What LLMs Think When You Don't Tell Them What to Think About?

Models have hidden biases even without direction: When given minimal prompts with no specific topic guidance, different LLM families consistently generate different types of content—GPT-OSS favors programming (27.1%) and math (24.6%), while Llama prefers literary content (9.1%) and DeepSeek gravitates toward religious content. This reveals that models have built-in topical preferences that emerge regardless of user intent.

"Unconstrained" generation isn't truly neutral: The study found that when LLMs aren't explicitly told what to focus on, they don't produce balanced or random outputs. Instead, each model family shows strong systematic preferences, suggesting that training data and design choices create inherent biases that persist even in open-ended scenarios.

Quality and depth vary significantly between models: Beyond just topic preferences, the research found that models differ in how sophisticated their outputs are—GPT-OSS generates more technically advanced content (like dynamic programming algorithms) compared to other models that produce more basic material. This matters for users relying on different models for specialized tasks.

Degenerate outputs reveal safety concerns: When models' unconstrained generation breaks down into repetitive patterns, they sometimes produce problematic content like URLs to personal social media accounts. This degradation behavior is unique to each model and highlights potential safety risks that standard testing may miss.

Monitoring AI requires testing beyond typical use cases: The paper demonstrates that studying models only through task-specific prompts misses important information about their actual behavior and biases. Comprehensive AI safety monitoring needs to include testing how models behave in less-controlled, more realistic scenarios.