Can AI detectors detect Claude, Gemini and DeepSeek, not just ChatGPT?

Updated October 31, 2026 · 3 min read

Short answer

Mostly, yes. The leading detectors are trained on text from many AI models, and independent tests in 2025 and 2026 found they catch plainly prompted Claude, GPT and Gemini text almost every time. In Epoch AI's July 2026 test of Claude Opus 4.8, GPT-5.5 and Gemini 3.1 Pro, no detector missed more than 2% of any model's text.

The gaps show up when the AI imitates a real writer's style, with less common models, and with some tools: in that same test about one in five of Gemini 3.1 Pro's style-imitation passages got past Pangram, and in 2025 Originality.ai missed up to 42% of Claude Opus 4 text at its strictest setting.

What do independent tests show for each AI model?

Share of AI text each detector missed (lower is better). Model and detector versions are the ones tested at the time.

Model	Pangram	GPTZero	Originality.ai	Study
GPT-4.1	1.3%	0.25%	16.1%	Chicago Booth, 2025 (0.5 threshold)
Claude Opus 4	2.2%	2.9%	37.3%	Chicago Booth, 2025
Claude Sonnet 4	1.5%	0.85%	18.8%	Chicago Booth, 2025
Gemini 2.0 Flash	2.7%	1.5%	2.7%	Chicago Booth, 2025
Claude Opus 4.8, imitating an author	5%	8%	12%	Epoch AI, 2026
GPT-5.5, imitating an author	5%	10%	27%	Epoch AI, 2026
Gemini 3.1 Pro, imitating an author	20%	14%	14%	Epoch AI, 2026

With plain prompts (no style imitation), Epoch AI found that no detector missed more than 2% of text from any of its three models. The imitation results come from giving each model five samples of a real author's work and asking it to write like them.

Hardest single case: Gemini 3.1 Pro imitating scientific writing evaded Pangram 48% of the time and GPTZero 36%. For Originality.ai it was GPT-5.5 scientific writing (39%) (Epoch AI, 2026).
GPT-5, Gemini 3.1 Pro and Claude Opus 4.6 (Graphite, industry study, 2026): Pangram, Copyleaks and GPTZero missed under 1% of GPT-5 and Claude text. Gemini was hardest: 3.5% missed by Pangram and 4.9% by Copyleaks.
DeepSeek and Grok (JAIT, 2026, six texts per model): Copyleaks and Originality.ai scored 100% on ChatGPT-4 Turbo, DeepSeek, Gemini and Grok. Grok was hardest for several tools: Turnitin 80.0%, ZeroGPT 74.5%, Grammarly 38.2%. GPTZero scored 86.7% on DeepSeek and 100% on Grok.
AI-polished text (PeerJ, 2025): GPTZero gave abstracts polished by Gemini 2.0 Pro a mean AI score of 55.50%, versus 19.79% for abstracts polished by ChatGPT o1.

What do the detector companies claim?

These come from each company's own tests or documentation:

Company	Claim	Date
Turnitin	Its English detector covers models including Claude Sonnet 4.5/4.6 and Opus 4.6, DeepSeek V3.2, Gemini 2.5 Pro and Gemini 3/3.1 previews, GPT-4o through GPT-5.4, Grok-4.1, Llama and Mistral. No per-model accuracy published.	Oct 2026
GPTZero	Its 4o model catches over 99% of text from GPT-5.6, Gemini 3.6, Grok 4.5 and Claude 5, with a false positive rate under 0.03%.	Oct 2026
Pangram	Pangram 4 missed 0.34% of 520,000 outputs from 26 models: lowest for Anthropic models (0.195%), highest for Grok 4.3 (0.605%).	Jul 2026
Originality.ai	96 to 99%+ accuracy on DeepSeek V3.2 and V4, depending on its model; 99% on Grok 4.6.	Aug to Oct 2026
Copyleaks	Above 98% accuracy per model on English text from GPT-4/ChatGPT, Claude, Gemini, DeepSeek, Llama and others.	Jun 2025

Why can newer models be harder to detect?

Detectors learn from examples. The RAID benchmark (ACL 2024) found "a clear bias towards domains and models that the detectors have trained on." A model released last week may not be in a detector's training data yet.
Style imitation works. When a model copies a specific writer's voice, its text looks less like generic AI output. That's where every detector in Epoch AI's test lost the most accuracy.
Tools differ by model. The same detector can be strong on one model family and weak on another, as OriginalityAI was on Gemini versus Claude Opus 4 in 2025.

What does this mean for checking a text?

A good detector will usually catch unedited Claude, Gemini, GPT or DeepSeek text of a few hundred words.
Be more cautious with text from very new models, text written to imitate someone, and short passages.
AI Checker Plus checks text from ChatGPT, Claude, Gemini, Llama, DeepSeek, Grok and other popular models. Like every detector, it can lag behind models that were released very recently.
Don't treat any result as proof. What the accuracy studies say.

Questions people also ask

Can Turnitin detect Claude?

Turnitin's FAQ (October 2026) lists Claude Sonnet 4.5 and 4.6 and Claude Opus 4.6 among the models its English detector covers, but it doesn't publish accuracy per model. We found no independent test of Turnitin on Claude text.

Can GPTZero detect Gemini?

Usually. In the Chicago Booth study (2025), GPTZero missed 1.5% of Gemini 2.0 Flash text at a 0.5 threshold. In Epoch AI's 2026 test it missed 14% of Gemini 3.1 Pro text written to imitate a real author's style. GPTZero itself reports over 99% recall on Gemini 3.6 on its own benchmark.

Is Claude harder to detect than ChatGPT?

It depends on the detector. In 2025, OriginalityAI missed 37.3% of Claude Opus 4 text versus 16.1% of GPT-4.1 text at a 0.5 threshold. In Epoch AI's 2026 style-imitation test, Claude Opus 4.8 was the easiest of the three models for Pangram and GPTZero (5% and 8% missed).

Can AI detectors detect DeepSeek?

In a small 2026 study (six DeepSeek texts per tool), Copyleaks and Originality.ai scored 100%, Turnitin 97.7% and GPTZero 86.7%. Originality.ai and Pangram also publish their own DeepSeek results, which are vendor tests.

Sources

By the AI Checker Plus team. We research and draft these guides with help from AI tools, then check every claim against the sources listed above.