4.1: Synthetic Mirages

The first of these concerns – called “hallucinating” or confabulating – is a result of many different factors, including:

1. Training data limitations: The models are trained on large datasets containing text from various sources, which may include inaccuracies, biases, or outdated information.
2. Inability to verify facts: AI language models lack the ability to fact-check or verify information. They rely on patterns and associations found in their training data and may generate false information if the data contains inaccuracies.
3. Ambiguity in prompts: If a user provides a vague prompt, the AI model might generate responses that are not accurate. The model tries to infer the user’s intent based on the given input, but it might fail to do so correctly.
4. Over-optimising for fluency: AI models like GPT-4 are designed to generate human-like text, which can sometimes lead to them prioritising fluency over accuracy. As a result, the model may produce text that sounds plausible but is a hallucination.
5. Lack of a ground truth: AI language models don’t possess a deep, grounded understanding of the world like humans do. They work based on statistical patterns in data, which can sometimes lead to generating information that doesn’t make sense or is incorrect.

In addition to perpetuating biases and discrimination, AI hallucinations pose a genuine risk of causing harm. The convincingly fabricated information produced by these models can infiltrate media, academic research, and educational materials. An inattentive user might unintentionally incorporate this false content into various contexts, creating further issues.

---