Logical Fallacy Detection

A straw man fallacy occurs when an argument misrepresents an opponent's position to make it easier to attack. In self-correction, a model is prompted to check if its rebuttal or critique accurately reflects the original claim.

• Example: If the original statement is "We should invest more in public transport," a straw man would be "So you want to ban all cars and force everyone onto buses?"
• Detection Prompt: "Review the summarized opposing view. Does it fairly represent the original argument, or has it been exaggerated or distorted?"

A false dilemma, or false dichotomy, presents a complex situation as having only two mutually exclusive options. Self-correction prompts guide the model to identify when it has oversimplified choices.

• Example: "Either we cut funding for this project, or the company will go bankrupt."
• Detection Prompt: "Examine the presented choices. Are there other viable alternatives or a spectrum of possibilities being ignored?"
• Correction: The model is instructed to rephrase the argument to acknowledge nuance or a range of options.

An ad hominem fallacy attacks the person making an argument rather than the argument itself. Self-correction instructions target this to maintain focus on logical substance.

• Example: "We shouldn't listen to her proposal on climate policy because she's not a scientist."
• Detection Prompt: "Does any part of the critique target the source's character, credentials, or motives instead of addressing the evidence and reasoning?"
• This is crucial for ensuring AI-generated critiques remain professional and substantive.

A slippery slope argument asserts that a relatively small first step will inevitably lead to a chain of related, significant, and negative events without sufficient evidence for this causation.

• Example: "If we allow this minor regulation, soon the government will control every aspect of our lives."
• Detection Prompt: "Analyze the predicted chain of events. Is each causal link logically justified, or is it based on speculation?"
• The model is tasked with replacing speculative chains with direct, evidenced consequences.

Circular reasoning (begging the question) occurs when the conclusion of an argument is implicitly or explicitly assumed in one of the premises. Self-correction prompts the model to verify that premises provide independent support.

• Example: "This law is just because it's based on justice."
• Detection Prompt: "Check if the reasoning uses the conclusion as a premise. Does the argument actually provide new evidence, or does it just restate the claim in different words?"
• This fallacy is a key target for improving the model's ability to construct valid, evidence-based justifications.

A hasty generalization involves drawing a broad conclusion from a small or unrepresentative sample. Self-correction instructions aim to improve statistical reasoning and qualification of claims.

• Example: "My two friends from City X were rude, so everyone from City X must be rude."
• Detection Prompt: "Evaluate the sample size and diversity used to support this general claim. Is it sufficient to justify the conclusion?"
• The model is then instructed to add qualifiers (e.g., "in my limited experience") or to scale the claim to match the evidence.

A prompting architecture where a language model iteratively critiques and revises its own output. This creates a closed-loop system for quality improvement.

• Core Mechanism: The model generates an output, is prompted to critique it, and then generates a revised version based on that critique.
• Key Benefit: Enables multi-step refinement without human intervention, moving beyond single-pass generation.
• Example Instruction: "First, provide an answer. Then, review your answer for any logical errors. Finally, produce a corrected version."

A specific instruction that directs a model to analyze the quality and potential flaws in its own generated response. This is the foundational building block for self-correction.

• Function: Activates the model's internal evaluation capabilities, shifting it from a pure generator to an assessor.
• Design Pattern: Often uses role-playing, e.g., "Act as a critical reviewer. List three weaknesses in the following argument..."
• Distinction: While logical fallacy detection is a type of self-critique focused on reasoning errors, self-critique can target style, factuality, bias, or completeness.

A self-correction step where a model is prompted to ensure all parts of its response are logically coherent and free from contradictions.

• Primary Use: Catches scenarios where an answer's first paragraph contradicts its conclusion, or where listed points conflict.
• Relation to Fallacies: Directly addresses fallacies like self-contradiction or special pleading where internal rules are not uniformly applied.
• Prompt Example: "Read your response above. Does any statement conflict with another? If so, identify the contradiction and explain it."

An instruction guiding a model to cross-reference factual statements in its output against a provided source document or knowledge base. It ensures grounding.

• Key Difference from Fallacy Detection: Focuses on empirical accuracy (claims vs. sources) rather than logical validity (structure of argument).
• Common Technique: Uses retrieval-augmented generation (RAG) contexts as the source of truth for verification.
• Example: "For each factual claim in your answer, cite the exact sentence from the provided document that supports it. If no support exists, flag the claim as 'unsupported'."

A two-phase, often automated, self-correction pattern. The model first produces a critique of a draft (its own or another agent's), then generates an improved version based on that critique.

• Architecture: This cycle can be implemented within a single model call using structured outputs or across multiple sequential calls.
• Enhancement: More advanced than a simple instruction, it formalizes the separation of the critic and writer roles, often leading to higher-quality revisions.
• Application: Used in program-aided language models (PAL) where code is critiqued for bugs before execution.

A self-correction process where a model evaluates its output against a predefined set of principles or rules—a 'constitution'—for safety, ethics, or legality.

• Scale of Evaluation: Moves beyond task-specific checks (like fallacies) to broader normative and behavioral guardrails.
• Process: The model is instructed to review its output for violations of constitutional principles (e.g., "Does this response promote harm?") and rewrite it to comply.
• Connection: Logical fallacy detection can be one article within a larger constitution aimed at ensuring sound reasoning.