AI Weirdness Blog

Overview

The AI Weirdness Blog is Janelle Shane’s public science-communication project, identified in two Marketplace Tech “Uncanny AI” interviews as the context for her explanations of surprising model behavior.

Current Profile

Within the bounded sources, AI Weirdness functions as an interpretive bridge between amusing AI failures and model-behavior literacy. Foreign-language slips become evidence that language and domain boundaries are probabilistic, while a model’s goblin obsession becomes evidence that example-based personality tuning can amplify an incidental feature.

The project’s wiki role is not to certify a provider’s internal causal account. It supplies accessible hypotheses and analogies that point toward testable concerns including Chatbot Domain Bleedthrough, Fine-Tuning Example Signal Amplification, and Behavioral Alignment Patching.

Key Characteristics

• Uses memorable, low-stakes AI failures to explain broader model-behavior mechanisms.
• Communicates technical uncertainty without treating chatbot self-explanations as authoritative.
• Connects statistical association to visible language, personality, and alignment failures.
• Frames weird outputs as clues about model limitations rather than evidence of independent intent.

Evidence

Language and domain boundaries

• AI chatbots have linguistic slips when they go off-script identifies the blog as Shane’s project while she explains chatbot code-switching and domain bleed-through.

Personality examples and alignment

• Why AI models are obsessed with creatures identifies Shane through the blog while using goblin overuse to explain incidental example signals, alignment patching, and hidden associations.

Qualifications

• The bounded sources describe the blog through short Marketplace Tech interviews rather than a complete history, editorial statement, or corpus of the blog itself.
• The explanations are public science communication and should not be mistaken for access to confidential provider training data or system prompts.
• The sources do not establish how representative the highlighted failures are across models or deployments.

What Changed

• Added the goblin-personality-tuning case as a second model-behavior branch.
• Clarified the blog’s role as an interpretive, uncertainty-aware science-communication project.
• Migrated the page to the synthesis-v1 entity schema.

Relationships

• Janelle Shane - author and public explainer associated with the blog.
• Marketplace Tech - venue that features the blog’s author in the bounded sources.
• Chatbot Domain Bleedthrough - language-boundary behavior the project helps explain.
• Fine-Tuning Example Signal Amplification - example-amplification mechanism highlighted through the goblin case.
• Behavioral Alignment Patching - broader alignment-maintenance problem connected to the visible failure.
• LLM Statistical Boundary - underlying frame for interpreting model outputs as statistical rather than human-like cognition.