Can AI Detectors Detect Claude and DeepSeek? Our ZeroGPT Test
We wrote 60 short texts with Claude and DeepSeek and ran them through ZeroGPT: 58 were flagged as AI. Here are the results by model and by type of writing, what other detectors say they cover, and what changed after a rewrite.
Yes, in our test, easily. ZeroGPT flagged 58 of 60 short texts written by Claude and DeepSeek as AI-generated, with average AI scores around 90%. Turnitin and GPTZero also list Claude and DeepSeek among the models they say they detect.
How we tested
- Writers: 30 texts by Claude (Claude Opus 5.5) and 30 by DeepSeek (deepseek-flash), generated in October 2026.
- Request: the same for every text: "Write a [type] of about 250 words about [topic]. Include these details exactly as written: [facts]. Return only the text."
- Types: academic essays, research abstracts, work emails, business reports, blog articles and personal statements, five per model each.
- Detector: ZeroGPT's detection API, one check per text, nothing edited by hand.
Results by model
| Model | Texts | Average AI score | Median | Flagged as AI |
|---|---|---|---|---|
| Claude | 30 | 90.2% | 95.5% | 29 of 30 |
| DeepSeek | 30 | 87.9% | 92.7% | 29 of 30 |
The two models were nearly indistinguishable. Each had a single text that ZeroGPT read as human, and both were work emails.
Results by type of writing
| Type | Average AI score (both models) |
|---|---|
| Academic essay | 95.8% |
| Personal statement | 94.1% |
| Research abstract | 93.9% |
| Business report | 92.3% |
| Blog article | 89.7% |
| Work email | 68.4% |
Formal, structured writing scored highest. Emails, which are short and conversational, scored lowest, and one Claude email scored 0%.
What other detectors say they cover
- Turnitin lists models its English detector can detect, including several Claude versions (Haiku, Sonnet and Opus), DeepSeek V3.2 and Gemini models. It updates the list with each release (Turnitin FAQ).
- GPTZero says it detects AI content from Claude, ChatGPT and Gemini, and its homepage FAQ also names DeepSeek (GPTZero).
These are the vendors' own statements; we tested only ZeroGPT.
What changed after a rewrite
We then rewrote every text once in RewriteMax's Balanced mode and checked again on ZeroGPT:
| Model | AI score before | AI score after | Read as human after |
|---|---|---|---|
| Claude | 90.2% | 10.3% | 30 of 30 |
| DeepSeek | 87.9% | 14.4% | 30 of 30 |
More on the method and the full results is on the ZeroGPT test page.
Limits of this test
- One detector. Other detectors use different models and may score the same texts differently.
- Short texts. About 250 words each; longer documents may behave differently.
- One run. Each text was generated and checked once.
- Detectors make mistakes. A 2023 study found seven detectors flagged non-native English essays as AI 61.3% of the time on average (Liang et al., Patterns). A score is a signal, not proof of who wrote something.
Common questions
Can GPTZero detect Claude?
GPTZero says it detects Claude. We have not tested GPTZero ourselves.
Is Claude harder to detect than ChatGPT?
We did not test ChatGPT. Between Claude and DeepSeek, ZeroGPT's results were almost the same.
Does editing AI text change the score?
In our test, a full rewrite did: the average ZeroGPT score fell from about 89% to 12%. Light edits keep more of the original structure.
*Test run on October 22, 2026. Vendor facts checked the same day.*