Janelle Shane

Overview

Janelle Shane is an AI author and science communicator associated with the AI Weirdness blog. In three Marketplace Tech “Uncanny AI” episodes, she interprets odd chatbot behavior as evidence of statistical association, imperfect salience, and fragile behavioral control rather than human-like intent.

Current Profile

Shane’s role across the bounded sources is to make model-behavior failures legible through concrete analogies. She separates memory from judgment when a chatbot resurfaces an irrelevant personal fact, separates fluent language from sealed domain boundaries when a model code-switches, and separates intended personality from learned example signals when a model becomes preoccupied with goblins.

Her explanations consistently support a cautious operational view: a plausible model-generated explanation is not proof of cause, accurate recall is not appropriate recall, and a prompt-level fix does not demonstrate that the underlying association has been removed. The newest source extends this stance into Behavioral Alignment Patching, Fine-Tuning Example Signal Amplification, and Automated Hiring Proxy Discrimination.

Key Characteristics

• Explains technical model behavior through accessible analogies without attributing human motives to the system.
• Treats fluent or correct output as compatible with poor domain boundaries, salience, or behavioral control.
• Distinguishes plausible explanations from verified causal accounts of model internals.
• Connects low-stakes AI oddities to higher-stakes privacy, safety, bias, and alignment risks.
• Emphasizes that training examples and human behavior can carry unintended associations into model outputs.

Evidence

Statistical and domain-boundary explanation

• AI chatbots have linguistic slips when they go off-script has Shane explain foreign-language slips through multilingual training data, probabilistic token prediction, and domains that are not cleanly walled off.

Memory and social judgment

• Uncanny AI: Why AI bots remember random, sometimes useless information has Shane distinguish storing a personal detail from judging when it is relevant, proportionate, or sensitive to mention.

Example signals and behavioral patching

• Why AI models are obsessed with creatures has Shane explain goblin overuse through personality examples and small reused fine-tuning data, then connect patching to hidden side effects and human-bias replication.

Qualifications

• These are public-facing explanations in short interviews, not provider-authored technical postmortems or direct inspections of model weights, prompts, and training datasets.
• Shane explicitly treats some causal explanations as plausible rather than confirmed when the relevant internal evidence is unavailable.
• The newest source’s hiring examples describe a general discrimination mechanism without naming or auditing a particular deployed system.

What Changed

• Added Shane’s explanation of incidental example signals becoming over-weighted during personality tuning.
• Extended her profile from memory and language-boundary failures to behavioral alignment and proxy discrimination.
• Migrated the page to the synthesis-v1 entity schema.

Relationships

• AI Weirdness Blog - science-communication project through which Shane is identified.
• Marketplace Tech - interview venue for the bounded “Uncanny AI” episodes.
• Megan McCarty Carino - host who frames Shane’s model-behavior explanations.
• Chatbot Memory Salience Failure - failure mode Shane explains through remembered details used without proportion.
• Chatbot Domain Bleedthrough - language and domain-boundary failure Shane explains through mixed training data.
• Behavioral Alignment Patching - alignment-maintenance pattern Shane describes as repeated symptom-level repair.
• Fine-Tuning Example Signal Amplification - post-training mechanism used to explain the goblin case.