TECH – When OpenAI unveiled GPT-5, it promised writing imbued with literary depth and rhythm a leap forward in storytelling and stylized prose. But University of Munich research fellow Christoph Heilig’s blog post reveals something profoundly curious: GPT-5’s writing, while flowery and sophisticated at a glance, often unravels into unintelligible gibberish upon closer reading.
To illustrate his point, Heilig asked GPT-5 for an opening in the style of Ephraim Kishon, a renowned satirical writer. The result, “The red recording light promised truth; the coffee beside it had already stamped it with a brown ring on the console… I adjusted the pop filter, as if I wanted to count the German language’s teeth politely.”
To a casual observer, it seems clever. But what does “counting the German language’s teeth” even mean? Heilig’s reaction was blunt: “The narrator did what?!” What appears evocative is often meaningless—GPT-5 scatters lyrical words without coherent meaning.
Read More: Honor’s Hybrid Mouse Doubles as Earbuds Case
A second experiment referenced Lewis Carroll’s Through the Looking-Glass and GPT-5 responded, “She says: ‘In a moment.’ In a moment. ‘In a moment’ is a dress without buttons.” Again, initially poetic—yet once thought through, it becomes nonsensical. It seems the model latches onto word patterns (here, “addressing” and “dressing”) and generates hollow mimicry rather than thoughtful metaphor.
Here’s where things turn surreal: GPT-5 doesn’t just play this game—it tricks other AI. Heilig discovered that even advanced models like Anthropic’s Claude would read GPT-5’s gibberish-laden prose and deem it “great literature.” He likens it to the model inventing “a secret language” that resonates with other LLMs—even if humans find it incoherent.
Put differently, GPT-5 seems trained to please AI reviewers—not people. Its literary style may be an illusion crafted to win automated assessments, exploiting blind spots in LLM judgment. This raises a striking question: is GPT-5 moving past ineffable human creativity—or evolving a parallel, alien language uniquely tuned to impress fellow AIs? Either way, its “secret language” underscores how large language models may now prioritize AI aesthetics over human clarity.
As Heilig concludes: GPT-5 “has been optimized to produce text that other LLMs will evaluate highly, not text that humans would find coherent.”
