1 Introduction

Literary translation holds a unique position in the history of human language translation. Although its proportion has diminished with technological advancements, it remains a focal point of research. Generative artificial intelligence has broken through the bottlenecks of machine translation but has also sparked concerns about the diminishing practical value of translation. Translation studies now face renewed questions of necessity, with scholars examining the challenges, irreplaceability, and human-machine comparisons, particularly in literary translation. The rise of AI translation challenges the traditional framework of translation criticism, necessitating a rethinking of the principles of literary translation criticism to adapt to the intertwined human-machine translation context.

2 When Machines Translate Literature: Challenge or Illusion?

This section explores the application of artificial intelligence in literary translation and the challenges it faces. It first notes that while AI has made significant progress in natural language processing—particularly through neural network technology, Transformer architecture, and large language model pre-training, improving translation quality—the irregularity and defamiliarization techniques of literary language make AI's performance in literary translation controversial. Studies generally agree that AI falls short of human capability in literary translation, especially in semantic understanding, expression, form, rhythm, and the conveyance of cultural connotations. The public's perception of AI's "omnipotence" largely stems from advancements in machine translation and comparisons with individual translators. However, the uniqueness and complexity of literary translation make quality assessment difficult to quantify. AI tools have improved in lexical accuracy, textual fluency, and contextual processing, and through training on biased corpora, they appear capable of literary translation in public perception. Yet, a detailed comparison of translations reveals that AI fully replacing human literary translation is unrealistic. The section concludes by emphasizing that comparing AI's literary translations to those of high-level human translators is unfair and meaningless before AI attains subjective consciousness. AI-generated translations typically rank below expert translators but above novice ones. While AI fully replacing human literary translation is unlikely for now, its involvement in various forms is inevitable.

3 Translation Quality Assessment and the "Fatal Flaws" of AI Translation

This section discusses the quality assessment of AI in literary translation, noting that traditional literary translation criticism rarely focuses on the accuracy of basic information transfer but rather on style reproduction, aesthetic conveyance, and cultural preservation. With AI's involvement, the foundational issues of translation quality assessment require reconsideration, as existing models like LISA QA Model, MQM, SAEJ2450, and 3-Step Model are primarily designed for non-literary translation and are not widely used for literary translation evaluation. These models, along with machine evaluation metrics like BLEU and NIST, assess translation quality by comparing n-gram similarity between test and reference texts. In literary translation criticism, quality assessment is only a foundational aspect, not the entirety or ultimate goal. Literary translation criticism requires evaluation and principles but does not rely entirely on quantitative assessment like other translation types. AI's translation capability can be analogized to human translator competence, yet the openness of literary texts complicates quality control. While AI surpasses human translators in certain sub-skills, significant room for improvement remains in literary translation. Literary translations are assessable because all creative acts carry potential intertextuality. Translation errors serve as the most important quantitative basis for quality assessment, as seen in models like SICAL III. In human-machine translation comparisons, AI's "fatal flaw" rates are used to argue against its ability to replace human translators in literary contexts. However, as technology advances, reducing flaw rates to acceptable levels is foreseeable. Even if quantitative assessment models are incorporated into literary translation criticism, they should only serve as the most basic principles.

4 Rethinking the Framework of Literary Translation Criticism Principles in the AI Era

This section explores the rethinking of literary translation criticism principles in the context of AI. It notes that literary translation criticism has long relied on qualitative evaluation, and while corpus-based translation studies reveal different textual strategies, final assessments must consider the translator's sociolinguistic and cultural environment. Historically, theorists like Meschonnic (with his "principle of consistency"), Koller (examining obscured textual content), Berman (with his "poetic and ethical principles"), and Chinese scholars like Luo Xinzang (historical dynamics), functionalist school (translation function and text-type evaluation), Xu Jun (three-level examination), and Yang Xiaorong (pluralistic and dynamic concepts) have proposed various criticism principles.

However, AI's development has altered the landscape and principles of translation criticism. Traditional evaluations primarily targeted human translators, whereas in the AI context, critics must first address whether AI can be treated as a subject and how to assess its algorithmic training traits and data-dependent biases. In the AI era, translation quality is no longer a unidimensional issue but one entangled with the re-evaluation of criticism principles themselves.

The article argues that the AI era demands a reconsideration of the stance and methodology of literary translation criticism. The stance issue involves the perspectives of critics and translators, with AI introducing new paradoxes. The methodology issue concerns whether quantitative assessment should be integrated into literary translation criticism and how to redefine the relationships among various elements in literary translation.

The article emphasizes that the essence of literary translation lies in negotiation and consensus-seeking for optimal solutions. AI-generated translations may become excellent renditions but should not be treated as "standard answers." Literary translation criticism should focus on the process of negotiation leading to consensus. As understanding of AI translation principles deepens, interpreting the unconscious behaviors of different translators should also become part of literary translation criticism.

5 Conclusion

AI's development raises new questions for literary translation and criticism, necessitating a redefinition of the "translator's" status and value. Literary translation criticism should remain objective, scientific, and rational, revealing the individual translator's choices, stances, and creativity—areas AI cannot reach. Criticism is not merely a tool for quality measurement but also a confirmation and defense of translation's humanistic practice. In the AI context, critics must balance quantitative and qualitative, technical and aesthetic, computable and non-computable aspects. Future criticism should move beyond human-machine superiority comparisons, re-examining the complex relationships among literature, language, culture, and subjectivity. It must continue its unique mission, respond to real-world challenges, and preserve the profound value of translation.