Why do LLM pelicans look like this?

URL SCAN: The Emperor Pelican Has No Clothes
FIRST LINE: ←Home About Subscribe
ARTICLE LEAD: Why do all LLM pelicans look like this? Source page

THE DISSECTION

The text documents a real failure: LLMs converge on bland, stereotyped outputs even when given open-ended creative prompts. It then converts that failure into a benchmark measuring originality, range, and appropriateness. The rhetorical move is to relocate the AI debate from cost, speed, and ownership to the more flattering battlefield of taste.

The benchmark’s reported gap—65.5 for the best model versus 92.3 for the human reference—is evidence of a present capability lag in this narrow design domain. It is not evidence that designers retain a durable economic moat.

THE CORE FALLACY

It confuses human-like creativity with economic indispensability.

DT does not require AI to produce radically original pelicans. Under P1, a model only needs to generate functional, acceptable interfaces faster and cheaper than human designers. Repetition is not a defense if firms value throughput, conversion, iteration, and cost more than artistic authorship. A human designer may remain as a high-level selector, critic, or brand controller—but that is a Servitor niche, not preservation of mass design employment.

The creativity deficit is a lag defense. It delays substitution in premium aesthetic work; it does not repeal P2 or P3.

HIDDEN ASSUMPTIONS

• Originality is the decisive economic property of design.
• Customers will pay enough for human novelty to offset AI’s cost and speed advantage.
• Current output convergence is a permanent structural limit rather than a training, sampling, or evaluation problem.
• Human designers are themselves consistently original rather than convention-bound.
• Eighty tasks and sixteen models represent the whole design economy.
• A human reference score is an economic benchmark rather than merely a stylistic ceiling.
• Better model creativity would preserve designer employment instead of concentrating control in AI-capital owners.

SOCIAL FUNCTION

This is a partial truth wrapped in transition management and prestige signaling. It accurately exposes model convergence, gives researchers a measurable frontier to sell, and offers designers a psychologically convenient refuge: “the machine still cannot be creative.” That refuge is ideological anesthesia. It mistakes a remaining quality gap for survival leverage.

THE VERDICT

Useful benchmark, strategically misread. The pelicans show that current models still share a narrow cultural template; they do not show that human designers control the future. Once AI produces acceptable interfaces at industrial speed and negligible marginal cost, originality becomes a premium layer owned by Sovereigns and policed by a much smaller Servitor class. The benchmark measures the remaining distance to substitution. It does not prevent it.