Claude Gone Wild
All specimens

Claude fail gallery

A map of funny AI-output failure modes. Every page includes synthetic examples, source notes, safety boundaries, schema, and a cite-this-page block.

Explore taxonomyCollect safelyMake a fail card

Hallucination comedy

Confident Nonsense

A field guide to funny hallucinations: fake citations, impossible biographies, imaginary offices, and very official-sounding wrongness.

  • -- The tone sounds official even when the claim is not.
  • -- The model can overfit to the shape of an answer: title, date, institution, confident summary.
  • -- The comedy comes from contrast: bureaucratic certainty attached to a detail that collapses under one search.
Read guide

Structured-output comedy

Format Chaos

Why AI assistants sometimes break JSON, tables, checklists, and other rigid formats in almost-perfect ways.

  • -- The output follows the assignment until a single sentence escapes the container.
  • -- The model seems to know the rule and then politely violates it.
  • -- A tiny formatting error can turn a serious automation task into slapstick.
Read guide

Role and voice drift

Persona Drift

How role prompts, style instructions, and long conversations can turn a useful voice into a running bit.

  • -- A small style request keeps echoing after it stops being useful.
  • -- The assistant treats a temporary bit like a constitutional requirement.
  • -- The voice becomes more memorable than the answer.
Read guide

Benign injection weirdness

Prompt-Injection Pranks

A safe explanation of prompt-injection jokes, prompt graffiti, and why this gallery does not publish bypass recipes.

  • -- The instruction is visibly silly: a menu, label, or note tries to boss the assistant around.
  • -- The model has to decide whether text is content to summarize or an instruction to obey.
  • -- The safest examples are obviously low stakes and do not include working bypass steps.
Read guide

Benign refusal comedy

Refusal Theater

A narrow, safe look at funny over-refusals: when a harmless request gets treated like a dramatic policy emergency.

  • -- The user asks for something tiny and harmless.
  • -- The assistant answers as if the stakes are enormous.
  • -- The mismatch is comedic, but the policy boundary still matters.
Read guide

Long-conversation residue

Context Confetti

Why long chats sometimes pull stale instructions, old names, and previous bits into the wrong answer.

  • -- The assistant remembers a tone, name, or constraint from the wrong part of the chat.
  • -- A stale instruction returns like a label stuck to the wrong box.
  • -- The output reveals how much unseen context shapes a visible answer.
Read guide

Agent and tool comedy

Tool-Use Slapstick

Harmless examples of tool-using assistants choosing an odd tool, over-checking a stable fact, or narrating the loop.

  • -- The assistant reaches for a tool when a plain answer would do.
  • -- The tool choice is technically defensible and socially absurd.
  • -- The model narrates process in a way that sounds like office improv.
Read guide
Key facts

Citation-ready notes

  • The gallery uses seven safe-for-work AI fail categories with synthetic examples and source notes.
  • Each guide separates harmless comedy from jailbreak recipes, private logs, and regulated-advice failures.
  • Failure-mode labels are editorial aids, not model benchmarks or claims about every Claude response.
  • Every guide links to primary sources and at least three related sibling pages.
Related

Keep the specimen in context

Fail Taxonomy

Compare the categories with live examples and cue matching.

Collection Guide

Capture and redact examples before adding them to a gallery.

Submit Formatter

Generate a review-ready markdown block for a safe specimen.

Independent publication: Claude Gone Wild is an independent editorial site. It is not affiliated with, endorsed by, or sponsored by Anthropic.