Best AI detectors: how to choose one you can trust
We make an AI detector, so we will not rank ourselves first or score other tools on our own tests. Instead, here is what to look for, and the facts each vendor publishes, with links and dates so you can check them.
Five questions to ask any detector
- How often does it call human writing AI? This false-positive rate matters more than headline accuracy, because a false accusation does the most harm.
- Was that measured in your language? A rate measured in English says little about Spanish.
- Does it show its evidence? Highlighted passages let you judge the result instead of trusting a number.
- Does it say when it is unsure? A borderline text should come back as unclear, not as a confident percentage.
- What happens to your text? Unpublished work should not be stored or used for training.
What the main detectors publish
Turnitin
- Who can use it: Sold to schools and universities; no plan for individuals. (Turnitin, checked 2026-11-10)
- Paid price: Institutional licence; a 2025 report put the AI add-on at about $3.05–3.57 per student per year in California. (The Markup, checked 2025-06-26)
- Spanish: Separate Spanish model since 12 September 2024. (Turnitin guides, checked 2026-11-11)
- Length rules: Needs at least 300 words; scores between 1% and 19% are shown only as an asterisk. (Turnitin guides, checked 2026-11-11)
- What it publishes about accuracy: States a false-positive rate under 1% for documents with 20% or more AI writing (English). (Turnitin, checked 2026-11-11)
GPTZero
- Who can use it: Open to anyone; owned by Superhuman (formerly Grammarly) since June 2026. (Superhuman, checked 2026-11-10)
- Free use: Without an account, up to 10,000 characters per scan; about 10,000 words a month with a free account. (GPTZero, checked 2026-11-10)
- Paid price: Premium shown at €18.99 a month in euros. (GPTZero, checked 2026-11-10)
- Spanish: Detects 27 languages; reports 95.7% accuracy and a 5.6% false-positive rate on Spanish. (GPTZero blog, checked 2026-11-10)
- What it publishes about accuracy: States 99.3% accuracy and a 0.24% false-positive rate overall. (GPTZero, checked 2026-11-10)
ZeroGPT
- Who can use it: Open to anyone, no login needed. (ZeroGPT, checked 2026-11-10)
- Free use: Up to 15,000 characters per check without an account. (ZeroGPT, checked 2026-11-10)
- Paid price: Monthly and annual plans, 40% off when billed annually; prices are not shown on the plan page without signing in. (ZeroGPT, checked 2026-11-11)
- Spanish: Interface in six languages including Spanish; the languages it detects are not listed. (ZeroGPT, checked 2026-11-11)
- What it publishes about accuracy: Calls itself "the most Advanced and Reliable" detector; its plan page gives no measured accuracy figures. (ZeroGPT, checked 2026-11-11)
Pangram
- Free use: 2,000 words a day after signing up. (Pangram, checked 2026-11-10)
- Paid price: $20 a month for 300,000 words. (Pangram, checked 2026-11-10)
- Spanish: Publishes a Spanish false-positive rate in its model card. (Pangram, checked 2026-11-10)
Copyleaks
- Free use: Up to 25,000 characters per scan free. (Copyleaks, checked 2026-11-10)
- Paid price: $16.99 a month for about 25,000 words. (Copyleaks, checked 2026-11-10)
- Spanish: Supports 30 languages including Spanish, with explanations in Spanish. (Copyleaks, checked 2026-11-10)
Facts about other products come from their own pages or from the reports linked, on the date shown. Prices and features change; check the source before deciding.
Our measured numbers
Rendered from the measurement file of Engine 2026.11-oculus-v0. We do not measure other vendors' detectors, so there are no competitor accuracy figures on this page.
| English human texts called Likely AI | 1.2 % (0.9–1.7) |
| English AI texts called Likely AI | 89.2 % (87.1–91.0) |
| Spanish human texts called Likely AI | 1.4 % (1.0–2.0) |
| Spanish AI texts called Likely AI | 99.5 % (98.8–99.8) |
What independent studies found about detectors in general
- Seven detectors flagged 61% of 91 human TOEFL essays written by non-native English speakers as AI-written (Liang et al., 2023). (Patterns / arXiv, checked 2026-11-11)
- A test of 14 detection tools found them "neither accurate nor reliable", and easy to fool with paraphrasing (Weber-Wulff et al., 2023). (International Journal for Educational Integrity, checked 2026-11-11)
- An audit by the University of Chicago's Becker Friedman Institute found large differences between commercial detectors, with only one meeting a 0.5% false-positive cap (Jabarian & Imas). (Becker Friedman Institute, checked 2026-11-10)