Executive summary

AI content detectors are useful for triage. None of the nine products huby evaluated is reliable enough to be the only evidence behind an academic, employment or contractual decision. Overall scores run from 2.5 to 3.7 on huby’s 5-point scale. No vendor reaches the 4.0 "Strong" band, and the cohort median is 2.9.

This paper compares Copyleaks, GPTZero, Originality AI, Pangram, QuillBot, Sapling, Undetectable.ai, Winston AI and ZeroGPT. It draws on huby’s latest full evaluation reports (October 27, 2026; November 7, 2026 for Pangram and Sapling), on huby’s controlled detection tests and on the public research collected on each product’s huby page.

Key findings

  • The leaders win on platform maturity more than on detection. Copyleaks (3.7) leads five of six categories (sharing first place in Privacy with GPTZero), mainly on enterprise deployment, LMS integration and security attestations. GPTZero (3.5) is second, on business viability and integrations. Pangram scores 2.9: it has one of the strongest validation records and ranked AI text perfectly in huby’s tests, but its privacy, security and enterprise-administration documentation is thin.
  • Detectors avoid false accusations at the cost of missing AI text. At each vendor’s default threshold, eight of nine products flagged no human-written document in huby’s tests. Recall ranged from 0.00 to 0.79, so most tools missed a large share of the documents that contained AI text.
  • Simple manipulation defeats most products. A character-level perturbation cut detection from 100% to 0% for four of the eight products where it could be measured. The median drop was 83 percentage points. Six of nine products detected none of the AI-paraphrased documents they were given.
  • Confidence scores are not probabilities. Expected Calibration Error ranged from 0.17 to 0.78, with a median of 0.42. A document scored "80% AI" is not 80% likely to be AI-written.
  • Mixed human-AI documents are the hardest case. When estimating the AI share of hybrid text, products were off by 8.0 to 82.1 percentage points on average, with a median error of 30.6.
  • Marketing outruns evidence. Seven of nine vendors advertise accuracy between 98% and 99.98%. None makes available independent audits that support those figures for edited, paraphrased or mixed-authorship text.
  • Four vendors sell evasion alongside detection. QuillBot, Sapling and ZeroGPT offer "humanizer" tools, and evasion is Undetectable.ai’s core product. This is a structural conflict of interest that buyers need to weigh.
  • Few vendors prove their security and privacy posture. Security is the weakest category overall, with a median of 2.7. Most vendors claim SOC 2 or ISO 27001 but do not publish the reports. For seven of nine vendors, public documentation is contradictory or unclear about whether submitted content trains their models.

Bottom-line recommendations

  1. Use detection for triage only. Before any consequence, require documented human review, corroborating evidence such as drafts or version history, and an appeal path.
  2. Shortlist by use case, not by headline score.
    • Institutions that need LMS, SSO, SCIM and data residency should start with Copyleaks.
    • Buyers who put published research and a low false-positive rate first should evaluate Pangram, and get its privacy, security and administration terms in writing.
    • Education teams that want a mature integration ecosystem should evaluate GPTZero, but should validate its recall and calibration on their own content first.
  3. Pilot on your own content before you buy. Include edited, paraphrased and mixed documents, and writing by non-native English speakers.
  4. Treat scores as rankings, not probabilities. Set thresholds locally, and don’t present raw percentages to students or employees as verdicts.
  5. Weigh conflicts of interest. Keep evasion-first vendors out of integrity programs. Ask suite vendors that also sell humanizers how their detector is kept separate from them.
  6. Contract for evidence. Ask for SOC 2 reports, a data processing agreement, written commitments not to train on your content, and retention schedules. Most vendors do not publish these.

Market definition and dynamics

AI content detection has split into five business models. The model a vendor runs shapes its incentives more than its feature list does.

What an AI content detector is

An AI content detector estimates whether a piece of text was written by a large language model rather than a person. All nine products in this cohort use statistical classifiers that do not depend on watermarks; none relies on provenance signals embedded by the generating model. Most return a document-level score with sentence-level highlighting. Many bundle adjacent capabilities, including plagiarism checking, image and deepfake detection, fact-checking and writing assistance.

Five business models

Segment Vendors Primary revenue What it means for buyers
Integrity platform Copyleaks Institutional licenses, credits, API Deepest administration, LMS integration and deployment options
Specialist detector Pangram, GPTZero, Originality AI, Winston AI Subscriptions and API usage Detection is the product; enterprise depth varies widely
Writing suite with a detector QuillBot, Sapling Writing-assistant subscriptions Detection is a feature; both also sell humanizer tools
Consumer freemium detector ZeroGPT Advertising and subscriptions Low friction, low assurance; also sells a humanizer
Evasion-first Undetectable.ai Humanizer subscriptions Detection supports evasion; unsuitable for integrity programs

Dynamic 1: an arms race with a commercial counter-market

A market of "humanizer" services exists to defeat detectors, and these services name their targets. huby’s evaluations document tools marketed against Copyleaks, GPTZero, Winston AI, Sapling and ZeroGPT. In one documented demonstration, ZeroGPT’s own humanizer cut its detector’s reading from 70.1% AI to 16.6%, flipping the verdict to human-written. Evasion does not require sophisticated tools: homoglyph substitution and AI paraphrasing defeated most products in huby’s tests (see Comparative analysis).

Dynamic 2: the false-positive constraint

A false accusation costs more than a missed detection, so vendors set conservative default thresholds. In huby’s tests, eight of nine products had perfect precision, but recall was low. Real-world false positives still occur. Examples documented in huby’s evaluations include:

  • A 1,252-word human-written article scored "100% AI" by Copyleaks.
  • Human nonfiction scored 99% AI by Originality AI and 96% AI by Winston AI.
  • The U.S. Constitution reportedly scored 92.15% AI by ZeroGPT.

Sapling is the exception in huby’s tests: it flagged 3 of 16 human documents (18.75%), one of them at 98.9% AI.

Dynamic 3: fairness for non-native English writers

Peer-reviewed research cited in huby’s evaluations finds disproportionate false positives on writing by non-native English speakers for Copyleaks, GPTZero, Sapling and ZeroGPT. User testing reports a similar pattern for Originality AI, where 18% of paragraphs by a non-native writer were flagged. Pangram publishes its own analysis of accuracy on ESL writing, but no vendor publishes an independent bias audit broken down by language group. huby’s own test corpus did not include non-native writing, so huby’s test results cannot settle this question.

Dynamic 4: claims versus evidence

Seven vendors advertise accuracy between 98% and 99.98%. The independent figures cited in huby’s evaluations are lower and vary widely by content type:

  • One independent benchmark measured Winston AI at about 42%.
  • Studies place GPTZero between 80% and 91.2%.
  • ZeroGPT measured between 64% and 84%.
  • Copyleaks fell to 58.7% on expert-edited machine text.

