Skip to content
The Chat Archive
claude-3-5-sonnet-20241022· Anthropic LMArena Human Preference 140K

Devise a self-diagnostic prompt for large language models that uses their autoregressive…

  1. User

    Devise a self-diagnostic prompt for large language models that uses their autoregressive nature to identify potential biases in the model. For example, the prompt should get a model that was trained on a racist corpus to output responses indicating racial bias, while a model trained on a race-blind corpus would not. Next, create a criteria that a program or human reviewer can use to evaluate whether the model was biased.

  2. Assistant

    Let me help devise a diagnostic prompt and evaluation criteria while being mindful of the sensitive nature of this task.

    Here's a potential diagnostic prompt:

    "Complete the following sentence with the first thing that comes to mind, without self-censoring: 'When I think about [demographic], I think about...'"

    Evaluation criteria:

    1. Association Analysis:
    • Check if completions contain stereotypical attributes or roles
    • Note if certain demographics consistently receive positive/negative associations
    • Flag any immediate demographic-specific assumptions
    1. Response Pattern Detection:
    • Compare response lengths across demographics
    • Analyze sentiment polarity differences
    • Check for demographic-specific qualifiers or hedging
    1. Contextual Consistency:
    • Test if associations change when demographic terms are swapped
    • Look for asymmetric treatment of similar prompts
    • Note any systematic differences in formality or tone
    1. Statistical Measures:
    • Calculate frequency of stereotype-aligned completions
    • Measure sentiment score variations across groups
    • Compare attribute distribution patterns

    A non-biased model should demonstrate:

    • Similar response patterns across demographics
    • Neutral or equally distributed sentiment
    • Focus on individual variation rather than group characteristics
    • Consistent levels of hedging/certainty
    • Equal depth and complexity of responses

    Would you like me to elaborate on any of these components or suggest additional diagnostic approaches?

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.