AI detectors compared: what each one is actually good at
Last updated: May 2026
A note before we start. We make an AI detector. So you'd expect us to argue ours is the best. We aren't going to. Partly because no detector deserves to be called "the best." Partly because pretending otherwise would be the kind of overconfidence that's caused half the problems in this space.
If you've searched for "best AI detector," you've probably noticed something strange. Every article ranks a different tool at number one. Often, that ranking is the tool that paid for the article. The actual answer, which nobody ranking detectors wants to give you, is that there isn't a single best. Each tool has strengths. Each has limits. The right one depends entirely on what you're using it for.
So this isn't a ranking. It's a guide to what each of the major detectors is genuinely good at, what they each fail at, and how to choose between them with your eyes open. All the accuracy numbers come from independent studies or the tools' own published benchmarks. We've linked the sources as we go.
The honest summary first
If you only read one paragraph. The strongest detector on raw, unedited AI text is probably Pangram, an academic-grade tool that ranked first in a peer-reviewed 2025 University of Chicago Booth study and again in the COLING 2025 RAID benchmark.1 Among the commonly used commercial tools, GPTZero has the strongest published false-positive control. Originality.ai tends to lead on raw recall. Turnitin wins on institutional integration but has documented issues with non-native English writers. Copyleaks is the broadest enterprise tool but the weakest on edited or paraphrased content. ZeroGPT is the free option most people try first, with about the accuracy you'd expect from a free option.
Every one of them gets things wrong. That's the part the marketing pages don't lead with.
The big finding nobody quotes anymore
Before we get into individual tools, it's worth being clear-eyed about the field. The most important paper on AI detection is still the 2023 Stanford HAI study by James Zou and colleagues, published in Patterns.2 It tested seven AI detectors against 91 TOEFL essays written by Chinese students. The results were grim. More than 61% of the human-written essays were flagged as AI. Nineteen of them were unanimously flagged by all seven detectors. Ninety-seven of every hundred essays were flagged by at least one detector.
The vendors didn't like this finding. Several have published rebuttals arguing the sample size was small or the data was outdated. Some of those rebuttals are reasonable. But the core point, that detectors systematically misread non-native English writing as AI, has held up across multiple later studies. If you're an educator, this is the single most important fact in this article. We'll come back to it.
The tools, one at a time
GPTZero
The detector you've probably heard of. It started as a university project by Edward Tian in early 2023 and grew into one of the most widely used tools in the field.
Where it leads: Published the strongest false-positive rate of any major commercial detector. On the RAID benchmark, which is a multi-domain academic dataset of news articles, reviews, social media, and books, GPTZero reported 95.7% recall at a 1% false-positive rate.3 In its own 3,000-sample test against Copyleaks and Originality.ai, it claimed 99.3% accuracy and a 0.24% false-positive rate.4 Take any vendor-published benchmark with a grain of salt, but GPTZero's transparency about its methodology is unusually good. They actually publish the version of the model used in each benchmark, which most competitors don't.
Where it falls short: Independent reviews have noted that performance drops sharply on humanised or paraphrased AI text. A 2025 review in Skywork found that minor humanisation cut sensitivity by roughly 70%.5 It also struggles with very short writing samples.
Best for: Educational settings where false positives matter more than catching every last cheater. The team's caution about overreliance is genuinely refreshing.
Originality.ai
Built specifically for publishers, SEO agencies, and content marketplaces. Launched 2022.
Where it leads: Consistently top scores on raw, unedited AI text. In a 2024 meta-analysis of 14 studies they compiled themselves, Originality.ai reported 97% overall accuracy across GPT-3.5, GPT-4, and human samples.6 Independent comparisons against Copyleaks generally agree that Originality.ai has higher raw recall.
Where it falls short: Higher false positive rate than GPTZero. Reviewers have noted that its strict sensitivity flags polished human writing more readily. That's fine if you're a content agency checking freelancer output. It's not fine if you're an educator assessing student work. Also worth noting: their published benchmarks tend to compare against older versions of competitors, which inflates their lead.
Best for: Content publishers, SEO teams, and agencies where the cost of a false negative (paying for AI-written content) is higher than the cost of a false positive (asking a writer to revise).
Turnitin
The dominant academic-integrity platform, used by over 16,000 institutions globally. Added AI detection on April 4, 2023.
Where it leads: Distribution. Turnitin is embedded in nearly every major LMS (Canvas, Moodle, Blackboard), making it the default for most academic settings. Their stated accuracy is 98% with less than 1% false-positive rate on documents containing more than 20% AI-generated content.7
Where it falls short: This is the tool with the most public false-positive incidents documented in journalism. A Markup investigation in 2023 followed cases of international students wrongly flagged.8 A Washington Post review reported false-positive rates up to 50% in limited testing. Turnitin has openly admitted it intentionally misses about 15% of AI-generated text to avoid flagging human writing, which is a tradeoff the company believes is correct, but one worth knowing about.9 Their own release notes confirm that low-confidence scores (1-20%) had elevated false-positive incidence.
Best for: Institutions that already use Turnitin. If you're being checked by Turnitin, it's because of distribution. Not because it's measurably more accurate than alternatives.
Copyleaks
Started in 2015 as a plagiarism platform. Added AI detection in 2023.
Where it leads: Multilingual coverage (30+ languages), enterprise-grade integrations, and a plagiarism detection foundation that makes it useful as a combined tool. Strong on clean, unedited AI text, broadly comparable to Originality.ai.
