applications and recruiters: register plate with the page title and its register ID
The view plate: register ID and title, drawn from the index data.

Why each card is in this view

  1. AIDI-D-0009Scribbr

    Walters 2023 classified 95% of 42 human papers correctly with 2% incorrect (publisher PDF). The detector page says "we do not store or share your data" and offers unlimited free checks (scribbr.com, read 8 November 2026).

  2. AIDI-D-0005Winston AI

    0 false accusations in 18 human-written documents in Weber-Wulff 2023, and RAID 2024 reports 0.55% at threshold 0.5 (arXiv). Scans take up to 200,000 characters on paid plans, and the privacy policy says submitted text and images are stored until the report or account is deleted (pricing, privacy policy, read 8 November 2026).

  3. AIDI-D-0001GPTZero

    0.00% at threshold 0.5 in RAID 2024, but 2% of 42 in Walters 2023, 1 of 10 in Perkins 2024, 1 of 5 in Elkhatat 2023 and 9 of 18 in Weber-Wulff 2023, the widest spread in the index. The widget takes 10,000 characters per scan without an account, and the FAQ says dashboard inputs are stored while API calls are not (FAQ, read 8 November 2026).

  4. AIDI-D-0003Copyleaks

    0 of 42 in Walters 2023, 1 of 5 human controls in Elkhatat 2023 and 5 of 10 in Perkins 2024 (arXiv). The web platform needs at least 255 characters (copyleaks.com, read 8 November 2026).

  5. AIDI-D-0010Sapling

    0 of 42 in Walters 2023 with 31% returned uncertain. The product page says results are more accurate after about 50 words and the API flags texts under 500 characters as low confidence; the free check covers the first 2,000 characters, and the API accepts zero_retention: true (sapling.ai, API docs, read 8 November 2026).

  6. AIDI-D-0004Originality.ai

    0 of 42 in Walters 2023 with 2 uncertain, and 0.25% at threshold 0.5 in RAID 2024. A 100-word minimum applies, and the privacy policy allows scan history to train models unless the user opts out (pricing, privacy policy, read 8 November 2026).

  7. AIDI-D-0011ZeroGPT

    0 false accusations in 18 in Weber-Wulff 2023 and 0 of 10 in Perkins 2024, 2% of 42 in Walters 2023, 1.42% at threshold 0.5 in RAID 2024. The pricing page recommends at least 150 to 200 words; the terms of use say a ZeroGPT output must not be "the sole basis for a decision that could materially affect another person's rights, education, employment, admission" (pricing, terms, read 8 November 2026).

  8. AIDI-D-0019GPT-2 Output Detector

    0 false accusations in 10 human samples in Perkins 2024, but 2 of 18 (11.1%) in Weber-Wulff 2023 and 2.91% (large model) to 6.59% (base model) at threshold 0.5 in RAID 2024 (arXiv 2403.19148, arXiv 2306.15666, arXiv 2405.07940). OpenAI released the RoBERTa weights under the MIT licence to run locally, the hosted Hugging Face demo is paused, and OpenAI's own release note puts the detection rate on GPT-2 text at about 95% and calls that "not high enough accuracy for standalone detection" (model card, GitHub, read 9 November 2026). The model was trained on GPT-2 output and predates the generators a candidate would use today.

  9. AIDI-D-0030BrandWell

    Weber-Wulff 2023 Table 9 records 0 false accusations in 18 human documents for the tool tested as Content at Scale, in the same test in which it missed 36 of 36 AI documents, and Walters 2023 Table 4 records 10% of 42 human papers called AI (arXiv 2306.15666, DOI 10.1515/opis-2022-0158, read 9 November 2026). The detector page has redirected to a Moxby marketplace Mod since 29 August 2026, so a recruiter runs it inside that browser extension, and the listing itself says "human-written text can be flagged, so results are not proof of authorship" (moxby.com, read 9 November 2026).

  10. AIDI-D-0017GLTR

    RAID 2024 Table 4 records a false positive rate of 99.3% on human-written documents at the repository's default threshold of 0.5, falling to 0.05% at 0.95, and Chaka 2023 found it classified 1 of 21 chatbot responses as machine-generated (arXiv 2405.07940, DOI 10.37074/jalt.2023.6.2.12). GLTR is a free, open-source visual aid from 2019 that colours each word by its rank under GPT-2, with a public demo and no account (gltr.io, GitHub, read 9 November 2026); it returns no verdict of its own, and the 99.3% figure shows what a naive threshold on its output does to human text.

