Séance de travail des membres du CRIHN 2018-2019: Richard Jean So

Value of the Information & Strength of the Argument

The presentation provides substantial value by connecting technical aspects of machine learning to long-standing questions in literary theory and poetry. So argues convincingly that the ability of a machine to fool human readers forces us to reconsider assumptions about the unique qualities of human creativity, such as ambiguity, emotional resonance, and aesthetic complexity. He grounds his argument in concrete examples, including the Duke student’s poem and scholarly applications of language models, and uses clear analogies to explain how classifiers work. The reasoning is logically structured, moving from specific observations to broader theoretical implications, and he openly acknowledges his own limitations, which strengthens the credibility of his claims. The dialogic format with the audience adds depth, though some arguments remain exploratory rather than fully conclusive.

Scientific Rigor, Source Quality, Title Accuracy

The talk maintains a high level of scientific rigor, with a clear distinction between established concepts (Turing test, language models) and personal interpretations. While specific sources are not always formally cited, So references key figures like William Empson and concepts like the Turing test, and he transparently discusses the origins of the example poem. The title is well-aligned with the content, as the talk genuinely addresses poetry in the age of AI, both in terms of generation and analysis. The presentation is not overly technical, but it is intellectually honest about the limitations of machine learning and the tensions it introduces in the humanities, without overstating claims.

The title accurately reflects the content, which examines the impact of AI on poetry and literary analysis, both in terms of creation and scholarly interpretation.

8/10

The talk offers a thoughtful, balanced exploration of AI and poetry, with clear explanations of machine learning concepts and appropriate caveats about the speaker's expertise. While not all sources are explicitly cited, the reasoning is sound and builds on established theories.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 02:00Presentation of a poem and the initial question: is it human or machine-written?
• 05:00Discussion of assumptions about human creativity versus machine limitations, referencing Empson and Romantic notions of poetry.
• 10:00Reveal that the poem was written by an algorithm and introduction to the Turing test.
• 20:00Explanation of machine learning through classifiers and language models, using spam filtering as an example.
• 30:00Application of language models to literary analysis and cultural analytics, searching for patterns across large corpora.
• 40:00Discussion of tensions between computational methods and traditional humanistic approaches, including the role of semantics and figurative language.

• Zachary Snow's machine-generated poem — The poem that fooled human editors, presented at the start and discussed throughout the talk.
• Turing Test — Mentioned as a theoretical framework for machine intelligence and passing as human.
• William Empson's 'Seven Types of Ambiguity' — Referenced to illustrate the idea that machines cannot achieve ambiguity in language.
• Raptor (1980s machine-generated poetry book) — Mentioned in the discussion as an early example of machine-generated poetry.