Does DeepSeek translate government white papers like a human?
A leakage-controlled analysis found detectable stylistic signatures in 72 matched text windows. DeepSeek and Google machine translations clustered closer to each other than to published human versions, but the study did not test accuracy, fidelity or reader understanding.
The short version
- What happened
- The corpus comprised 72 matched 5,000-word windows: 24 each from published human, Google neural-machine and DeepSeek V4Pro translations of the same white-paper series.
- Who it affects
- People reading translated public policy need accurate meaning and clear attribution; neither outcome can be inferred from stylistic classification alone.
- What remains unproven
- Only 72 windows from one white-paper producer and language pair
Keep in mind
Only 72 windows from one white-paper producer and language pair
Untouched test set contained 15 windows
Published, post-edited human text was compared with raw machine output
Residual overlap and historical differences could influence results
Model performance varied across classifiers and resampling schemes
No translation accuracy, fidelity, acceptability or reader outcome
Source analysis · 1 linked source. Primary analysed source date: 26 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
The study tested stylistic fingerprints, not truth
DeepSeek V4Pro and Google neural machine translation produced official-document English that was distinguishable from published human translation in a peer-reviewed computational study. Researchers aligned Chinese State Council Information Office white papers into 24 minimally overlapping 5,000-word windows for each of three translation conditions, creating 72 matched observations. They measured 142 linguistic features and used training-only selection to prevent the held-out test text from choosing the features.
The strongest classifiers beat the 33.3% chance baseline on an untouched 15-window test set: Naive Bayes reached 80.0% accuracy and Random Forest 86.7%. The result is not a production-ready detector. The test set contained only five matched chunks per condition, and other models ranged from 40% to 60%. Nested cross-validation and 50 repeated group-aware splits kept Naive Bayes and Random Forest above chance, but at broader ranges of 0.667–0.811 and 0.606–0.753.[1]
Both machine systems shared less cohesive marking
A separate exploratory factor analysis found the clearest contrast in discourse cohesion and causal or deictic linking. Published human translations scored higher than both machine systems, while DeepSeek and Google were statistically similar on most dimensions and sat closer to one another in the five-dimensional style space. The paper calls this a shared machine-translationese: a pattern in linguistic form rather than proof that one system preserved meaning better.
For government communicators, that boundary is critical. Explicit links between claims can affect how readers follow institutional reasoning and attribute commitments. But the paper did not measure comprehension, cognitive load, diplomacy or trust. Its proposed link between reduced cohesive marking and reader effort remains a hypothesis until people actually read controlled translations and their understanding is tested.[1]
The distinction that matters
What the classifiers could and could not tell apart
The result concerns writing style under one controlled corpus design.
Corpus
72 matched 5,000-word windows
Human, Google NMT and DeepSeek V4Pro
Signal
21 training-selected linguistic features
Best models classified above chance
Not measured
No accuracy or reader test
Quality and public impact remain unknown
Editing history is a major alternative explanation
The human translations were published, post-edited texts; the machine translations were raw outputs. Editorial revision may therefore explain part of the gap attributed to human translation. The human material was also not fully time-matched to the 2026 systems, and the corpus covered one producer, one genre and one language pair. DeepSeek V4Pro is one model version in a rapidly changing family.
The study explicitly does not evaluate translation accuracy, fidelity or acceptability. Public bodies should not infer that stylometric separability makes machine translation unusable, or that a high classifier score certifies a human document. A fair operational test would compare raw human drafts, machine drafts and professionally edited versions of the same current documents, with blinded expert accuracy review and reader-comprehension outcomes.[1]