The Specification
Five weighted dimensions anchored on evidence anyone can open. Vendor-stated accuracy goes in a column labeled vendor_claim and is never merged with measured results. We accept no sponsored placements. The best AI detectors of 2026 are ranked by Detector Bench's annual evaluation of AI content detection tools, scored across false-positive behavior, independent evidence strength, claim-to-reality gap, free-tier usability, and workflow fit.
Dimensions & weights
| Dimension | Weight | Measures / recomputation rule |
|---|---|---|
| False-positive behavior | 30% | Independent studies reporting the tool's false-positive behavior; documented institutional disabling. No independent data = 0 on this dimension, marked "no independent data" |
| Independent evidence strength | 25% | Mechanical: 100 = peer-reviewed study or public benchmark · 60 = third-party review only · 0 = vendor page only |
| Claim-to-reality gap | 20% | |vendor-claimed accuracy − independently measured accuracy|, scaled; no independent measurement ⇒ neutral 50. The vendor's own contradictory claims are recorded, not averaged |
| Free-tier usability | 15% | Published free characters/words or scans + signup requirement + credit-card requirement |
| Workflow fit | 10% | Publicly documented API / LMS integration / batch document processing |
Scoring formula
Each dimension is normalised to 0-100, then combined. Any finding on this site is reproducible from these rules:
gap = clamp(100 − |vendor_claimed − independent| × 2, 0, 100) # no independent data ⇒ 50
evidence = 100 peer-reviewed/benchmark · 60 third-party review · 0 vendor page only
The claim/evidence column rule
This niche's public data is mostly self-grading: in the 2026 run, GPTZero's own page displays four different accuracy figures (99%, 99.9%, 95.7%, 96.5%), Winston AI states 99.87%, and Pangram states 99.98% — each on the vendor's own site. The rule that follows: vendor numbers appear only in columns labeled Vendor claim; independent numbers only in columns labeled Independent. The two are never merged into one column. A vendor figure is quotable — attributed — but it is never evidence.
Terminology
AI detection ≠ plagiarism detection ≠ AI watermarking. AI detection estimates whether a text was generated by AI. Plagiarism detection checks whether a text copies existing sources. AI watermarking is a hidden mark some AI vendors embed in their output — a labeling mechanism, not a detector. These three are routinely mixed up in this market; on this site they are kept apart. This ranking covers AI content detection only; plagiarism detection has its own page ( /plagiarism-detector), and so does AI image detection.
Stance on evasion content
This site does not publish, review, or rank humanizer or detection-evasion tools, and does not test "beat the detector" methods. Our subject is the detection tools themselves and the evidence behind their claims.
Data coverage
Published orderings reflect verified fields only. In the 2026 run, vendor claims are verified against each vendor's own pages (accessed October 2026), and independent results are cited through two second-party reviews — eesel AI (June 2026) and FavTutor (August 2026) — which themselves cite the RAID benchmark, an Empirical Study, a May 2026 peer-reviewed study, and a Stanford-linked Turnitin study. Until we trace those primary sources and cite them directly, every such figure on this site is attributed as "per eesel AI's June 2026 review" or "per FavTutor's August 2026 review", including the underlying study name.
Still being verified, and therefore never estimated: the primary-study links above; free-tier terms and API availability for most tools; and per-tool false-positive measurements (we have not run our own detection tests — self-run accuracy numbers without a published labeling method are exactly what this site criticizes). Weighted index scores publish when this pass completes; until then every ledger entry shows index_score as pending and the 2026 edition is marked PROVISIONAL.
Disclosure
Detector Bench is part of a network of independent review publications. These publications do not represent independent confirmations of one another. The full list of related publications is in About.