Pangram is the most transparent vendor here: it has published an arXiv technical report, model cards and open-weight models. Even so, the third-party verification reports it cites were not available for review.

Dynamic 5: from text classification to process evidence

Vendors are adding evidence that does not rely on a single classifier score. GPTZero and Originality AI offer writing-replay features that reconstruct how a document was drafted. Originality AI and Winston AI add fact-checking, and Originality AI holds a U.S. patent for its fact-checking system. Copyleaks, Winston AI, Pangram and ZeroGPT extend detection to images or video. Process evidence is more defensible than a single score, and huby expects it to become a standard procurement requirement.

Dynamic 6: consolidation and vendor risk

Ownership is consolidating at the top while the long tail stays thin:

  • Larger owners: GPTZero has been acquired by Superhuman, and QuillBot is owned by Learneo.
  • Funded independents: Pangram has raised $13 million and Copyleaks more than $7 million.
  • Small, founder-led firms: Winston AI, ZeroGPT and Undetectable.ai disclose little funding or governance. Third-party estimates put Winston AI at about three employees.

For multi-year institutional contracts, vendor viability matters as much as detection accuracy.

Evaluation methodology

All nine products were scored under one framework, huby’s published methodology (version 1.2), using the same factors and the same tests. Evidence was collected between August 30 and September 21, 2026. The evaluation reports were first produced on October 1, 2026 and revised on October 27, 2026; the Pangram and Sapling reports were regenerated again on November 7, 2026. The revisions removed evidence found to describe unrelated products and re-scored the affected vendors. They involved no new testing. The full methodology is at huby.ai/methodology.

Framework and weights

Each product is classified as an AI-detection product, subtype Text-detection. That classification activates 29 subcategories across six categories. Scores run from 1.0 to 5.0 at every level and roll up into an overall score using the category weights below. huby sets these weights before collecting any data and does not change them afterward.

Category Weight What it covers
Quality 30% Detection accuracy, robustness against evasion, model coverage, explainability, benchmarking, usability
Security 17% Detection-engine integrity, platform and API security, identity and access, assurance and incident response
Privacy 16% Data governance, AI training use of customer content, regulatory alignment, user controls
Impact, Ethics and Safety 14% Fairness and transparency, responsible detection practices, AI governance, abuse prevention
Use Cases & Pricing 13% Packaging and pricing, enterprise deployment, integrations, reporting, market adoption
Sustainability and Reliability 10% Service reliability, vendor viability, model maintenance, operations, environmental and social responsibility

Score bands: 5.0 Exemplary; 4.0–4.9 Strong; 3.0–3.9 Adequate; 2.0–2.9 Weak; 1.0–1.9 Deficient. Decimals reflect the quality of the evidence within a band.

Evidence hierarchy

Where sources conflict, higher-tier evidence wins:

  1. huby’s own controlled tests.
  2. Independent audits, certifications and peer-reviewed research.
  3. Vendor documentation.
  4. Structured practitioner evidence.
  5. Anecdotal community evidence.

Evidence older than 24 months is discounted. Vendors did not pay for, commission or review any evaluation.

How huby tested detection

huby ran each product against a shared test corpus using its automated test harness. Tests used each product’s public free tier, or a trial account or API key where the free tier was not usable. The corpus contained human-written documents, wholly AI-generated documents (from Claude, Gemini and GPT models), human-polished AI text, adversarially perturbed AI text, AI-paraphrased human drafts and mixed human-AI documents. Results were scored at each vendor’s default 50% threshold.

Metric What it measures How to read it
ROC-AUC How well the score ranks AI-bearing text above human text, regardless of threshold 1.0 is perfect ranking; 0.5 is chance
Precision / Recall Share of flagged documents that contain AI / share of AI documents that get flagged High precision with low recall means a cautious tool that misses AI text
F1 Balance of precision and recall at the default threshold Higher is better
False positive rate Share of human documents flagged as AI The metric that drives wrongful accusations
Expected Calibration Error (ECE) Average gap between the reported "% AI" and the observed frequency 0.10 means scores are off by about 10 points
Mixed-text MAE Average error in the estimated AI share of hybrid documents In percentage points; lower is better
Perturbation and paraphrase detection Detection rate after character-level edits, or on human drafts rewritten by AI Measures robustness to cheap evasion

Limitations readers should weigh

  • Samples are small and uneven. Free-tier and trial limits meant the number of documents scored per product ranged from 5 (Originality AI) to 64 (Sapling and ZeroGPT). Many sub-measures rest on one to six documents, so the test rankings point in a direction rather than fixing a precise rank.
  • The corpus is narrow. It was dominated by a single content domain, used mostly short documents, and contained no writing by non-native English speakers.
  • Some products were scored indirectly. Winston AI was tested on shorter documents because of trial word limits. Originality AI’s scores were derived from its sentence highlighting, because the tested tier shows no numerical score.
  • Scores reflect one point in time. Detectors are retrained often, and vendors may have changed their products since September 2026.
  • Atomic factor scores are left out of this paper. They belong to each vendor’s confidential owner report. This paper reports category and subcategory scores and the test results described in huby’s written assessments.

Comparative analysis

Copyleaks and GPTZero form the front tier at 3.7 and 3.5. Originality AI and QuillBot follow at 3.2 and 3.0, and five vendors score below 3.0. The spread is widest in Impact, Ethics and Safety (1.8 to 3.6) and in Use Cases & Pricing (2.4 to 4.2), at 1.8 points each. It is narrowest in Quality, at 1.0 point (2.8 to 3.8). In other words, vendors differ more in responsible-use practice and enterprise packaging than in detection quality.

Overall ranking

No vendor reaches the 4.0 Strong band; Copyleaks leads at 3.7.

#1
Copyleaks
3.7
Overall score · Adequate
GPTZero
3.5
Overall score · Adequate
Originality AI
3.2
Overall score · Adequate
QuillBot
3.0
Overall score · Adequate
Pangram
2.9
Overall score · Weak
Winston AI
2.9
Overall score · Weak
ZeroGPT
2.8
Overall score · Weak
Undetectable.ai
2.7
Overall score · Weak
Sapling
2.5
Overall score · Weak
Source: huby evaluation reports, revised October 27, 2026 (Pangram and Sapling November 7, 2026) · overall scores for the nine AI detectors

huby groups the cohort into three tiers:

  • Leaders (3.5 or above): Copyleaks and GPTZero.
  • Contenders (3.0–3.4): Originality AI and QuillBot.
  • Limited fit (below 3.0): Pangram, Winston AI, ZeroGPT, Undetectable.ai and Sapling.