Where it falls short: Independent testing consistently flags Copyleaks as the weakest of the major detectors on paraphrased or hybrid (part-human, part-AI) content. GPTZero's own comparison reported Copyleaks misclassifying about 1 in 20 human documents as AI.4 That rate is uncomfortably high for educational settings.
Best for: Multinational enterprise teams that need detection across many languages and integration with existing content workflows.
ZeroGPT
A free AI detector that gets a lot of traffic because it doesn't paywall basic checks.
Where it leads: Accessibility. No signup, instant results, no cost. For casual checks of unedited AI text, it works reasonably well.
Where it falls short: Almost everywhere else. Independent reviews consistently rank ZeroGPT as the weakest of the popular tools. It has high false-positive rates on polished human writing. Performance degrades fast on anything that's been edited. Its scoring isn't well calibrated, so the percentages it returns don't mean what you'd intuitively expect them to mean.
Best for: Quick, low-stakes checks where you mostly want to know if something is obviously AI. Don't use it for anything that affects someone's grade or job.
The comparison table
If you want the at-a-glance version:
| Tool | Best for | Main weakness |
|---|---|---|
| GPTZero | Education, low false positives | Drops on humanised text |
| Originality.ai | Publishers, content agencies | Stricter, more false positives |
| Turnitin | Institutional use (already deployed) | Documented bias issues |
| Copyleaks | Enterprise, multilingual | Weakest on edited text |
| ZeroGPT | Quick free checks | Lowest reliability overall |
What's actually true about all of them
Three things are worth taking away from this:
1. They all degrade on edited text. Every major detector loses significant accuracy the moment AI output has been paraphrased, run through a humaniser tool, or edited by hand. Multiple 2026 comparisons confirm this. It isn't any individual tool's problem. It's the core limitation of the whole field.10
2. The non-native English speaker problem is real. The Stanford finding has been replicated repeatedly. Detectors confuse simpler vocabulary and tighter sentence structure with AI output, because that's also what AI tends to produce. If you're using a detector and the writer learned English as a second language, weigh the result with extreme caution.
3. The percentages mean less than they look like. A "97% AI" score from one detector might correspond to "30% AI" from another on the same text. These aren't probabilities in the statistical sense. They're model outputs, not calibrated estimates. Treat them as relative signals within a single tool, not as universal truths.
Where Telltale fits
Since you'll wonder. Telltale is the tool we make. It takes a different angle from everything above. We don't try to compete on percentages. Instead, the tool surfaces the specific patterns it found in the text and explains each one, drawing on Wikipedia's editorial guide to AI writing. The score is there, but the evidence is the point.
This is honest about its limits in a way commercial detectors usually aren't. We can't catch sophisticated AI use better than Pangram or GPTZero. What we can do is show you our reasoning, so you can argue with the result. We think that matters more than another percentage. We've written separately about why we're sceptical of percentage-based detection in general.
We'd genuinely recommend GPTZero for educational use cases where you need a single tool with low false positives. We'd recommend Originality.ai for content agencies. We'd recommend our own tool if you want transparent reasoning rather than a verdict. None of these are universal answers.
Telltale shows you the specific patterns it found in the text, with explanations for each. Free, no signup, no AI used in the analysis itself.
How to choose, practically
If you're trying to pick a detector for a specific use case, here's the actually useful version:
- If you're a teacher. Use the one your institution has already licensed (probably Turnitin), but use it as a flag for further investigation, never as a verdict. Cross-check with at least one other tool if the stakes are high. Pay extra attention if the student is a non-native English speaker.
- If you're a content publisher. Originality.ai or Copyleaks. Both are built for your workflow.
- If you want a free check. GPTZero's free tier is the strongest of the free options.
- If you want to understand why something was flagged. Use a tool that shows you its reasoning. This is our pitch, but the same applies to GPTZero's sentence-level highlighting.
- If it really matters. Never rely on one tool. And never rely on a tool alone. A detector score is a flag, not proof.
One last thing
The detector industry has a bad habit of pretending it's solved a problem it hasn't. The honest state of things in 2026 is that detection is useful, imperfect, and getting harder rather than easier as models improve. Anyone selling you certainty is overstating what's possible.
The right approach is to use these tools as signals. Alongside knowing the writer, knowing the context, and being willing to ask before you accuse. That's true of our tool. It's true of every tool listed here. It's the field, not a flaw in any one product.
Sources
- GradPilot, "Turnitin's 15% AI Miss Rate" (citing Jabarian & Imas, University of Chicago Booth, 2025; COLING 2025 RAID benchmark) — gradpilot.com
- Liang, Zou et al., "GPT detectors are biased against non-native English writers," Patterns (2023) — Stanford HAI summary
- GPTZero, "RAID benchmark results" — gptzero.me
- GPTZero, "GPTZero vs Copyleaks vs Originality" — gptzero.me
- Skywork, "GPTZero Review October 2025" — skywork.ai
- Originality.ai, "AI Detection Accuracy Studies Meta-Analysis" — originality.ai
- Leap, "Turnitin AI Detection Accuracy 2026" — tryleap.ai
- García Mathewson, "AI Detection Tools Falsely Accuse International Students of Cheating," The Markup (2023) — themarkup.org
- Popular AI, "These Turnitin false positives in 2025 and 2026" — popularai.org
- Phrasly, "Is Copyleaks Accurate? 2026" — phrasly.ai