AI hallucinations may occur for text, audio, and image outputs generated by Large Language Models (LLMs) and AI tools including chatbots and image generators. Hallucinations occur when patterns or objects are incorrectly perceived by the model and the resulting output may contain surreal, nonsensical, or inaccurate elements. AI won’t request clarification but may make up substitutions.
Consequences and Risk:
AI hallucinations may result in generating misleading information, the spread of misinformation, or outputs that reflect training data bias.
Causes of AI Hallucinations:
- Overfitting where the model is unable to make predictions outside of its training data.
- Insufficient training data including data bias, data inaccuracies, lack of data diversity and representative datasets covering real-world scenarios, and/or a lack of well-structured data.
- High model complexity.
- AI models that lack a well-defined purpose or limitations of use, including the incorporation of domain-specific constraints, rules to guide the generation process, or a lack of constraints that limit possible outcomes.
- Lack of testing and refinement of AI models including ongoing evaluation of outcomes.
- Lack of human intervention and review to identify and correct hallucinations.
- Lacking the incorporation of adversarial training techniques to defend against potential attacks designed to improve the model’s resilience to unexpected inputs.
Some AI models are designed intentionally to allow for hallucinations. Applications of such models may occur in creative pursuits.
