Tornado Alley (PROG)

[janice shell]

Sure they might make mistakes, but no way are they not relatively reliable.

Check this out:

Confabulation: The Surprising Value of Large Language Model Hallucinations

Peiqi Sui, Eamon Duede, Sophie Wu, Richard So

https://aclanthology.org/2024.acl-long.770/

An interesting perspective, but...

This, however, is a good explanation of why they do what they do:

Are AI Models Hallucinating—Or Just Confabulating?

June 10, 2025

https://cosimameyer.com/post/are-ai-models-hallucinating-or-just-confabulating/

Note this especially:

From a technical perspective, this phenomenon is inherent to how these models operate. The generation process relies on probability: at each step, the model predicts the most likely next word based on the preceding context.14 Essentially, with each new word generated, the model assigns a probability to it, favoring words with higher probabilities. If the probability at any point drops too low, the model stops generating text. While this probabilistic approach is fundamental, it can lead to factually incorrect outputs—a consequence generally undesirable, unless the goal is creative or fictional writing.

There's much, much ore on the subject...