Claude IDs merch, ChatGPT hallucinates history

Multimodal models still struggle with high-confidence hallucinations when faced with niche visual data they weren't explicitly trained to identify.

• –Claude's "home field advantage" resulted in instant recognition of its own visual branding assets, suggesting strong internal alignment.
• –ChatGPT's hallucination was surprisingly deep, mapping abstract icons to specific historical milestones like fire and steam engines.
• –The confusion between "Artifacts" (Claude) and "Canvas" (ChatGPT) indicates cross-model feature bleed in the training or fine-tuning datasets.
• –This failure mode highlights a lack of "I don't know" thresholding in vision models when interpreting abstract designs.
• –For developers, this demonstrates the necessity of grounding vision outputs in a known domain for critical classification tasks.