LLM Bias

An LLM Bias is an AI system bias that causes large language models to systematically favor certain outputs, perspectives, or patterns over others in ways that deviate from objective or balanced responses.

• AKA: Language Model Bias, LLM Systematic Bias, Large Language Model Prejudice, LLM Output Skew.
• Context:
  • It can typically arise from Training Data Biases in corpus selection.
  • It can typically manifest in Output Generation through preference patterns.
  • It can typically affect Decision Making in downstream applications.
  • It can typically require Bias Mitigation Techniques for fairness improvement.
  • It can often reflect Societal Biases from human-generated text.
  • It can often be measured using Bias Detection Methods and fairness metrics.
  • It can often persist despite Fine-Tuning Attempts at bias removal.
  • It can range from being an Explicit LLM Bias to being an Implicit LLM Bias, depending on its manifestation clarity.
  • It can range from being a Demographic LLM Bias to being a Linguistic LLM Bias, depending on its bias domain.
  • It can range from being a Harmful LLM Bias to being a Benign LLM Bias, depending on its impact severity.
  • It can range from being a Detectable LLM Bias to being a Hidden LLM Bias, depending on its observability level.
  • ...
• Example:
  • Content Generation Biases, such as:
    • LLM Plausibility Bias favoring plausible-sounding over accurate content.
    • LLM Frequency Bias favoring common patterns over rare cases.
    • LLM Length Bias favoring verbose outputs over concise responses.
  • Demographic Biases, such as:
    • LLM Gender Bias in pronoun associations.
    • LLM Cultural Bias in value judgments.
  • Task-Specific Biases, such as:
    • LLM Confirmation Bias reinforcing initial assumptions.
    • LLM Anchoring Bias overweighting early information.
  • ...
• Counter-Example:
  • Random Error, which lacks systematic pattern.
  • Human Cognitive Bias, which stems from biological cognition rather than algorithmic training.
• See: AI System Bias, LLM Plausibility Bias, LLM Error, Bias Mitigation, Fairness in Machine Learning, Training Data Bias, Algorithmic Bias.

---