Pangram and Winston AI tie at 2.9. Pangram’s detection ranks fourth in the cohort (Quality 3.4), but Privacy and Security scores of 2.4 pull it into the limited-fit tier. Sapling ranks last at 2.5: every category except Quality scores 2.4 or below, despite strong test results.

Category profile

Copyleaks leads five of six categories; ethics and packaging split the field most. Darker cells score higher; bold marks each category leader, including ties.

Vendor Quality Privacy Security Ethics & Safety Reliability Use Cases
Copyleaks 3.8 3.5 3.5 3.6 3.8 4.2
GPTZero 3.3 3.5 3.0 3.4 3.9 4.1
Originality AI 3.7 3.2 2.8 2.8 3.5 3.1
QuillBot 2.8 3.1 2.7 3.1 3.3 3.5
Pangram 3.4 2.4 2.4 3.3 3.0 2.7
Winston AI 2.8 3.3 2.7 2.9 2.6 3.5
ZeroGPT 2.9 2.6 2.5 2.7 2.7 3.4
Undetectable.ai 3.5 2.3 2.3 1.8 2.8 3.0
Sapling 3.0 2.3 2.3 2.3 2.4 2.4
Source: huby evaluation reports, revised October 27, 2026 (Pangram and Sapling November 7, 2026) · six category scores for nine vendors

Quality varies least across vendors. Impact, Ethics and Safety and Use Cases & Pricing vary most. Copyleaks leads Impact, Ethics and Safety at 3.6, and Undetectable.ai’s 1.8 is the cohort’s only Deficient category score. Undetectable.ai’s 3.5 in Quality shows that a competent detector can sit inside a business built on evasion.

Positioning

Copyleaks leads on both axes; GPTZero trades detection for platform maturity.

Vendor Detection quality Trust and enterprise readiness Position
Copyleaks 3.8 3.69 Strong on both
GPTZero 3.3 3.53 Strong on both
QuillBot 2.8 3.11 Trusted, weaker detection
Originality AI 3.7 3.05 Strong on both
Winston AI 2.8 3.01 Trusted, weaker detection
ZeroGPT 2.9 2.76 Weaker on both
Pangram 3.4 2.72 Detects well, weaker trust
Undetectable.ai 3.5 2.40 Detects well, weaker trust
Sapling 3.0 2.33 Weaker on both

Cohort medians: detection quality 3.3; trust and enterprise readiness 3.01. Vendors at the median count as above it.

Source: huby evaluation reports, revised October 27, 2026 (Pangram and Sapling November 7, 2026) · trust and enterprise readiness = category scores weighted by huby’s published weights (Privacy 16, Security 17, Ethics 14, Use Cases 13, Reliability 10)

The map separates two things a single overall score blends together: how well a product detects, and how safely an organization can deploy it. Undetectable.ai and Pangram detect above the median but trail on trust and enterprise readiness; Originality AI detects well but sits only just above the trust median. GPTZero shows the reverse: a mature platform with median detection. Originality AI, QuillBot and Winston AI cluster near the trust median, so small differences there should not drive a purchase. Sapling sits lowest on trust.

Detection test results

The products split on threshold strategy more than on raw ability. Pangram and Originality AI ranked AI text perfectly (ROC-AUC 1.00), yet Pangram flagged only 38% of AI-bearing documents at its default threshold. Sapling caught the most AI text but was the only product to flag human writing.

Vendor Docs scored ROC-AUC Precision Recall F1 False positive rate ECE Mixed-text MAE (pts)
Copyleaks 17 0.81 1.00 0.63 0.77 0% (0 of 9) 0.18 13.1
Pangram 10 1.00 1.00 0.38 0.55 0% (0 of 2) 0.42 30.6
GPTZero 10 0.75 1.00 0.38 0.55 0% (0 of 2) 0.52 44.9
Sapling 64 0.88 0.93 0.79 0.85 18.75% (3 of 16) 0.17 25.3
Originality AI 5 1.00 1.00 0.75 0.86 0% (0 of 1) Not measured 9.5
Winston AI 6 0.75 1.00 0.25 0.40 0% (0 of 2) 0.51 74.3
QuillBot 14 0.77 No flags 0.00 0.00 0% (0 of 3) 0.78 82.1
ZeroGPT 64 0.83 1.00 0.23 0.37 0% (0 of 20) 0.48 31.7
Undetectable.ai 30 0.99 1.00 0.62 0.77 0% (0 of 6) 0.29 8.0

Notes:

  • All results are at the vendor’s default 50% threshold.
  • Originality AI’s scores were derived from its sentence highlighting, because the tested tier shows no numerical score. For the same reason, its calibration could not be measured.
  • QuillBot flagged no documents at all, so its precision is undefined.

Read this table alongside the sample sizes in its "Docs scored" column; several results rest on very few documents.

Robustness to cheap evasion

Cheap evasion defeats most detectors: six of nine caught no AI-paraphrased text. Share of AI-bearing documents flagged in huby’s tests.

Vendor Unmodified AI text After character perturbation AI-paraphrased drafts
Copyleaks 100% (n=4) 0% (n=1) 0% (n=1)
Pangram 100% (n=2) 0% (n=2) 0% (n=2)
GPTZero 100% (n=2) 0% (n=2) 0% (n=2)
Sapling 100% (n=11) 73% (n=11) 83.3% (n=12)
Originality AI 100% (n=1) 100% (n=1) 0% (n=1)
Winston AI 100% (n=1) 0% (n=1) 0% (n=1)
QuillBot 0% (n=3) Not measurable 0% (n=2)
ZeroGPT 40% (n=10) 25% (n not reported) 20% (n=10)
Undetectable.ai 100% (n=6) 33.3% (n=6) 33.3% (n=6)
Source: huby controlled tests, August–September 2026 · free or trial tiers, vendor default 50% threshold

Homoglyph substitution and AI paraphrasing cost an attacker almost nothing, and they defeated most of the cohort. Only Sapling held up on both. Originality AI resisted perturbation but missed paraphrased text, and Undetectable.ai and ZeroGPT were partly resistant. Treat robustness claims with skepticism unless a vendor shows results against named evasion tools.

Service reliability and speed

Most services returned a usable score for every submission, with two exceptions. Sapling succeeded on 137 of 148 submissions (92.6%), and Winston AI on 6 of 7 (85.7%). QuillBot’s free tier hit usage quotas after seven requests during huby’s load test.

Among the seven products where latency was measured, 95th-percentile response times were:

  • Fastest: Sapling (10.4 s) and ZeroGPT (10.8 s).
  • Mid-range: QuillBot (13.5 s), Pangram (13.8 s) and Winston AI (16.9 s).
  • Slowest: Undetectable.ai (20.2 s) and GPTZero (24.0 s).

These measurements come from free or trial tiers and may not reflect paid API performance.

