The Ledger · Entry 018 · Creation
Poetry and literary prose
ContestedSince 2024Criterion performanceLast reviewed 2026-08-17
Event
A study in Scientific Reports finds that non-expert readers cannot tell GPT-3.5 poems from those of ten well-known poets (46.6 percent accuracy) and rate the machine poems higher on several qualities.
The readers were not experts, the poems were short, and the study measured recognition and preference, not what a poem is for. The authors note that the AI poems were more accessible, and accessibility is what untrained readers reward. What is contested is whether a text without an author who lived it, meant it and answers for it is the same kind of object; publishers, prizes and readers currently say no, and the market for named human authors has not moved. Fluency is ceded; authorship is not.