AI detectors promise a simple verdict: human or machine. The reality is messier — independent benchmarks show accuracy swinging between 60% and 95% depending on the tool, the text, and how much it was edited. This guide shows how the tools work, which ones are worth using, a workflow that gives you the most reliable answer possible, and how to avoid the false positives that have embarrassed schools and publishers.
How AI detectors actually work
Most detectors are not looking for anything as obvious as a watermark. They measure statistical patterns in the writing:
- Perplexity — how "surprised" a language model is by each word. AI-generated text is typically low-perplexity: it picks the most predictable next word, so it reads as safe and smooth. Human writing spikes and dips in surprising places.
- Burstiness — how much sentence length and structure vary. Humans write a long, winding sentence, then a short one. AI tends toward uniform, medium-length sentences with similar rhythms.
- Stylometric signals — vocabulary diversity, transition-word frequency, and structural tics that cluster differently in machine output.
Some providers also experiment with watermarking — subtle statistical biases baked into generated text. OpenAI reportedly built such a system, then shelved it after concluding that simple paraphrasing would defeat it.
The key implication: detectors score probability, not proof. Every output is a guess with a known error rate.
The hard truth: no detector is reliable on its own
Before picking a tool, internalise what independent research actually found:
- On the RAID benchmark (a large independent evaluation covering more than 6 million AI generations across 11 models and 11 adversarial attacks), Originality.ai scored the highest overall accuracy at 85% at a 5% false-positive rate, and reached 96.7% on paraphrased AI content. GPTZero was unusually robust to adversarial attacks but its base accuracy at the same false-positive rate trailed at 66.5%.
- A 2026 peer-reviewed study (Hadra et al.) testing 192 texts found 61% overall accuracy for Turnitin and 69% for Originality.ai. Both struggled most with mixed human-and-AI writing.
- The false-positive problem is real and documented. A widely cited study (Liang et al., 2023) found an average false-positive rate of about 61% on TOEFL essays across seven detectors — formal, constrained English by non-native speakers statistically resembles AI output. After the essays were lightly polished, the rate fell to 11.6%.
- Paraphrasing is the universal solvent. Krishna et al. (NeurIPS 2023) showed that automated paraphrasing dropped DetectGPT's accuracy from 70.3% to 4.6% at a constant 1% false-positive rate. On heavily edited or "humanised" AI text, most tools collapse to single-digit detection rates.
Several universities — Vanderbilt, Michigan State, Northwestern, and UT Austin in 2023, Waterloo in 2025 — switched off Turnitin's AI detection feature over exactly these reliability concerns. OpenAI discontinued its own AI text classifier for similar reasons. Treat every detector score as a screening signal, never a verdict.
The tools worth using in 2026
With those caveats, some tools are clearly more useful than others. Here are the ones with independent backing:
1. GPTZero — best free starting point
GPTZero offers a free tier of 10,000 words per month, which covers most casual use. It gives sentence-level highlighting so you can see exactly which passages look machine-like, and its Writing Replay feature records the writing process as proof of authorship. It performed best in a 2026 study on non-native-simulated text (90% accuracy) and is unusually robust to adversarial tricks. Weakness: it struggles with paraphrased AI content compared with Originality.ai.
2. Originality.ai — best for publishers and marketers
Credit-based pricing (no free tier), but it bundles plagiarism checking, fact-checking, and readability scoring into every scan. It topped the independent RAID benchmark overall and caught 96.7% of paraphrased AI content — the strongest paraphrase detection available. Best fit: content teams checking freelance submissions at scale.
3. Copyleaks — lowest false-positive rate
In independent comparisons, Copyleaks consistently posts the lowest false-positive rate of the major tools, which matters if being wrong about human text is your biggest risk. It offers a small free tier (1,200 words/month) and API access.
4. Winston AI — best for documents and media
Beyond text detection, Winston AI can scan images with OCR and includes deepfake detection. Its free trial covers 2,000 words with no credit card. Handy for educators dealing with scanned or handwritten submissions.
Free quick checks: ZeroGPT and Crossplag are free and fine for a rough first pass, but independent tests rank them well below the tools above. Use them as a sniff test only. Skip for general use: Turnitin is institution-only (its parent company says it deliberately leaves about 15% of AI writing unflagged to keep document-level false positives under 1%) — treat its score as triage, not evidence.
A reliable detection workflow: the 6-step method
Here is the practical process that gets you closest to a reliable answer:
Step 1 — Check the text length. Detectors are unreliable on short text. Under 200–300 words, even the best tools produce noisy results. If the passage is short, combine it with more of the author's writing before drawing conclusions.
Step 2 — Run at least two different detectors. No single tool sees the whole picture. GPTZero plus Originality.ai, or GPTZero plus Copyleaks, is a solid pair because they use different detection approaches. Run the same text through both.
Step 3 — Require agreement. One tool flagging AI at 60% confidence means little. Require both tools to flag the same passages before treating the result as meaningful.
Step 4 — Read the sentence-level highlights. All the major tools highlight which sentences look machine-generated. Real AI-assisted text usually shows clusters of flagged sentences (whole paragraphs), not scattered single sentences. A human essay with five isolated flagged sentences is far less suspicious than one where three consecutive paragraphs light up.
Step 5 — Account for the known bias triggers. Non-native speakers, formal academic writing, technical documentation, and text polished with grammar aids all inflate AI probability scores. If any of these apply, discount the score heavily.
Step 6 — Look for process evidence, not just the score. Version history, drafts, and tools like GPTZero's Writing Replay tell you far more than any percentage.
Why paraphrased text beats every detector
If you need to understand the ceiling of these tools, this is it. Automated paraphrasing — running AI text through a second model or a paraphraser like QuillBot — reshapes the statistical fingerprints detectors look for. The landmark NeurIPS 2023 study found paraphrasing dropped one detector's accuracy from 70.3% to 4.6%, and commercial "humaniser" tools exist specifically to exploit this.
Practical consequence: detectors work best on raw, unedited AI output and worst on anything a person has touched or paraphrased. If you are a publisher checking freelancer work, know that a determined writer can evade every detector on the market — which is why contracts and subject-matter spot-checks matter as much as scanning.
For teachers and managers: policy beats policing
The universities that disabled AI detection did not give up on academic integrity — they shifted strategy. Consider the same:- Never base a disciplinary decision on a detector score alone. Scores understate and overstate in the same document.
- Set expectations in writing. State clearly what AI assistance is allowed (brainstorming, grammar checks) and what is not — most disputes come from ambiguous policy, not sophisticated cheating.
- Design AI-resistant assessment. Process documentation, in-class drafting, and personalised prompts are all harder to fake than a take-home essay.