A ChatGPT detector scores text, not people. Paste in a paragraph. The tool returns a probability that a language model wrote it. That number is a statistical estimate. It is never a verdict.
Here is how the tools work, where they fail, and which ones deserve your time.
Section 1: Top AI Detectors Technical Matrix
Vendor-claimed accuracy runs high. Independent tests run lower. Treat the figures below as marketing ceilings, and confirm current limits on each vendor’s pricing page before you commit.
| Tool | Claimed Accuracy | Models Targeted | Core Focus | Free Tier |
|---|---|---|---|---|
| GPTZero | ~99% (vendor claim) | ChatGPT/GPT-series, Claude, Gemini, Llama | Education, sentence-level analysis | Limited monthly word allowance |
| Originality.ai | ~99% (vendor claim) | GPT-series, Claude, Gemini, plus paraphrase detection | Publishers, SEO agencies | No meaningful free plan; pay-per-credit |
| Copyleaks | ~99% (vendor claim) | GPT-series, Claude, Gemini, multilingual | Enterprise, LMS integrations | Small number of free scans |
| Winston AI | ~99% (vendor claim) | GPT-series, Claude, Gemini | Publishers, document scoring | Short free trial in credits |
Model coverage shifts fast. New releases such as GPT-5-class systems force vendors to retrain classifiers. Check the changelog before you trust any tool on text from a brand-new model.
Section 2: What Is a ChatGPT Detector and Why Does It Matter?
It is a probabilistic classifier. Engineers train it on huge sets of human writing and machine writing. Then it learns the statistical gap between them.
The output is a likelihood. A score of 87% AI means the text resembles machine samples the classifier saw in training. It does not mean the author is proven. It does not mean 87% of the words came from a model.
Why it matters:
- Editors vet freelance submissions
- Teachers screen essays
- Publishers protect search rankings
- Hiring teams check cover letters
The stakes are uneven. A wrong score costs a blogger a rewrite. It can cost a student a disciplinary hearing. Use the number as a prompt for a conversation, never as the final word.
Section 3: The Underlying Science: Perplexity and Burstiness Explained
Perplexity measures surprise. Feed a sentence to a reference language model. Ask how predictable each word was. Predictable words give low perplexity. Surprising words give high perplexity.
Language models favor the likely word. Generation works by sampling from probability distributions. The system leans toward high-probability choices. The result reads smooth, safe, and statistically expected. That is low perplexity.
Burstiness measures rhythm. It tracks how much sentence length and structure vary across a passage. Machine text tends to settle into an even cadence. Medium sentence. Medium sentence. Medium sentence. That is low burstiness.
Humans write unevenly. We ramble, then snap. A twelve-word sentence sits next to a three-word one. We pick odd verbs. We digress. Our text spikes in both perplexity and burstiness.
A simple mental model:
- Low perplexity + low burstiness → likely machine
- High perplexity + high burstiness → likely human
- Mixed signals → the gray zone, where most errors live
Modern detectors go further. Many now use fine-tuned transformer classifiers rather than raw perplexity alone. They also look at token-level patterns, vocabulary distribution, and stylistic fingerprints. Perplexity and burstiness remain the clearest way to understand the logic.
Section 4: Deep Scans and False Positives
False positives are the central flaw. Detectors misread any writing that happens to look statistically “predictable.”
Non-native English writers get hit hardest. Stanford researchers published a widely cited 2023 study showing that several detectors disproportionately flagged essays by non-native English speakers as AI-generated. Limited vocabulary and simpler syntax lower perplexity. The classifier reads that as machine output.
Formal business writing triggers flags too. Legal memos, compliance documents, technical manuals, and policy briefs all follow rigid conventions. Rigid conventions produce predictable text. Predictable text scores as AI.
Edited AI text can slip through. A human who rewrites a draft heavily may drop the score below the threshold. False negatives are as real as false positives.
Sentence-by-sentence highlighting helps. Better tools color each sentence by its probability of being machine-written. Use that map to diagnose a flag:
- Isolated red sentences usually mean boilerplate or a quoted definition
- Clusters of red deserve a closer read
- A mostly green document with one red patch is rarely worth an accusation
Check the version history. If a document lives in Google Docs, its revision log shows drafting over time. That evidence often beats any detector score.
Section 5: Best AI Detection Tools Tested
GPTZero is built around education. It popularized the perplexity and burstiness framing. Its sentence-level highlighting makes false-flag triage easy. It also offers a Google Docs integration and a Chrome extension aimed at teachers who want to inspect writing history. Best for: classroom and academic review. Watch for: occasional over-flagging of formal prose.
Copyleaks targets institutions. It supports many languages and plugs into learning management systems and enterprise workflows through an API. A Chrome extension is available for quick checks. It pairs AI detection with plagiarism scanning in one report. Best for: schools and organizations that need scale and multilingual coverage. Watch for: a free tier that runs out fast.
Quillbot’s AI detector is the low-friction option. It is free to use in a browser, needs little setup, and gives a quick read on whether a passage looks machine-written. It is a fast first pass, not a forensic tool. Best for: writers who want a quick self-check before submitting. Watch for: less granular diagnostics than the enterprise platforms.
Integration availability changes. Confirm current extensions and add-ons in each tool’s documentation before you build a workflow around them.
The Practical Takeaway
Run two detectors, not one. Disagreement between tools signals uncertainty. Agreement is still not proof.
Pair the score with context. Compare the text with the author’s earlier work. Review drafts. Ask a question about the content.
Never punish on a number alone. Detectors estimate probability. Evidence of authorship comes from process, not from a percentage.