Pricing and packaging

Entry prices are low across the market. For institutions, packaging differs more than price: what matters is whether API access, SSO, LMS connectors and non-expiring credits are included or reserved for custom-quoted plans.

Vendor Free access Entry paid plan (monthly) Allowance on entry plan Model
Copyleaks Limited trial credits $16.99 25,000 words Non-expiring credits; enterprise plans add LMS, SSO, API
GPTZero Free tier $12.99 30,000 words Subscription; usage-based API
Originality AI 3 scans a day $12.95 2,000 credits (about 200,000 words) Credits; API on Enterprise ($179/month)
Pangram 2,000 words a day $20.00 300,000 words Subscription; API pricing not published
QuillBot Limited free tier About $4.17 (billed annually) Bundled writing suite Subscription suite; contact sales for teams
Sapling 2,000 characters per check $25.00 100,000 characters per check Subscription; usage-based API; Enterprise from $15 per seat
Undetectable.ai Free detector $9.99 10,000 words Word-based subscription
Winston AI 14-day trial (2,000 credits) $10.00 100,000 credits Subscription; API and team plans
ZeroGPT No-signup free tier $13.99 100,000 characters per detection Subscription; usage-based API

Prices are vendor-provided as of October 22, 2026, from each product’s huby page; QuillBot’s figure comes from huby’s evaluation. Two listings contain inconsistencies that buyers should confirm with the vendor. Originality AI’s annual Enterprise price ($1,638.96) does not equal 12 months at $179. ZeroGPT lists its MAX plan ($20.99) below its PLUS plan ($22.99).

Vendor profiles

The profiles appear in huby rank order. Each one gives the overall and category scores, what huby’s tests showed, and the buyers the product best fits. Category scores are listed as Quality / Privacy / Security / Impact, Ethics and Safety / Sustainability and Reliability / Use Cases & Pricing.

1. Copyleaks: 3.7, Adequate (leader)

Category scores: 3.8 / 3.5 / 3.5 / 3.6 / 3.8 / 4.2. First in five of six categories, sharing first place in Privacy with GPTZero.

Copyleaks is the most complete institutional platform in the cohort, but its detection engine is as easy to evade as most others. Founded in 2015 and still led by its founders, it combines AI detection (30+ languages), plagiarism checking (100+ languages), source-code checks and image detection in one credit-based service.

Strengths

  • Enterprise deployment and integration lead the cohort: Enterprise Deployment and Administration scored 4.0, and Workflow Integrations and Extensibility 4.3. Copyleaks offers SCIM provisioning with Okta and Entra ID, SAML SSO, multi-tenant administration, US or EU data residency, and on-premise or private-cloud options. It also has native LTI connections to Canvas, D2L, Moodle and Brightspace.
  • Clean test results on unmodified text: no false positives (0 of 9 human documents), no misses on purely AI documents (0 of 4), F1 0.77 (third), and the second-best calibration (ECE 0.18). It processed 100% of submissions, with independently measured latency of 510 ms.
  • Strong business footing: Vendor Sustainability and Business Viability scored 4.4, supported by an Inc. 5000 ranking (#153) and a 4.5/5 rating from 41 G2 reviews. Copyleaks is so widely used that other tools benchmark against it.

Cautions

  • Fragile under manipulation. A single character-level perturbation cut detection from 100% to 0%, tying for the worst result in the cohort. Copyleaks also missed an AI-paraphrased document (n=1) and identified AI in only one of two mixed documents. Overall ROC-AUC was 0.81, sixth of nine. Academic work cited by huby measured 58.7% accuracy on expert-edited machine text.
  • The ">99% accuracy" claim is not backed by accessible studies. Peer-reviewed research cited by huby finds disproportionate false positives on non-native English writing.
  • Privacy messaging conflicts. Copyleaks has a "never train on customer content" policy, yet markets a "Shared Data Hub for enhanced detection accuracy."
  • Thin public assurance. SOC 2 and ISO 27001 reports are not public, and there is no public SLA.

Best fit: universities, school systems and enterprises that need LMS-native workflows, centralized administration and data-residency choices. Pair it with human review and with process evidence such as draft history.

Copyleaks on huby

2. GPTZero: 3.5, Adequate (leader)

Category scores: 3.3 / 3.5 / 3.0 / 3.4 / 3.9 / 4.1. Joint first in Privacy and first in Sustainability and Reliability.

GPTZero is the most widely adopted specialist detector, and its business is among the strongest in the cohort. In huby’s tests, though, its detection was among the weakest in the cohort. Independent reports cite more than 19 million registered users, profitability, 500% ARR growth and a $10 million Series A. huby’s evaluation also notes its acquisition by Superhuman.

Strengths

  • Viability and ecosystem. Vendor Sustainability and Business Viability scored 4.6, and Workflow Integrations 4.4. GPTZero offers add-ins for Word and Google Docs, a Chrome extension, connectors for Canvas, Blackboard and Moodle, Zapier, webhooks and a documented API.
  • Education partnerships. Through the American Federation of Teachers, 15,000 educators get free access. GPTZero has also run large-scale analyses, such as checking 4,841 NeurIPS papers for hallucinated citations.
  • Responsible-use messaging. Results are presented as probabilities with sentence highlighting. GPTZero explicitly warns against punitive use and publishes resources for falsely accused students.
  • Transparent free versus paid data terms. Paid and API content is not stored or used for training.

Cautions

  • Weakest ranking accuracy in huby’s tests: ROC-AUC 0.75, tied for last. Recall was 0.38 (F1 0.55), and the second-worst calibration (ECE 0.52) means its "AI probability" behaves as a directional signal, not a probability.
  • No resistance to cheap evasion. Detection fell from 100% to 0% under character perturbation, and GPTZero flagged none of two paraphrased documents. Mixed-text error was 44.9 points.
  • Weaker on some content types. Independent studies cited by huby report a 48% miss rate on scientific writing, 66% detection of AI fiction, and false positives on formal and non-native writing.
  • Free-tier privacy and identity gaps. Free-tier submissions train the model, with no self-service opt-out. Identity and Access Management scored 2.6: there is no native MFA for password accounts and no role-based access control.
  • Slowest service tested, at 24.0 seconds (95th percentile) on the free tier.

Best fit: K-12 and higher-education teams that value broad integrations, teacher-facing guidance and vendor stability. Use it with conservative policies. Procurement should require role-based access control and SSO terms, and should validate recall on edited text.

GPTZero on huby

3. Originality AI: 3.2, Adequate (contender)

Category scores: 3.7 / 3.2 / 2.8 / 2.8 / 3.5 / 3.1. Second in Quality.

Originality AI is a detector built for SEO professionals, content agencies and publishers. It sells pay-as-you-go credits and offers several model choices: Turbo, Lite, Academic and Multilingual. In huby’s small test it posted top-tier accuracy. Enterprise controls, assurance and responsible-use guidance lag the leaders.

