What do you we already know?

• Implicit knowledge
  • Polyani
• Explicit knowledge
  • Encapsulated in the academic literature

What is the research hypothesis?

• What do we not know collectively as a human?
  • Don’t try to be a journalist

Research Methodology

• Strength and weakness of competing research methodologies? Single domain expertise is a big risk factor with AI

Mokyr

• “Useful knowledge”

  • knowledge that helps people understand or manipulate the world around them.

Prescriptive Knowledge ($\lambda$-knowledge):

• Knowledge of how to do something.

It is the recipe, technique, or invention itself.

• Example:
  • Adding a spoonful of starter curd to warm milk will transform it overnight. No understanding of microbiology is required; one simply has to follow the steps.

Prescriptive Knowledge ($\lambda$-knowledge):

• Knowledge of how to do something.

It is the recipe, technique, or invention itself.

• Example: Yoghurt
  • Adding a spoonful of starter yoghurt to warm milk will transform it overnight. No understanding of microbiology is required; one simply has to follow the steps.

• Example: Social Networks
  • How do social networks work over long distance?
  • How do you shape beliefs?

Prescriptive Knowledge ($\lambda$-knowledge):

• How

• 

What is the relationship between the different layers

• dynamic complementarity

1. Concept Explanation & Literature Mapping

• Tools can explain complex methods (e.g., ANOVA, mixed-effects models, p-values), suggest study resources, or help identify analytical frameworks used in published papers.

2. Specialised Scientific Discovery

• Platforms like Anthropic Claude Science help integrate fragmented databases, literature, and analytical tools in fields like biomedicine and formal sciences.

1. Concept Explanation & Literature Mapping

• Tools can explain complex methods (e.g., ANOVA, mixed-effects models, p-values), suggest study resources, or help identify analytical frameworks used in published papers.

2. Specialised Scientific Discovery

• Platforms like Anthropic Claude Science help integrate fragmented databases, literature, and analytical tools in fields like biomedicine and formal sciences.

Hallucinations and Methodological Misalignment

• LLMs predict text and code sequences based on patterns; they are not trained statisticians. They may apply the wrong statistical test to your data structure or misinterpret edge cases while sounding completely confident.

Causal and Contextual Blind Spots

• AI struggles with deep causal logic, study design nuances, and qualitative context, often favoring overly complex or popular methods over appropriate ones (complexity and popularity bias).

Keep a Human-in-the-Loop

• Treat AI as a research assistant, not the principal investigator.
  • You are fully responsible for the methodology, ethics, and validity of the findings.

Inspect and Run Code Independently

• Do not rely on black-box answers.
  • Have the AI generate clean scripts (Python/R), and execute and verify them outside the chat interface.