  11. AIDI-D-0045GPTKit

    0 of 114 in Orenstrakh 2023, the only detector of eight at zero (arXiv 2307.07411, Table 8, read 9 November 2026), and 4.5% accuracy on manipulated AI text in Perkins 2024, so a cover letter run through a paraphraser passes; content is removed from the server after detection per the FAQ.

  12. AIDI-D-0046PlagiarismCheck

    0 of 18 in Weber-Wulff 2023 with 19 of 36 AI documents undetected (arXiv 2306.15666, Tables 9 and 10, read 9 November 2026); the privacy policy keeps checked documents until deleted or a year of inactivity for individual users.

  13. AIDI-D-0049Fast-DetectGPT

    23.2% of human documents flagged at the naive 0.5 threshold in RAID 2024, about one application in four at that setting (arXiv 2405.07940, Table 4, read 10 November 2026); a recruiter running the open code locally sends no text anywhere.

  14. AIDI-D-0053Binoculars

    0.00% of RAID's human-written documents at threshold 0.5 (arXiv 2405.07940, Table 4, read 10 November 2026), free to run from the repository or the hosted Space; DAMAGE 2025 prints 5.40% of human student essays flagged at the default threshold with no essay count, and 28.23% of humanized AI essays caught, so a cover letter rewritten through a humanizer will mostly pass (arXiv 2501.03437, Tables 3 and 7). Check For AI and Writefull, added the same day with a 1 of 18 each from Weber-Wulff 2023, stay out as Writer does: both tools are withdrawn and no free tier exists at the read date.

What the evidence says for recruiters

The study written for this reader is Liang 2023. Seven detectors scored 91 human-written TOEFL essays and 88 US 8th-grade essays; the paper reports an average false positive rate of 61.22% on the TOEFL set against near-perfect accuracy on the 8th-grade set, and 18 of the 91 TOEFL essays were flagged as AI by all seven detectors. GPTZero, Sapling, Originality.ai and ZeroGPT were among the seven tested (Liang 2023, arXiv, read 8 November 2026). A candidate who learned English as a second language and writes carefully is the profile that study flagged.

The 0 figures in the table come from small samples of a single text type. Walters 2023 scored 42 human-written first-year composition papers, and Weber-Wulff 2023 scored 18 human-written documents, 9 of them machine-translated (Walters 2023, Weber-Wulff 2023, read 8 November 2026). Perkins 2024 scored 10 human samples and recorded 5 false accusations for Copyleaks, 1 for GPTZero and 0 for ZeroGPT (Perkins 2024, arXiv, read 8 November 2026). Elkhatat 2023 scored 5 human controls and recorded 1 false positive each for GPTZero and Copyleaks (Elkhatat 2023, Springer PDF, read 8 November 2026). None of the nine study records scored cover letters or job applications.

Sapling's own caveat describes the cover letter exactly: "The shorter the text is, the more general it is, and the more essay-like it is, the more likely it is to result in a false positive". The sentence is on the Sapling product page, read 8 November 2026. No detector score should be the only evidence in a misconduct or hiring decision.

For voice rather than text, the Pindrop vs Resemble Detect comparison sets the two vendors that name hiring and meeting integrations on the same fields, read 10 November 2026.

