AI Poetry Translation Fails to Capture Nuance, Highlighting Human Limitations

Artificial intelligence has revolutionized translation, but a recent test of ChatGPT’s ability to render a 1953 Urdu poem into English revealed persistent gaps in machine understanding of literary artistry. The experiment, led by poet Adeeba Shahid Talukder and journalist Lisa Lock, demonstrated that AI struggles to convey the emotional depth, cultural context, and poetic structure of works like “Mulāqāt” by Faiz Ahmed Faiz, a cornerstone of South Asian literature.

Why Do AI Models Struggle With Poetic Translation?

Poetry requires more than linguistic accuracy—it demands cultural literacy, metaphorical interpretation, and emotional resonance. According to literary critic Gayatri Chakravorty Spivak, translating poetry involves a “worlding” of language, where the translator must bridge not just words but entire worldviews. AI systems, including ChatGPT, often produce translations that are technically correct but emotionally flat.

In a test of Faiz’s “Mulāqāt,” ChatGPT translated the opening lines as “This night is the tree of pain,” omitting the rich metaphors and emotional weight of the original. Human translator Talukder rendered the same passage as “This night is the dark, lush tree of a grief greater than you and I,” capturing the poem’s anguish and poetic density.

What Are the Key Limitations of AI in Literary Translation?

AI translation tools face three major challenges when handling poetry:

1. Metaphor Interpretation: Poets like Faiz use extended metaphors that lack direct equivalents in other languages. ChatGPT’s translation of “caravans of torch-bearing stars” as “thousands of torches in place of stars” stripped the imagery of its cultural and emotional layers.
2. Sentence Structure: Complex, nonlinear sentence structures in poetry confuse AI models. For example, ChatGPT misinterpreted the Urdu phrase “apnā” (a reflexive pronoun) as “my,” altering the meaning of lines about “a thousand moons” weeping.
3. Emotional Nuance: AI lacks the embodied experience to grasp the emotional subtleties of poetry. While Talukder’s translation describes “drops of silence” falling on a lover’s brow, ChatGPT rendered this as “tiny diamonds,” a literal but emotionally hollow substitution.

How Does Human Translation Differ?

Human translators bring cultural expertise and creative intuition to their work. Talukder, who has translated Faiz’s poetry for international audiences, emphasized that “poetry is not just language—it’s a dialogue between the poet’s world and the reader’s.” This involves understanding historical context, such as Faiz’s imprisonment and his themes of political struggle and personal loss.

Urdu literary critic Shamsur Rahman Faruqi noted that Urdu poets “treat metaphor as fact,” a convention AI fails to recognize. For instance, the metaphor of a “tree of grief” in “Mulāqāt” becomes a foundation for subsequent imagery, like “dew of silence” transforming into “silver pearls.” AI models, lacking this cultural framework, produce fragmented, disconnected lines.

What Are the Implications for AI in Creative Fields?

The experiment underscores a broader debate about AI’s role in creative industries. While tools like ChatGPT excel at translating legal documents or technical manuals, they falter when faced with art that relies on ambiguity and cultural specificity. As poet A.K. Ramanujan wrote, “Only poems can translate poems,” a sentiment echoed by Talukder, who argues that “AI can’t replicate the human soul’s engagement with language.”

For now, literary translation remains a domain where human expertise is irreplaceable. As AI continues to evolve, its ability to handle poetry may improve—but for the foreseeable future, the art of translation will rely on the nuanced, culturally aware work of human poets and scholars.