← News

ResearchSources: arXiv · 16 Nov 2026

Researchers test whether AI can write an infinite poem

An arXiv paper reports two experiments using couplets to probe how far large language models can extend a poem, finding interesting output that still trails human creativity.

In plain words

A research paper asks whether AI can keep writing a poem indefinitely. The authors used a strict repeating verse form where only the end words change, and had large language models generate new couplets. The AI produced interesting word pairings, but the authors judged it below human-level creativity in handling poetic relationships.

What was announced

The paper, posted on arXiv, describes two experiments built on couplets whose end words carry the poetic relationship within a strict anaphoric, litany-like structure. In the first experiment, the authors used manual prompting and co-creation to extend an original poem, which they then developed into an animated webpage installation and interactive artwork. The second experiment reached an automated output of 1,000 couplets by combining a large language model with semantic similarity, and the authors report that their second method showed a higher capacity for both semantic and quantitative output.

The authors conclude that the experiments expose limits in current LLMs' word-level generation capacity, while suggesting that poetry may itself revive by incorporating AI workflows. The paper also includes commentary and a brief survey of computational poetry experiments, discussing authorial intention and the distinction between natural and artificial poetry, and argues why the couplet form suits testing AI's poetic output.

What it does not show

  • The evidence is the paper's abstract only; the full article, including methods, models used, and evaluation details, was not available for review.
  • The claim that AI falls short of human-level creativity reflects the authors' own qualitative judgment, not an independent evaluation.
  • No specific LLM names, benchmarks beyond the 1,000-couplet output figure, or comparison results are given in the supplied passages.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.