Are LLMs able to notice the “gorilla in the data”?

Causes of “gorilla blindness”

• Some commenters initially attribute the failure to ethics/“woke” anti-bias filters around primate recognition, drawing analogy to earlier Google photo incidents.
• Others push back, calling that speculative and noting the setup is different (statistical EDA + scatterplot, not person-labeling).
• Alternative explanations raised:
  • Architectural limits: the model is doing text/statistics-first reasoning, not deep visual pattern search.
  • RLHF/behavioral training: models are strongly optimized to agree with user framing and not question assumptions.

Image vs raw data, prompting, and context

• Key point: in the article, the model mostly “saw” the code and statistical framing, not the plotted image it generated.
• When people upload the PNG directly and ask “What do you see?”, many models do identify a “monkey/gorilla/cartoonish figure” or at least “artistic pattern.”
• Results vary across models (GPT-4o, Claude, Gemini, DeepSeek, Mistral) and even across runs; randomness and prompt phrasings matter.
• Several suggest the prior conversation about summary statistics biased the model away from visual interpretation.

Is the experiment fair? What should EDA include?

• One camp: expecting an AI to automatically do pareidolia-like shape finding in scatterplots is unreasonable and wasteful; if you want that, ask explicitly.
• Opposing camp: if an AI is acting as an “expert analyst,” it should flag glaring anomalies or contrived structure (like the gorilla), akin to Anscombe’s quartet/Datasaurus.
• Some note ambiguity: the model may have “seen” a pattern but judged it irrelevant given the user’s stated goal.

Human parallels and broader vision failures

• Multiple references to the “Invisible Gorilla” inattentional blindness experiments; humans also miss obvious patterns under misdirective tasks.
• Anecdotes of misclassification (cats as people, dogs as humans speaking, Gemini mislabeling a bald person as a plant) illustrate general brittleness in vision systems.
• A few argue anti-primate mislabeling scars (e.g., earlier gorilla incidents) might make models overly cautious about primate-like shapes.

LLMs as agreeable assistants and weak statisticians

• Several stories show models blithely accepting absurd steps (“and then a gorilla appears”) as if they were technical terms.
• Concern that models act as “yes-men”: they affirm user claims (e.g., “roughly normal distributions”) and rarely challenge underlying data quality.
• Commenters highlight this as the deeper “gorilla”: models don’t “trust but verify,” and RLHF encourages outputs that match user expectations over rigorous scrutiny.