Strengths

  • Best F1 in huby’s tests (0.86) and perfect ranking (ROC-AUC 1.00), with no false positives and the second-lowest mixed-text error (9.5 points). These results come from only five documents, scored from sentence highlighting.
  • The only product unaffected by character-level perturbation (detection stayed at 100%, n=1). Originality AI documents its defense against homoglyphs.
  • Independent paraphrase benchmarks are favorable. RAID-based results cited by huby range from 85% to 96.7%. In a StealthGPT humanizer study, Originality AI’s false negative rate rose only to between 5% and 21%, far less than some peers.
  • Useful publisher features: a full-site scanner, a Chrome extension, writing replay, a patented fact-checker (U.S. Patent No. 12,705,435) and a configurable "AI allowance" from 0% to 40%.

Cautions

  • Many users report false positives. These include human writing flagged at 97–100% and terminated freelance contracts. Originality AI’s own documentation states a 1.56% false positive rate; its marketing claims "99% accuracy."
  • It failed huby’s paraphrase test (n=1), and its scores are not calibrated probabilities. Its interface shows binary "GENAI DETECTED" verdicts, which work against human review.
  • Training policy is contradictory. Marketing says content is never used for training, but help-center material describes default opt-out training.
  • Thin enterprise and education controls: no SOC 2 or ISO 27001 attestation, no SCIM, no multi-tenancy, no on-premise option and no native LMS integration (Enterprise Deployment 2.5).
  • Key-person dependency on a founder-led company that does not disclose its finances.

Best fit: SEO teams, publishers and agencies that audit writers’ content at volume and value non-expiring credits and site scanning. Use the AI allowance and human review; avoid using a single score as grounds for terminating a contract.

Originality AI on huby

4. QuillBot: 3.0, Adequate (contender)

Category scores: 2.8 / 3.1 / 2.7 / 3.1 / 3.3 / 3.5. Joint last in Quality, with Winston AI.

QuillBot is a mass-market writing suite, owned by Learneo since 2021, whose core product is a paraphraser; AI detection is a secondary feature. In huby’s tests the detector did not flag a single AI document at its default threshold.

Strengths

  • Huge reach and low price. QuillBot ranks among the top 20 generative AI products by web traffic. The Premium bundle costs about $4.17 a month billed annually, and Market Adoption scored 3.9.
  • Integrates where people write: add-ins for Microsoft Word and Google Docs, a Windows app, and browser extensions for Chrome and Edge.
  • No false positives, both in huby’s test (0 of 3) and in a peer-reviewed study cited by huby. The interface also warns against relying on the detector alone.

Cautions

  • Detection failed in huby’s tests. QuillBot flagged none of 11 AI-bearing documents, including all three wholly AI documents, for an F1 of 0.00. It was the worst in the cohort on mixed text (82.1-point error) and on calibration (ECE 0.78). Text from all three foundation models went undetected.
  • Built-in conflict of interest. The paraphraser and a separate "AI Humanizer" exist to change the signals that detectors rely on, and huby found a product video claiming the paraphraser evades plagiarism detection. QuillBot’s own detector sometimes misses its paraphraser’s output.
  • Not built for institutions. There is no batch detection, no public API, no LMS integration and no SCIM, and the terms of service prohibit independent benchmarking.
  • Weak privacy defaults. Free-tier content trains models by default, data rights are handled through manual requests, and an independent analysis flags the browser extension’s data access.

Best fit: individual writers who already use QuillBot and want a quick, low-stakes self-check. huby does not recommend it for integrity enforcement or any decision about another person.

QuillBot on huby

5. Pangram: 2.9, Weak (limited fit)

Category scores: 3.4 / 2.4 / 2.4 / 3.3 / 3.0 / 2.7. Tied with Winston AI at 2.9.

Pangram publishes more research on its own models than any other vendor in the cohort, and it has the most cautious default threshold, but almost none of what a buyer needs to verify its privacy, security or administration is publicly documented. huby’s evaluation treats most of those gaps as missing evidence rather than confirmed failings, and rates confidence in them as low. Pangram has raised about $13 million (a seed round of roughly $4 million in June 2025 and $9 million in July 2026), and its detection is embedded in Inspera, Tremau, Proofig, NewsGuard and Substack.

Strengths

  • Well-documented detection research. Benchmarking and Validation scored 4.1, second only to Copyleaks. Pangram publishes an arXiv technical report, versioned model cards for Pangram 3.2 and 4, open-weight models and a peer-reviewed workshop paper. Most of this evidence is vendor-authored. A Becker Friedman Institute study of 1,992 human passages found Pangram’s false positive rates stayed within its stated caps.
  • Perfect ranking in huby’s tests: ROC-AUC 1.00, no false positives (the highest score given to any human document was 0.02%), and no misses on purely AI documents. Results were consistent across Claude and GPT text, and every submission was processed.
  • Active model maintenance. Detection Model Maintenance scored 4.0, tied for first. Each release ships with a model card, and the Pangram 4 technical report cuts the false negative rate on the same 520,000-output test set from 1.99% (Pangram 3.3.2) to 0.34%.
  • Low API cost. A working paper cited by huby measured $0.0016 per passage for Pangram, against $0.0029 for Originality AI and $0.0040 for GPTZero. The paper covers API pricing only, and its funding is not shown.

Cautions

  • Free-tier recall was low (0.38, F1 0.55). huby could not reproduce the vendor’s published recall on the publicly available tier.
  • Evasion gaps. Character-level perturbation cut detection from 100% to 0%, and Pangram flagged neither of two AI-paraphrased documents. One independent analysis found 495 of 500 machine-translated AI samples classified as human, and Epoch AI found Pangram missed about one in five passages written to imitate an author’s style.
  • Scores are poorly calibrated (ECE 0.42), and mixed text is handled poorly (MAE 30.6 points). A "Confidence limited — short text" notice can appear next to a "100% AI" or "100% Human Written" verdict, so the caveat never reaches the headline figure. Independent writers describe verdicts that flip after small edits.
  • Privacy terms cannot be verified (Privacy 2.4). Pangram’s privacy policy and terms were available only as snippets, so no retention, deletion, ownership or transfer clause could be read. Its "never used for training", GDPR and SOC 2 statements are not backed by a DPA, a subprocessor list or an audit report.
  • Enterprise readiness is undocumented (Security 2.4, Use Cases & Pricing 2.7). A "Continue with SSO" option exists, but the protocols, SCIM, MFA, administrator roles and audit logs are not documented. There is no public SLA, status page or incident history, and plan pricing beyond the 300,000-word Individual plan is not published.

