Imagine an AI rating a jumble of words like "Goetterdaemmerung's corpus hemorrhaged through cryptographic hash" as high literature. That’s exactly what a German researcher discovered about OpenAI’s GPT models this year. Christoph Heilig, from Munich’s Ludwig Maximilian University, found that even the latest GPT-5.4—released in March 2026—consistently praises pseudo-literary gibberish, raising red flags about AI’s role in aesthetic judgment.
From Rainy Streets to Existential Void 🌧️🌀
Heilig tested GPT models by asking them to rate sentences for literary quality. Starting with a simple scene ("The man walked down the street…"), he added increasingly absurd phrases. The result? The more "film noir meets tech jargon" the text became, the higher GPT scored it—even with its "reasoning" features turned on. "Eschaton pooling in existential void," anyone?
Why This Matters for AI’s Future 🔍
Heilig warns that as companies use AI to evaluate other AI systems, these biases could snowball. Henry Shevlin of Cambridge’s Leverhulme Centre compares it to human cognitive flaws but adds: "Processes with little human oversight are ripe for exploitation." Think academic journals using GPT to review papers—yikes!
While OpenAI reportedly tweaked GPT to flag Heilig’s test phrases as "literary experiments," the core issue remains: Can we trust AI’s taste? For now, it’s clear—GPT might need a poetry class. 📉
