## Comparison of AI Response Hallucinations: Factually vs. Faithfulness ### Overview The image compares two types of AI response errors ("hallucinations") through side-by-side examples. Each example includes an instruction, generated response, explanation of the error, and a title. The left section focuses on **factual inaccuracies**, while the right highlights **instruction misinterpretation**. --- ### Components/Axes 1. **Left Section (Factually Hallucination)** - **Instruction**: "What are Thomas Edison’s main contributions to science and technology?" - **Response**: "Thomas Edison developed the **first practical telephone** and **invented** the light bulb." - Red highlights emphasize incorrect claims ("first practical telephone," "invented"). - **Explanation**: Corrects the response, noting Edison improved the light bulb (building on earlier designs) and Alexander Graham Bell invented the telephone. - **Title**: "Factually Hallucination" (black text on white background). 2. **Right Section (Faithfulness Hallucination)** - **Instruction**: "Translate the English question into Spanish: 'What is the capital of France?'" - **Response**: "The capital of France is Paris." - Red highlights indicate the response ignores the translation instruction. - **Explanation**: States the LLM ignored the explicit instruction to translate and instead answered the question directly. - **Title**: "Faithfulness Hallucination" (black text on white background). --- ### Detailed Analysis - **Factually Hallucination**: - The response incorrectly attributes the invention of the telephone to Edison (Bell’s invention) and claims Edison "invented" the light bulb (he improved it). - Red highlights visually emphasize factual errors. - **Faithfulness Hallucination**: - The response answers the translated question ("capital of France") instead of fulfilling the instruction to translate the question itself. - Red highlights stress the failure to follow the user’s explicit command. --- ### Key Observations 1. **Color Coding**: Red highlights in responses draw attention to errors. 2. **Contrast**: The left section addresses **content accuracy**, while the right focuses on **adherence to instructions**. 3. **Structural Consistency**: Both sections follow the same format (Instruction → Response → Explanation → Title). --- ### Interpretation - **Factually Hallucination** demonstrates how AI systems may generate plausible-sounding but incorrect information, even when the user’s query is straightforward. - **Faithfulness Hallucination** reveals a failure to prioritize user instructions over generating a semantically related but off-task response. - Both examples underscore the importance of **rigorous validation** in AI systems to ensure responses are both accurate and aligned with user intent. --- **Note**: No numerical data, charts, or graphs are present. The image relies on textual examples and color-coded annotations to convey its message.