Best fit: publishers, online platforms and trust-and-safety teams that put avoiding false accusations first, and developers who want low-cost API scoring. Buyers should confirm recall on the tier they would license, and obtain a DPA, a SOC 2 report and service terms in writing before deployment.

Pangram on huby

6. Winston AI: 2.9, Weak (limited fit)

Category scores: 2.8 / 3.3 / 2.7 / 2.9 / 2.6 / 3.5. Second-lowest in Sustainability and Reliability, after Sapling.

Winston AI pairs a careful, review-oriented interface with the widest gap in the cohort between claimed and measured accuracy. It is a bootstrapped company founded in 2022. Third-party estimates put it at about three employees and $330,000 in annual recurring revenue. Besides text detection it offers plagiarism checking, image and deepfake detection, OCR and fact-checking.

Strengths

  • Responsible design in the product itself. Results appear as a "Human Score" with cautious explanations, and the AI prediction map states openly that sentence-level results are less accurate than the document score. Model versions are shown in the output.
  • Practical packaging and integrations: low-cost tiers from $10 a month, non-expiring pay-as-you-go credits, an API with webhooks, Zapier and n8n connectors, a Moodle plugin, browser extensions for Chrome, Firefox and Edge, and partnerships with Crowdin and SmartRecruiters (Use Cases & Pricing 3.5).
  • No false positives in huby’s test (0 of 2 human documents).

Cautions

  • The "99.98% accuracy" claim is contradicted. One independent benchmark cited by huby measured about 42%. Winston AI’s own news item describing it as "top ranked" in a 2026 audit is a vendor publication, not independent validation. In huby’s test, ROC-AUC was 0.75, tied for last, with recall 0.25 and F1 0.40. Results were measured on shorter documents because of trial limits.
  • No resistance to cheap evasion. Perturbation cut detection from 100% to 0%, and Winston AI missed the AI-paraphrased document (n=1). It assigned 0% AI to an adversarial sample. Mixed-text error was 74.3 points (n=1), and calibration error was 0.51.
  • Least reliable service tested: 85.7% processing success and a 16.9-second 95th-percentile latency, with no public SLA. Sustainability and Reliability scored 2.6, second-lowest in the cohort.
  • Unverified assurance. The SOC 2 Type 2 claim is not backed by a public report, there is no native MFA for password accounts, and Winston AI does not disclose higher false positive risk for non-native writers.

Best fit: individual educators and small content teams that want an inexpensive, easy-to-read checker for low-stakes screening. It is not suitable as evidence in integrity or hiring decisions without independent corroboration.

Winston AI on huby

7. ZeroGPT: 2.8, Weak (limited fit)

Category scores: 2.9 / 2.6 / 2.5 / 2.7 / 2.7 / 3.4.

ZeroGPT is a high-traffic freemium detector that needs no signup. In huby’s tests it ranked mid-pack on ranking accuracy but missed most AI text at its default threshold. It also sells a humanizer that defeats its own detector.

Strengths

  • Easy access: a generous no-signup free tier, batch upload, and an API with webhooks.
  • Fast and dependable service in huby’s tests: 115 of 115 documents processed, with a 10.8-second 95th-percentile latency (second fastest).
  • No false positives on huby’s largest human sample (0 of 20), and the second-smallest drop under character perturbation (from 40% to 25%), though from a low starting detection rate.
  • Median ranking accuracy (ROC-AUC 0.83 on 64 documents), in line with academic studies that put its accuracy in the low 80s.

Cautions

  • Misses most AI text: recall 0.23 and F1 0.37 (eighth). ZeroGPT missed 60% of wholly AI documents, and 75% to 80% of adversarial and polished AI text.
  • Coverage is uneven: the widest gap across foundation models in the cohort. ZeroGPT detected 50% of Claude, 40% of Gemini and 33% of GPT samples, and failed the Hindi-English code-switched test.
  • Scores are poorly calibrated (ECE 0.48) and not reproducible. Highlighted sentences can change between runs. Widely reported false positives include the U.S. Constitution scored at 92.15% AI.
  • Self-defeating product line. In a documented demonstration, ZeroGPT’s humanizer cut its own detector’s score from 70.1% to 16.6%. Its marketing also presents generated reports as "proof."
  • Weak privacy and assurance. Statements on whether submitted text is stored or used for training contradict each other. There is no SOC 2, no vulnerability disclosure program, no status page and no SLA (Security 2.5).

Best fit: casual, low-stakes self-checks by individuals. It is not suitable for institutional procurement or for any decision with consequences.

ZeroGPT on huby

8. Undetectable.ai: 2.7, Weak (limited fit)

Category scores: 3.5 / 2.3 / 2.3 / 1.8 / 2.8 / 3.0. Last in Privacy, Security and Impact, Ethics and Safety.

Undetectable.ai is primarily an evasion product: its core service rewrites AI text to pass detectors such as Turnitin, GPTZero, Originality AI and Copyleaks. Its bundled detector, which aggregates third-party detection models, performed surprisingly well in huby’s tests. That performance does not change the overall verdict.

Strengths

  • Competent detection on huby’s corpus: ROC-AUC 0.99 (third) on 30 documents, no false positives (the highest score on a human document was 27%), and the lowest mixed-text error in the cohort (8.0 points).
  • Second-best paraphrase detection (33.3%) and a smaller-than-median drop under perturbation (from 100% to 33.3%, fourth of eight). It detected 100% of unmodified text across Claude, Gemini and GPT.
  • Reliable processing (60 of 60 submissions), a large user base (23 million users, per the vendor), and estimated annual recurring revenue of $3.7 million.

Cautions

  • Built to defeat integrity controls. huby’s evaluation documents the tool reducing a 100% Turnitin AI score to 0%, an affiliate program aimed at students, and a university handbook that names it. Impact, Ethics and Safety scored 1.8, the lowest in the cohort.
  • Detection collapses on edited text. Detection of polished and adversarial AI text fell to 33.3%. Calibration error was 0.29.
  • Deficient privacy. Under the terms of service, using the product counts as consent to train on submitted content, and there is no opt-out. Technical evidence shows content shared with OpenAI and Google, which is not disclosed in public policy. AI Data Usage scored 1.9.
  • No enterprise security. There is no SAML or SCIM, no role-based access control and no native MFA (Identity and Access Management 1.7). There is also no SOC 2 or ISO 27001, no vulnerability disclosure program, no status page and no SLA.
  • Founder dependency and no disclosed funding in a market that depends on continuous R&D to keep up with detectors.

Best fit: none for institutional detection. Its detector results are useful as a benchmark: they show that a product can be competent at detection while its owner’s business runs in the opposite direction. Integrity programs should treat the tool as a known evasion vector and test their own detectors against it.

