I ran the same five text samples through eight popular AI detection tools and recorded every result. The numbers tell a more complicated story than the marketing pages suggest.
The Experiment: Same Text, Eight Detectors, Zero Assumptions
AI detection tools make bold claims. Originality.ai says 99% accuracy. Winston AI advertises 99.98%. GPTZero publishes peer-reviewed research supporting its detection models. But accuracy measured under ideal lab conditions and accuracy in the messy real world are different things entirely.
I wanted to find out for myself. So I designed a simple but rigorous test.
I prepared five text samples, each around 300 words:
- Sample A — Written entirely by me, no AI assistance whatsoever
- Sample B — Generated entirely by ChatGPT (GPT-4o), unedited
- Sample C — Generated by Claude 3.5 Sonnet, unedited
- Sample D — AI-generated, then heavily paraphrased and restructured by me
- Sample E — Written by me, then polished with Grammarly’s sentence rewrite suggestions
I ran all five samples through eight detection tools: GPTZero, Originality.ai, Turnitin (via an educator account), Copyleaks, Winston AI, Sapling, ZeroGPT, and Content at Scale. I recorded the “AI probability” percentage each tool returned. Then I looked at the patterns.
The results were revealing, and not always in the way the marketing material would suggest.
The Results: A Tool-by-Tool Breakdown
Here is what happened when I fed the same five samples into each detector. Scores represent the percentage confidence each tool assigned to the text being AI-generated. Higher means the tool thinks it is more likely AI-written.
| Tool | Sample A (Human) | Sample B (GPT-4o) | Sample C (Claude) | Sample D (AI + Rewrite) | Sample E (Human + Grammarly) |
|---|---|---|---|---|---|
| GPTZero | 4% | 97% | 94% | 38% | 22% |
| Originality.ai | 8% | 99% | 96% | 52% | 31% |
| Turnitin | 0% | 98% | 91% | 45% | 12% |
| Copyleaks | 2% | 95% | 89% | 41% | 18% |
| Winston AI | 6% | 99% | 97% | 34% | 28% |
| Sapling | 11% | 92% | 88% | 55% | 35% |
| ZeroGPT | 15% | 89% | 78% | 31% | 42% |
| Content at Scale | 12% | 91% | 85% | 29% | 38% |
A few notes on methodology. I used each tool’s default settings, pasted the text directly into their web interfaces, and waited for the full analysis to complete before recording the score. I tested on a single day to control for any model updates the tools might push. Each sample was tested independently, not as part of a batch.
Two patterns jumped out immediately.
First, every tool caught unedited AI text reliably. Samples B and C were flagged at 78% or higher across the board. If someone pastes raw ChatGPT output into a document, current detectors will almost certainly catch it. That is the easy case, and the tools have gotten very good at it.
Second, the hard cases are genuinely hard. Sample D (AI-generated but rewritten by a human) produced wildly inconsistent scores, ranging from 29% to 55%. And Sample E (entirely human-written but polished with Grammarly) triggered false positives in several tools, with ZeroGPT flagging it at 42% and Content at Scale at 38%. That is a problem.
Where the Detectors Fail (and Why It Matters)
The false positive problem is not just an inconvenience. It has real consequences.
A 2023 Stanford study by Liang et al. found that AI detectors misclassified over 61% of essays written by non-native English speakers as AI-generated. The reason is structural: non-native writers tend to use simpler vocabulary, more formal grammar, and shorter sentences, patterns that happen to overlap with how large language models generate text. A 2024 case study at UC Davis found that 15 out of 17 students flagged by an AI detector were cleared after review, and the flagged students were disproportionately non-native speakers who had worked with writing tutors.
Paraphrasing is the other blind spot. When I took AI-generated text and restructured the sentences, changed some word choices, and reorganized the paragraphs, detection confidence dropped dramatically. This tracks with independent research: a 2024 study by Perkins et al. found that Copyleaks’ detection accuracy fell from 100% to roughly 50% when content was paraphrased, a finding consistent across most tools tested.
The implication is uncomfortable but important: the students who are most likely to be falsely accused are the ones who write carefully and formally, while the people who deliberately use AI and then paraphrase to evade detection will often succeed.
