Why Do LLM Pelicans Look the Same? Mode Collapse, Prompting, and the Design Creativity Debate

The short answer

LLMs asked to draw "a pelican riding a bicycle" produce nearly identical SVGs — same left-to-right composition, same sun, same road, same scarf — and the Hacker News thread on the article "The Emperor Pelican Has No Clothes" splits sharply over what that means. One camp calls it a real and under-discussed mode-collapse problem in post-trained models; the other camp argues that a bare prompt with no creative direction should produce a plain, literal answer, and that the sameness is instruction-following working as intended.

What the source material actually contains

The Hacker News submission is titled "Why do LLM pelicans look like this?" and links to an article at arcfu.com titled "The Emperor Pelican Has No Clothes." The scraped page content for that link is not the article text — it is an announcement thread for Design Creativity Bench, a benchmark from Kombai that evaluates AI models on UI design tasks. That announcement states:

16 frontier models were evaluated across 80 open-ended UI tasks, each adapted to two product domains with human designer references.
Three dimensions were measured: Originality (how distinct a model's designs are from other models' outputs for the same brief), Creative Range (how much a model adapts when the product domain changes), and Appropriateness (how well the UI meets the brief's functional requirements).
The best model reported, OpenAI GPT-6 Astra, scored 65.5 against a human reference of 92.3.

The benchmark's framing — "Frontier AI models still fall significantly short of human designers on design creativity" — is the claim the comment thread reacts to, even though the pelican examples themselves come from the article the post links to.

The core claim: LLM outputs are mode-collapsed

The strongest technical version of the argument comes from a commenter who separates output diversity from model capability:

"All current post-trained LLMs are mode-collapsed compared to the post-training set distribution, and this doesn't get nearly enough attention from the labs and independent researchers. It's far from being a transformer-only problem, it's a common failure mode in all kinds of models. It's not a fundamental limitation either, unsupervised models have no problem with being as diverse as the training set allows."

That comment (from orbital-decay) makes three distinct points worth separating: mode collapse is a post-training artifact, it is not specific to transformers, and it is not an inherent ceiling — base or unsupervised models retain more of the training distribution's spread.

A related commenter (ozgung) points at the specific convergences that are hard to explain from the prompt alone: "The sun, the road, the green, the scarf, going right. Even the style. They're not obvious associations one would make given the prompt."

The counterargument: a simple prompt deserves a simple answer

The most repeated objection in the thread is that the pelican test measures literal instruction-following, not creativity.

"If I ask for a pelican riding a bicycle, I am expecting a relatively boring image of a pelican riding a bicycle, not a pelican flying through a thunderstorm while on a bicycle. That's just straight up not what I asked for."

That commenter (randomblock1) extends the argument to a concrete failure mode: ask for an SVG of a laptop and an unprompted "creative" model may return a retro or futuristic laptop that is harder to recognize and more likely to be rejected — so training pushes models toward the neutral, canonical answer.

A second commenter (riebschlager) frames it as user intent: "I want to use these tools to express my creativity, not its own... Fulfilling a simple request simply seems like a feature, not a bug." A third (wolttam) states the position plainly: "It's unsurprising and perhaps even desirable that different LLMs return similar results to an identical prompt. If you want variation, prompt for it."

Prompting does shift the distribution

The thread contains first-hand evidence that the sameness is prompt-sensitive rather than fixed. One commenter (joshkel) tested variations in Opus 5.5:

Adding "Be creative" produced a beach scene with fish in the bike basket and a lighthouse in the background — described by the tester as "not too creative."
Adding "Pick a unique and creative variation of this concept" produced a "midnight circus act" of a pelican riding a penny-farthing on a tightrope while balancing a fishing pole with fish, strung between two lighthouses.

Another commenter (ed) argues this is the intended behavior: "A good prompt will shift the distribution away from average... This is a feature, not a bug."

A third (tetraodonpuffer) proposes a concrete follow-up experiment: run the pelican benchmark twice, once with "be original" appended, and check whether every model converges on the same idea of originality. The same commenter draws an analogy to image generation, where base models are more seed-sensitive than distilled or "turbo" variants, and to human behavior: "if you ask 100 people to draw a horse, it is quite likely it will be a lot more similar to a 'here's a stock horse photograph' than a 'here's what picasso would draw when asked to draw a horse.'"

"Diversity collapse" may be the wrong direction of change

A commenter (joshkel) flags a definitional problem: the article's phrase "the LLM diversity collapse problem" implies models are getting less diverse over time, but the linked benchmarks suggest newer, more capable models are more creative. That tension is unresolved in the thread.

A separate commenter (vatsachak) draws a much stronger conclusion from the same observation: "The reason why LLMs cannot be purely pre-trained on LLM outputs is because LLM output diversity is low," adding that this implies "there is going be no functional difference between hypothetical Qwen 4.5 and Fable 5.5" and that "LLMs are not AGI." Those are the commenter's claims, not established results.

Biases inside the "diverse" outputs

Two commenters point out that the sameness runs deeper than subject matter.

spuz: "even in the diverse set, most of the images show the pelican riding from left-to-right. There's a ton of bias in our data and even in our brains that you would need to erase to get true diversity."
ks2048: "in 10/13 of the examples, there are 3 small horizontal lines behind the bike meant to convey motion. These look surprisingly similar and not necessarily something that directly follows from the prompt."

A related observation from aabajian concerns default rendering style: SVG requests appear to imply a polygonal, flat look, and explicitly asking for photorealism changed the result — "It made a polygonal SVG. I asked it to make it more photorealistic and it looked way better, still an SVG though."

The commercial context

One commenter (phil-martin) identifies the Design Creativity Bench announcement as marketing: "this is a clever ad for their paid AI design product" at kombai.com/features/design. Readers evaluating the benchmark's headline numbers should weigh that the benchmark and the product share an owner.

Why the thread thinks this matters beyond pelicans

The practical concern is not aesthetics. One commenter (jvanderbot) frames convergence as a competitive-intelligence and security problem: if every model implements "make me a thing that does X" the same way, the output resembles "a kind of stack-overflow, well commented, internship project" — "Same-y enough that you could train an adversary on it? Or hack it afterward?"

A second (jeanpah) generalizes the risk: "If it happens with the pelican, it will happen to everything else." A third (seidleroni) connects the pattern to Ethan Mollick's writing on AI having a specific "taste," and argues the practical response is human direction: "this is one way humans can provide a lot of input and direction."

Open questions raised but not answered

Does "be original" itself collapse? The proposed test — append the instruction and check whether all models converge on the same notion of originality — was suggested but not run in the thread.
How do humans compare? One commenter (electroweak) asks whether a grade-4 class given the same prompt would produce more or less variety than the models.
Are there better probes than pelicans? One commenter (xtracto) reports that asking models for origami instructions — "Write an svg with graphical instructions on how to make a pelican riding a bicycle in Origami using any number of A4 paper sheets" — is far harder for current models, which mostly manage only "a basic airplane that really floats."
Is the benchmark measuring the right thing? The Design Creativity Bench's three axes (Originality, Creative Range, Appropriateness) explicitly separate distinctness from brief compliance, which is the distinction most of the comment thread argues the pelican test blurs.

Bottom line

The pelican sameness is real and reproducible, but the thread does not support a single explanation. The evidence points to at least three contributing factors: post-training mode collapse toward the modal answer, reward for literal instruction-following on underspecified prompts, and shared training-data biases such as left-to-right motion and a default flat SVG style. The disagreement is about which of those dominates — and whether the fix is better prompting, better post-training, or accepting that a plain request should get a plain answer.