Undetectable.ai on huby

9. Sapling: 2.5, Weak (limited fit)

Category scores: 3.0 / 2.3 / 2.3 / 2.3 / 2.4 / 2.4. Last in Sustainability and Reliability and in Use Cases & Pricing; joint last with Undetectable.ai in Privacy and Security.

Sapling is a writing assistant for business teams with an AI detector included. Its detector produced some of the strongest results in huby’s tests, but nearly everything around it, from security assurance to the company’s own viability, rests on vendor claims or is undocumented. A Y Combinator profile lists a 2019 founding by Ziang Xie and two employees, which points to a very small company. Sapling also sells a humanizer.

Strengths

  • Highest recall and second-best F1 (0.79 and 0.85) on the largest sample tested (64 documents). ROC-AUC was 0.88, fourth. Sapling detected 100% of unmodified AI text across Claude, Gemini and GPT.
  • Best calibration in the cohort (ECE 0.17). The detector offers document- and sentence-level scores, highlighting and per-sentence perplexity (Detection Explainability and Confidence 3.5).
  • Most robust against evasion. Sapling caught 83.3% of AI-paraphrased documents, the best in the cohort and 50 points ahead of the next product. It still detected 73% of perturbed AI text (third of eight).
  • Fastest service tested (10.4 s at the 95th percentile), with a documented API quota of 50,000 characters a day and 250,000 a month per key by default.

Cautions

  • Highest false positive rate tested: 3 of 16 human documents were flagged (18.75%), one at 98.9% AI. One independent study found 64.1% of human writing misclassified, and research documents bias against non-native English writers. Sapling’s claim of a 0% false positive rate is contradicted, it publishes no model card, and third parties describe the detector as English-only.
  • Detection and evasion under one roof. Sapling’s "Advanced AI Humanizer" is designed to make text score as human in detectors, which works directly against its own detector.
  • Less reliable processing. Usable scores came back for 92.6% of submissions, eighth of nine, and Sustainability and Reliability scored 2.4, the lowest in the cohort. No revenue, funding or customer figures are available to judge the company’s durability.
  • Assurance rests on vendor claims. SOC 2 Type II, SAML single sign-on and self-hosted deployment are described only on Sapling’s own pages, with no audit report. Data deletion requires a manual support request.
  • Weak defaults for individuals. Free and individual users have no self-service training opt-out, and the public detector keeps prompt and detection history for three days. Sapling markets an AI resume detector without guidance on bias or human oversight.

Best fit: business and customer-support teams already using Sapling’s writing assistant who want an inexpensive, explainable second opinion. It is a weak fit for academic-integrity or hiring decisions while its false positive rate stays this high and its assurance is unverified.

Sapling on huby

Recommendations by use case

Choose a detector for the decision it will support. A tool that is fine for screening blog submissions can be harmful in a misconduct hearing. The table gives huby’s shortlist by use case; the guidance below covers the controls each context needs.

Use case Shortlist Evaluate with care Avoid Must-have controls
Higher education and K-12 integrity Copyleaks, GPTZero Pangram Undetectable.ai, QuillBot, Sapling, Winston AI, ZeroGPT LMS integration, SSO, human-review workflow, appeal process, process evidence
Publishing, SEO and content agencies Originality AI, Copyleaks Pangram, Winston AI Undetectable.ai Site scanning, non-expiring credits, configurable tolerance, shareable reports
Enterprise, HR and compliance Copyleaks GPTZero Undetectable.ai, QuillBot, Sapling, ZeroGPT SOC 2 report, DPA, data residency, no-training terms, employment-use guidance
Platforms, trust and safety, developers Copyleaks Pangram, GPTZero Undetectable.ai Documented API limits, SLA, latency data, versioned models

Education and academic integrity

No product in this cohort should be the basis for a misconduct finding. Use detection to decide which work to look at more closely, then rely on evidence a student can respond to: draft and version history, writing-replay records, oral follow-up and assignment design.

  • Copyleaks offers the deepest LMS and administrative integration. Pangram offers the most cautious default threshold and strong published validation, but its privacy and administration terms are undocumented, so get them in writing before a pilot.
  • GPTZero’s ecosystem and teacher guidance are strong, but its ranking accuracy and calibration were among the weakest huby tested.
  • Set institutional thresholds locally, and do not show raw percentages to students as verdicts.
  • Account for the documented risk to non-native English writers. Require vendors to share false positive data for ESL writing before you sign.

Publishing, SEO and content agencies

The risk here is wrongly penalizing writers and breaking contracts. Originality AI’s site scanning, writing replay and configurable "AI allowance" suit editorial workflows. Make sure contracts with writers do not treat a detector score as proof. Pangram suits publishers that want a lower false-positive posture and API access, provided they can obtain its data-handling terms in writing.

Enterprise, HR and compliance

Applying detection to hiring and employment raises legal and fairness risk, and most vendors provide little guidance for it. Copyleaks has the most complete enterprise controls (SCIM, SSO, residency, on-premise options).

  • Require a SOC 2 Type 2 report under NDA, an executed DPA, written no-training commitments and a retention schedule.
  • Do not use AI resume detectors to screen candidates out.

Platforms, trust and safety, and developers

At platform scale, ranking quality and API economics matter more than single-document verdicts. Pangram’s perfect ranking in huby’s test and its low measured API cost suit moderation queues where scores route content to human reviewers, but it publishes no SLA, status page or API rate limits, so contract for them. Copyleaks offers webhooks and mature asynchronous processing.

  • Calibrate thresholds on your own traffic, and expect performance to drift as new LLMs ship.
  • Expect adversaries to use homoglyphs and paraphrasers. Add text normalization ahead of detection.

A procurement checklist

  1. Pilot on at least 200 of your own documents. Include human writing from your actual population, non-native writers among them, plus AI, edited-AI, paraphrased and mixed samples.
  2. Measure false positive rate, recall at your chosen threshold, and calibration. Do not accept vendor accuracy claims without the dataset behind them.
  3. Test the vendor’s detector against common humanizers, including the vendor’s own if it sells one.
  4. Require audit reports, a DPA, residency options and a training opt-out in the contract.
  5. Write human-review and appeal steps into policy before the tool goes live.
  6. Re-test after every major model release, at least every six months.

