Why do LLM pelicans look like this?
A viral prompt asking AI models to “generate an SVG of a pelican riding a bicycle” has become an informal benchmark for how similarly large language models respond to underspecified creative tasks. Commenters note that most systems now produce nearly identical, generic scenes — same side-on view, same sunny sky, same accessories — and see this as evidence of “mode collapse” driven by shared training data and post-training alignment techniques. Some welcome this conformity for practical uses like coding and business graphics, while others worry it flattens innovation, reinforces clichés, and exposes why human taste and direction remain critical in AI-assisted creation.
Prompt, Pelican Benchmark, and “Diversity Collapse”
- Many note that different models give strikingly similar SVGs for “a pelican riding a bicycle”: side view, left-to-right motion, sunny sky, clouds, generic road/grass, scarf/helmet, motion lines.
- Some see this as evidence of an “LLM diversity collapse” where models converge on the same centroid output for underspecified prompts.
- Others argue this is expected and even desirable: given a vague request, the model should pick the most statistically typical, recognizable scene.
Creativity vs Predictability
- One camp wants models to be more imaginative by default; otherwise, AI-generated design, code, and writing all converge to bland sameness and may stifle innovation.
- Another camp prefers predictability: a simple prompt should yield a minimally surprising, generic result (like a standard PB&J or neutral laptop icon), not a wild reinterpretation.
- Several commenters stress that interesting results require richer prompts and iterative direction; “generic in ⇒ generic out.”
Role of Training, RLHF, and Synthetic Data
- Debate over whether LLMs can be trained purely on LLM-generated data; linked work suggests heavy reliance on human “seed” data plus synthetic augmentation, with ongoing human curation.
- Some claim RLHF/alignment and post-training compress diversity, pushing models toward a single style; base models are said to be more varied but less “safe” and more hallucination-prone.
- There is concern that increasing synthetic data and alignment loops will further narrow output space across models.
Comparisons to Humans and Style Convergence
- Comparisons to children drawing houses or generic scenes: humans also regress to clichés, but human groups still show more variation than current models on the pelican prompt.
- Noted biases: left-to-right motion, icon-like SVG style, common tropes (fish in basket, seaside, scarves) mirror patterns in training data and broader culture.
- Similar “sameness” is observed in corporate logos, UIs, and interior design; some see LLMs as amplifying existing real-world convergence.
Use Cases and Tools
- For code, mode collapse is seen as a feature: consistent, “boring” implementations and standard patterns are desirable.
- For art and design, sameness is a bug; suggested mitigations include: better prompts, external randomness/brainstorm tools, multi-model setups, and human-in-the-loop critics to steer style.