Skip to content
DeepSeekDetect
EN

How to tell if DeepSeek wrote it

DeepSeek tends to write with many em dashes, evenly sized sentences and a recognisable stock of phrases such as "rather than", "far more" and "a quiet". None of these proves anything alone, since people use every one of them, but clusters of them are worth noticing. The multipliers below compare fully AI-written DeepSeek-family text against human writing in the comparison corpus.

How to read the multipliers

A figure like "5.6×" means the phrase shows up that many times more often in DeepSeek-family text than in human writing of comparable kind. It describes a tendency across many texts, not a rule for one text. A phrase used once in a page says almost nothing.

English tells

Phrases that stand out against human writers

Phrases that separate it from other AI models

"a short" (2.7×), "the problem" (2.5×), "examine" (2.5×), "thing" (2.2×) and "kind of" (1.8×) appear more often than in other AI families, so they hint at DeepSeek rather than AI in general.

Structural markers

Chinese tells

Phrases that stand out against human writers

Phrases that separate it from other AI models

而不是 (3.3×), 那个 (2.7×), 东西 (2.6×), 心里 (2.8×) and 习惯 (2.2×) point toward DeepSeek over other families, a more conversational register than some other models.

Structural markers

"Not X but Y" contrasts at about 3.9× the human rate, and em dashes at about 3.5×.

How Qwen and GLM differ

Qwen leans on formal connectors such as "serves as" (26×), "ensures that" (20×) and "is crucial" (25×) in English. GLM leans on intensifiers such as "entirely" (11×), "ultimately" (10×) and "deeply" (9.4×). Both also use many em dashes. The checker shows these families separately; see DeepSeek versions for the models.

Why phrases are not proof

Treat any list as a reason to look closer, never as a verdict. This is a style signal, and it can point at a family without proving who wrote the text.

What Veread reads beyond phrase lists

Veread, Fileverity's text model, does not count phrases. It reads the whole text and the rhythm and structure behind it, marks the passages that read as AI-written, and compares the style with each model family. Of fully AI-written DeepSeek-family texts, 95.1% were caught, with 0.5% of human texts wrongly flagged; see the accuracy page. For documents and reports, use the full AI text detector.

Check a text yourself

Paste at least 100 words into the DeepSeek AI detector for a verdict, marked passages and a style similarity for DeepSeek, Qwen and GLM. One check is free.

To compare all seven model families, see the guide how each AI model writes.

Questions and answers

Can you tell DeepSeek writing by eye?

Sometimes, from clusters of signals such as frequent em dashes, even sentence lengths, and stock phrases like "rather than" and "a quiet". No single marker proves it, because humans use all of these too.

What are the most reliable DeepSeek tells in English?

Heavy em dash use, "rather than", "far more", "a quiet", "yet the" and "offers a", together with uniform sentence lengths. These appear many times more often than in human writing, but not in every text.

What are the tells in Chinese?

Phrases such as 往往, 并非, 而是, 在于, 节奏 and 后续, plus "not X but Y" contrast constructions and em dashes at several times the human rate.

Are these tells the same for Qwen and GLM?

No. Qwen leans on phrases like "serves as" and "ensures that"; GLM leans on "entirely", "highly" and "ultimately". Their phrase sets overlap with DeepSeek's only partly.

Is a phrase list enough to prove authorship?

No. Style is a signal. Use a detector that reads the whole text, and weigh the result with drafts and context.