Accuracy Claims vs Independent Tests
Vendor pages in this market display accuracy figures from 99% to 99.98%. Independent records on the same tools run lower — one 99% claim measured at 63.77% in a head-to-head study. This page keeps the two kinds of number in separate columns, forever, because merging them is how the market advertises its way past the evidence.
Why this page exists
Our October 2026 sampling of "best AI detector" sources found vendors ranking themselves: at least four of the top-10 results were vendors' own pages or blogs, three of them displaying accuracy claims between 99% and 99.98% that no independent record in our fact base corroborates. The one position nobody had taken was the boring one — print both kinds of number and label them. That is this page, and the rule behind our 2026 ranking.
The recorded columns
| tool | vendor claim (their page) | independent record (our sources) |
|---|---|---|
| GPTZero | 99% · 99.9% · 95.7% · 96.5% — four figures on one page | 95.7% on the RAID benchmark; 63.77% in a head-to-head study (per eesel AI, Jun 2026) |
| Winston AI | 99.87% · 10M+ users (vendor-stated) | **99% standardized accuracy** (May 2026 peer-reviewed study, per FavTutor) |
| Originality.ai | not on record in our crawl | 97% (empirical study, per eesel AI) · 98% (peer-reviewed table, per FavTutor) |
| Pangram | 99.98% · "3rd-Party Verified" (vendor label) | none found in this run |
| Turnitin | not on record in our crawl | 61.3% of non-native English essays flagged as AI (Stanford-linked study, per eesel AI); disabled by five named universities |
Provenance note: until each cited study is traced to the underlying paper, we attribute every figure to the review that reported it — the trace is an open item in our data coverage, not a silent shortcut.
How the gap becomes a score
The gap is 20% of every tool's total on this site. A vendor can win it back the honest way: publish a number a third party can measure, and let the two columns meet.
FAQ
Why do vendor accuracy claims and independent results differ?
Because they measure different things under different conditions: a vendor number is self-reported on the vendor's own terms, an independent number comes from someone else's sample. The recorded spread is wide — one vendor page displays four figures from 95.7% to 99.9%, one 99% claim was measured at 63.77% in a head-to-head study cited by eesel AI, and Pangram states 99.98% with no peer-reviewed support found in our crawl.
Which vendors have independent numbers on record here?
In this run: GPTZero at 95.7% on the RAID benchmark and 63.77% in a head-to-head study (both per eesel AI's June 2026 review), Originality.ai at 97% in an empirical study and 98% in the peer-reviewed table FavTutor cites, and Winston AI at 99% standardized accuracy in that same May 2026 peer-reviewed study. Entries without independent numbers — Pangram, ZeroGPT and others — say so on their panels instead of borrowing someone else's figure.
How should a school or publisher read an accuracy claim?
Ask three questions: is the number vendor-claimed or independently measured; what was the sample; and what does it say about false positives on human writing — the documented failure mode that led five named universities to disable one vendor's detector. An accuracy figure with no false-positive record is half an answer.