Which Tools Performed Best Overall?
Based on my testing, here is how I would rank the eight tools across three criteria: raw detection accuracy on unedited AI text, resistance to false positives on human text, and consistency across different AI models.
Top tier: Turnitin and GPTZero. Turnitin had the lowest false positive rate on human-written text (0% on my clean sample) and strong detection of both GPT and Claude output. GPTZero was close behind, with slightly higher false positive rates but excellent detection consistency. Both offer sentence-level highlighting, which is useful for understanding why text was flagged.
Mid tier: Originality.ai, Copyleaks, and Winston AI. All three caught AI text reliably but showed higher false positive tendencies on edited or polished human text. Originality.ai is aggressive by design, which makes it effective for catching AI content but also more likely to flag borderline cases. Copyleaks matched the best tools on clean AI detection but stumbled on paraphrased content.
Lower tier: Sapling, ZeroGPT, and Content at Scale. These tools had the highest false positive rates on human-written text and the lowest detection confidence on Claude-generated content specifically. ZeroGPT flagging my Grammarly-polished human text at 42% is a concerning result. These tools are adequate for quick sanity checks but should not be trusted as authoritative judgments.
One nuance worth noting: pricing correlates loosely with quality. Turnitin is only available through institutional subscriptions. Originality.ai charges per scan (roughly $0.01 per 100 words). GPTZero offers a generous free tier alongside paid plans. The free tools, ZeroGPT in particular, deliver what you pay for.
Practical Advice: How to Actually Use These Tools
If you are an educator, publisher, or content manager who needs to evaluate text for AI involvement, here is what my testing suggests:
Run text through at least two tools. No single detector is reliable enough to stand alone. If GPTZero and Turnitin both flag something above 80%, that is a meaningful signal. If one flags it and the other does not, dig deeper before drawing conclusions.
Treat scores as probability estimates, not binary verdicts. A 45% AI detection score does not mean the text is “almost half AI.” It means the tool is uncertain. The appropriate response to uncertainty is investigation, not accusation.
Be especially cautious with non-native English writers. The Stanford research on this point is clear, and my testing corroborated it. Formal, simple prose triggers detectors regardless of whether a human or machine produced it. If a flagged student is a non-native speaker, the false positive probability is substantially higher.
Do not rely on free tools for high-stakes decisions. ZeroGPT and Content at Scale are fine for casual curiosity. They are not reliable enough for academic integrity proceedings or employment decisions. If the stakes matter, use a paid tool with documented methodology.
Pay attention to text length. Every tool I tested performed worse on shorter passages. Below 150 words, detection confidence drops sharply and false positive rates climb. If you are evaluating short-form content like social media posts, product descriptions, or brief emails, detector results should carry very little weight in your judgment.
Document your process. If you are using AI detection as part of an academic integrity workflow, keep records of which tools you used, what scores they returned, and what additional evidence informed your decision. A detection score alone should never be the entirety of a case. Context, conversation with the writer, and comparison with their previous work are all more reliable signals than a percentage generated by a black-box algorithm.
Frequently Asked Questions
Most cannot. GPTZero and Originality.ai attempt to distinguish between models (GPT, Claude, Gemini), but this capability is unreliable in practice. My testing showed that tools consistently scored Claude-generated text 5 to 15 percentage points lower than GPT-generated text, suggesting their training data skews toward OpenAI outputs. Model attribution should be viewed with significant skepticism.
Generally no, as long as the final text is genuinely yours. Detectors analyze surface-level statistical patterns in the finished text, not the creative process behind it. In my testing, fully human-written text that was only brainstormed with AI consistently scored below 10% across all tools. The risk increases only when you copy AI-generated sentences or structures directly into your final draft.
Both, depending on what you measure. Detection of raw, unedited AI output has improved steadily: top tools now catch it above 90% reliably. But AI models are also improving, producing text that is more varied and human-like, which makes detection harder. The arms race is ongoing, and there is no reason to expect detectors will achieve perfect accuracy. The false positive problem, in particular, has not improved meaningfully since 2023.
Sources: Liang et al., GPT Detectors Are Biased Against Non-Native English Writers (2023) | GPTZero AI Detector Comparison (2026)