Sources

  • https://www.scribbr.com/ai-detector/ (read 8 November 2026)
  • https://gowinston.ai/pricing/ (read 8 November 2026)
  • https://gowinston.ai/privacy-policy/ (read 8 November 2026)
  • https://gptzero.me/faq (read 8 November 2026)
  • https://gptzero.me/news/best-ai-detectors (read 8 November 2026)
  • https://copyleaks.com/ai-content-detector (read 8 November 2026)
  • https://sapling.ai/ai-content-detector (read 8 November 2026)
  • https://sapling.ai/pricing (read 8 November 2026)
  • https://sapling.ai/docs/api/detector/ (read 8 November 2026)
  • https://originality.ai/pricing (read 8 November 2026)
  • https://originality.ai/privacy-policy (read 8 November 2026)
  • https://www.zerogpt.com/pricing (read 8 November 2026)
  • https://www.zerogpt.com/terms-of-use (read 8 November 2026)
  • https://www.degruyterbrill.com/document/doi/10.1515/opis-2022-0158/pdf (read 8 November 2026)
  • https://arxiv.org/pdf/2306.15666 (read 8 November 2026)
  • https://arxiv.org/pdf/2405.07940 (read 8 November 2026)
  • https://arxiv.org/pdf/2304.02819v3 (read 8 November 2026)
  • https://arxiv.org/pdf/2403.19148 (read 8 November 2026)
  • https://doi.org/10.37074/jalt.2023.6.2.12 (read 9 November 2026)
  • https://huggingface.co/openai-community/roberta-base-openai-detector (read 9 November 2026)
  • https://github.com/openai/gpt-2-output-dataset/tree/master/detector (read 9 November 2026)
  • http://gltr.io/ (read 9 November 2026)
  • https://github.com/HendrikStrobelt/detecting-fake-text (read 9 November 2026)
  • https://link.springer.com/content/pdf/10.1007/s40979-023-00140-5.pdf (read 8 November 2026)

Revision log: 2026-11: view created2026-11-09: GPT-2 Output Detector and GLTR rows added; the counts in the opening paragraph now follow the index data file at build time2026-11-09: six wave 3 cards checked; Writer noted as a withdrawn tool with a study-sourced rate and no free tier, compound counts 24 and 26; 2026-11-09: ten wave 4 cards checked, BrandWell added, DetectGPT noted as a study-rated open-source method with no hosted tool, compound counts 32, 35 and 112026-11-10: GPTKit, PlagiarismCheck and Fast-DetectGPT added (ninth wave), compound counts 35, 38 and 142026-11-10: Binoculars added (tenth wave), compound counts 41, 46 and 15; Check For AI and Writefull noted as withdrawn tools with a study-sourced rate and no free tier2026-11-10: twelfth-wave pages linked

FAQ: AI detectors for job applications

Which AI detectors fit screening job applications?

Nine cards carry a study-sourced false positive figure and a self-serve plan or free tier: Scribbr, Winston AI, GPTZero, Copyleaks, Sapling, Originality.ai and ZeroGPT, as read on 8 November 2026, and two free open-source research tools read on 9 November 2026, the GPT-2 Output Detector and GLTR, whose rows show how far one detector's rate can move with the threshold. The table shows each figure with its base so a recruiter can see how small the human samples were before treating any rate as a promise.

Can an AI detector tell if a cover letter was written with AI?

A detector returns a probability, and on short general text the vendors themselves say the probability is weakest. Sapling flags results under 500 characters as low confidence, Copyleaks needs 255 characters, Originality.ai needs 100 words and ZeroGPT recommends 150 to 200 words, all read 8 November 2026. A typical cover letter sits near those floors.

What false positive rate should HR expect?

The published range for one detector can run from 0 to 50% depending on the sample: GPTZero recorded 0.00% at threshold 0.5 in RAID 2024 and 9 false accusations in 18 documents in Weber-Wulff 2023, and GLTR recorded 99.3% at threshold 0.5 and 0.05% at 0.95 in the same RAID 2024 table. Expect the rate on your applicant pool to differ from every published figure, and measure before relying on one.

Are non-native English applicants flagged more often?

Liang 2023 found an average false positive rate of 61.22% on 91 TOEFL essays across seven detectors against near-perfect accuracy on 88 US 8th-grade essays, with 18 essays flagged by all seven (arXiv, read 8 November 2026). The study tested 2023 versions; vendors have since published claims of improvement, recorded on each card.

Is a cover letter long enough to score?

Sometimes. Copyleaks accepts 255 characters, Sapling scores from about 50 words with low confidence under 500 characters, Originality.ai requires 100 words and ZeroGPT recommends 150 to 200 words, all as stated on 8 November 2026. A 300-word letter clears every stated minimum in this view; a 120-word note does not clear them all.

Can a detector score be grounds to reject a candidate?

No, and the vendors say so. ZeroGPT's terms bar using an output as the sole basis for a decision affecting employment, and Sapling says no current detector should be used as a standalone check, both read 8 November 2026. A score is a reason to ask for a writing sample or a conversation, not a result.

Related reading