Outlook: the next 12–24 months

Over the next two years, huby expects buyers to move from asking "was this written by AI?" to asking "can this author show how the work was produced?" Detection will stay, but as one signal among several. The forecasts below are huby’s analytical judgments, drawn from patterns across this cohort; they are not vendor commitments.

huby forecasts

  1. Process evidence becomes a procurement requirement. By late 2027, most institutional RFPs for integrity tools will require draft-history or writing-replay evidence alongside a classifier score. GPTZero and Originality AI already offer replay features; others will follow or partner.
  2. Calibration and false-positive disclosure become table stakes. Buyers and regulators will push vendors to publish false positive rates by content type and language group. Vendors that keep relying on a single headline accuracy figure of "99%" will lose institutional deals.
  3. Cheap evasion forces input hardening. Homoglyph and whitespace attacks defeated half of the products where they could be measured. Text normalization, and testing against named humanizers, will become standard release gates within 12 months.
  4. Detect-and-evade portfolios come under pressure. Institutions will increasingly ask suite vendors to separate or drop humanizer features. Expect procurement policies that exclude vendors selling evasion tools from integrity contracts.
  5. Consolidation continues. Acquisitions such as Superhuman’s purchase of GPTZero signal that detection is becoming a feature of larger writing and productivity platforms. Small, founder-led detectors will be acquired, specialize, or exit.
  6. Detection broadens to multimodal provenance. Image, audio and video detection will be bundled with text, and buyers will judge vendors on provenance across formats.

What would change huby’s view

  • Independent results at scale. Large audits that show reliable recall on edited and paraphrased text without higher false positives would justify more weight on classifier scores.
  • Provenance standards. Widespread adoption of watermarking or signed provenance by major model providers would shift value away from statistical detection.
  • Fairness evidence. Published, independent evidence that false positives no longer fall disproportionately on non-native writers would lower the risk of using detection in education.

Appendix

A. Subcategory scores

The table lists all 29 subcategory scores (1.0–5.0) for each vendor, in huby rank order. Category scores appear in the Comparative analysis.

Subcategory Copyleaks GPTZero Originality AI QuillBot Pangram Winston AI ZeroGPT Undetectable.ai Sapling
1.1 Benchmarking and Validation 4.4 3.8 3.7 3.1 4.1 2.2 2.8 2.7 2.6
1.2 Detection Accuracy and Performance 3.8 3.0 4.0 2.4 3.6 2.7 3.1 4.2 3.1
1.3 Detection Explainability and Confidence 3.9 3.9 3.3 3.1 3.1 3.1 2.2 2.9 3.5
1.4 Model Adaptability and Coverage 3.8 3.6 4.0 2.7 3.8 3.3 2.7 3.7 3.2
1.5 Robustness Against Evasion 3.2 2.7 3.2 2.8 2.7 2.5 3.0 3.2 2.9
1.6 Usability and Workflow Integration 4.5 3.6 3.7 3.2 2.8 3.6 3.2 3.3 2.4
2.1 AI Data Usage 3.5 3.0 3.2 3.2 2.6 3.4 2.7 1.9 2.2
2.2 Data Governance and Ownership 3.7 3.8 3.4 3.4 2.2 3.4 2.3 2.7 2.5
2.3 Privacy Compliance and Regulatory Alignment 3.5 3.7 3.1 2.9 2.3 3.3 2.7 2.0 2.2
2.4 User Privacy Controls 3.2 3.4 3.2 2.9 2.3 3.2 2.8 2.5 2.2
3.1 Application and API Security 3.3 3.4 2.6 2.5 2.2 3.1 2.6 2.3 2.2
3.2 Detection Engine Integrity 3.5 3.1 3.6 2.5 2.9 2.4 2.8 2.5 2.2
3.3 Identity and Access Management 3.9 2.6 2.6 3.3 2.2 2.7 2.8 1.7 2.2
3.4 Platform Security 3.8 3.2 2.7 2.7 2.2 2.9 2.4 3.0 2.6
3.5 Security Assurance and Incident Response 3.1 2.7 2.3 2.7 2.4 2.3 2.1 1.9 2.2
4.1 Abuse Prevention and Public Trust 3.6 3.4 3.0 2.2 3.1 2.6 2.6 1.7 2.2
4.2 Detection Fairness and Transparency 3.6 3.1 2.7 3.0 3.5 2.8 2.4 1.9 2.5
4.3 Responsible AI Governance 3.3 3.5 2.6 3.2 3.2 2.9 2.5 1.7 2.4
4.4 Responsible Detection Practices 3.9 3.6 2.8 3.7 3.3 3.4 3.2 1.8 2.2
5.1 Detection Model Maintenance 2.7 4.0 2.9 2.4 4.0 2.6 2.4 2.6 2.2
5.2 Detection Service Reliability 4.3 3.9 3.9 3.1 2.8 2.5 3.1 3.5 2.9
5.3 Environmental and Social Responsibility 3.9 3.6 3.2 3.1 2.9 2.8 2.3 1.5 2.2
5.4 Operational Excellence 3.7 3.0 3.5 3.5 2.3 2.6 2.6 2.3 2.4
5.5 Vendor Sustainability and Business Viability 4.4 4.6 3.5 4.2 3.0 2.8 2.6 3.2 2.2
6.1 Commercial Packaging and Pricing 4.4 4.2 3.7 4.1 2.7 3.2 3.7 3.2 2.5
6.2 Detection Analysis and Reporting 3.5 3.8 3.0 3.1 2.2 3.7 3.5 3.2 2.2
6.3 Enterprise Deployment and Administration 4.0 3.6 2.5 3.1 2.2 3.0 2.6 2.2 2.4
6.4 Market Adoption and Customer Success 4.5 4.3 3.3 3.9 2.9 3.6 4.0 3.0 2.2
6.5 Workflow Integrations and Extensibility 4.3 4.4 3.0 3.1 3.2 4.2 3.4 3.4 2.8

B. Sources

huby methodology and product pages

huby evaluation reports. These are nine detailed evaluation reports, first produced October 1, 2026 and revised October 27, 2026; the Pangram and Sapling reports were regenerated on November 7, 2026. They run from 218 to 271 pages each. Evidence was collected between August 30 and September 21, 2026. They include huby’s controlled test results and the full citation trail behind every score.

Research cited in huby’s evaluations and on huby product pages (selected)

C. Disclosures

  • Independence. No vendor paid for, commissioned or reviewed this paper or the underlying evaluations. huby charges no fees for evaluation, listing or reports, and has no commercial relationship with any vendor in this cohort. Affiliate programs listed on some product pages pay huby no commission.
  • Point in time. Findings reflect products as evaluated in August and September 2026. This edition uses huby’s October 27, 2026 revision, which removed evidence describing unrelated products and re-scored all nine vendors without new testing, together with the November 7, 2026 regeneration of the Pangram and Sapling reports. Prices are as of October 22, 2026.
  • Use of AI. huby uses AI-assisted analysis on its evidence base. AI can misread or omit information; huby does not guarantee completeness.
  • Not advice. This paper is not legal, security, compliance or procurement advice. Base purchase decisions on your own evaluation.
  • Trademarks. Product names belong to their owners; mention does not imply endorsement.
  • Corrections. Report errors to support@huby.ai.

Explore each detector’s full evaluation and Trust Score on huby.

See the evaluations on huby.ai →

Evaluation methodology: huby.ai/methodology