Is human translation more conservative than machine translation?

What is it about?

Have you ever wondered how translators choose their words? Imagine you are translating a Chinese text into English. Would you write “seek” or “look for”? “Do not” or “don't”? These choices may seem small, but they reveal something interesting about how translators think and work. Our study explores whether translators, both humans and machine systems, tend to favour more formal, cautious language when translating. We call this tendency "conservatism". To investigate this, we compared four types of English texts: original English writing, professional human translations, translations produced by DeepL, and translations generated by Google Translate. The dataset covered more than four million words across four major text types: news reports, academic writing, fiction, and general informational texts. We focused on 21 common language choices. For example, do translators prefer single-word verbs or more phrasal verbs? Do they use full negative forms or contractions? The results revealed a clear pattern: human translators were the most conservative group. Across all text types, they consistently preferred more formal expressions. Even when translating fiction, a genre that often allows informal and conversational language, human translators tended to make the text sound slightly more formal than the original. AI translators showed a different pattern. DeepL and Google Translate displayed some conservative tendencies in news and fiction, but in academic and general texts they often produced language that was no more formal, and sometimes even less formal than that used by native English writers. We also found differences between machine systems. Overall, DeepL tended to be slightly more conservative than Google Translate, possibly because of differences in their training data and translation algorithms. Why does this happen? The answer may lie in human psychology. When people take on the role of a translator, they often feel responsible for producing a reliable and professional text. To avoid mistakes, misunderstandings, or criticism, they may unconsciously choose safer and more formal expressions. In other words, translators often engage in risk-avoidance behaviour. Machine systems, however, do not worry about being judged. They simply reproduce patterns learned from vast amounts of training data. As a result, their language choices reflect statistical patterns rather than concerns about professionalism or risk. These findings suggest that translation is not only a linguistic activity but also a psychological one. The words translators choose can reveal how they balance accuracy, readability, and the desire to avoid risk, something that still distinguishes human translators from today's machine systems.

Featured Image

Photo by Zach M on Unsplash

Photo by Zach M on Unsplash

Why is it important?

A First-of-Its-Kind Human–AI Comparison: Most previous research on translation conservatism has focused exclusively on human translators. This study is the first to systematically compare conservative language choices in human translation and AI-generated translation. Rather than simply evaluating translation quality, we investigate the underlying decision-making processes reflected in translators' linguistic choices. The study also compares two of the world's most widely used machine translation systems, DeepL and Google Translate, revealing systematic differences in how different AI models handle formality and stylistic variation. An Innovative Approach to Language Choice: Instead of examining isolated linguistic features, we focus on how translators choose between alternative expressions with similar meanings. This allows us to capture subtle preferences that reveal underlying translation behaviour. Methodologically, the study combines two advanced statistical techniques, profile-based correspondence analysis and mixed-effects logistic regression, to provide a robust picture of translation patterns. We also carefully control for the influence of genre (news, academic, fiction, and general texts) and linguistic feature type (lexical and grammatical choices). AI Translation Is Transforming the Industry: With the rapid development of AI tools such as ChatGPT, DeepL, and Google Translate, machine-generated translations are reaching increasingly high levels of quality. As a result, a growing number of people are asking: Do we still need human translators? Our findings suggest that the difference between humans and AI extends beyond accuracy. Human translators demonstrate a greater sensitivity to context, genre conventions, and appropriate levels of formality. These are areas where AI systems still face important limitations. A Shift in Translation Research: Translation studies has traditionally focused on the concept of “translation universals”, features believed to be common to all translated texts. Our findings challenge this assumption. We show that conservatism is not a universal characteristic of translation. Instead, it appears to be a behaviour particularly associated with human translators, reflecting a tendency to avoid linguistic risk and maintain professional credibility. AI systems, by contrast, do not make such judgments; they largely reproduce patterns learned from their training data. What appears to be a preference is often better understood as an algorithmic bias rather than a conscious choice. For the Translation Industry: The findings can help make post-editing more effective. Knowing that AI systems often struggle with appropriate levels of formality and stylistic adaptation allows human translators to focus their attention on these areas when revising machine-generated texts. More broadly, the study highlights that the value of human translators lies not only in producing accurate translations, but also in producing appropriate ones. Human expertise remains crucial for adapting language to specific audiences, purposes, and communicative contexts. For AI Development: The results provide useful insights for the next generation of translation technologies. Future AI systems need stronger awareness of genre, register, and stylistic conventions rather than relying solely on literal accuracy. Different communicative settings, such as academic writing, business communication, and literary translation, may require different levels of formality. Building this flexibility into AI systems will be an important challenge for future development. For Translator Education: The study also has implications for translator training. It encourages educators to pay greater attention to risk awareness, stylistic judgment, and context-sensitive decision-making. By understanding when a cautious, formal approach is appropriate, and when a translation should remain closer to the style of the original text, future translators can develop more sophisticated professional judgement and greater stylistic flexibility.

Perspectives