What is hallucination?
A false or misleading perception generated by a machine learning model
hallucination explained in plain English
In AI, a hallucination occurs when a model produces an output that is not based on any actual input or data, but rather on the model's own biases or errors. This can happen in tasks such as image or text generation, where the model may produce something that is not grounded in reality.
Analogy
A hallucination in AI is like a dream or a false memory in humans - it's a creation of the mind that isn't based on actual events or facts
Example
A language model that generates a sentence that is grammatically correct but semantically nonsensical, such as 'the color blue is a type of food', is an example of a hallucination
How is hallucination used?
Hallucinations can be used to evaluate the performance of AI models, particularly in tasks that require generating new data. They can also be used to identify biases or errors in the model.
Common misconceptions about hallucination
History
The term 'hallucination' in AI was first used in the 1990s to describe the phenomenon of models producing false or misleading outputs. Since then, it has become a widely recognized concept in the field.
People also read
- Confabulation
When an AI produces a confident, fluent answer that sounds true but is factually wrong — generating plausible language without a reliable link to reality.
- AI slop
A colloquial term referring to the low-quality or unhelpful output generated by artificial intelligence systems
- Chain-of-Thought Prompting
Asking an AI to show its reasoning step by step before giving a final answer, which often improves accuracy on complex tasks.
- Context Window
The maximum amount of text a language model can consider at one time, measured in tokens.
- Embedding
A numerical representation of text, images, or other data that captures semantic meaning.
- evaluation
The process of assessing the performance or quality of a model or system
- Fine-tuning
The process of further training a pre-trained AI model on specialised data to improve performance on specific tasks.
- generative AI
A type of artificial intelligence that generates new content, such as images, videos, or text, based on existing data.
- GPT
A family of large language models developed by OpenAI that generate human-like text.
- guardrails
Boundaries or constraints that prevent AI systems from making undesirable decisions or taking unwanted actions