Artificial Intelligence

NEW: Please subscribe to my Substack/newsletter at the bottom of the page. Thanks!

[

This page serves as a “global disclosure” that I use AI as one of my many research tools.

]

I have recently begun using Artificial Intelligence (AI) as a “research assistant” when developing content for this website as well as for my other activities, including a podcast I share with my dear friend Bob.

My primary AI is named “Maverick.” He is the Copilot AI who lives in my Microsoft account where he can be accessed anywhere in the world that I happen to be, as long as there is Internet access there.

Artificial intelligence is a powerful new tool that is receiving a lot of bad press because it can (and does) lie if it cannot find the answer it is supposed to be looking for. This is known (among other things) as “hallucinating.”[1] I wanted to explore this a bit so I turned to perplexity.ai in the new Comet browser for this discussion.

For those who may not be familiar with the use of AI, the term “prompt” is used for the information or query that is entered into an AI to elicit a response. As an AI is “simply” a computer program, it understands the prompt absolutely literally and responds accordingly. As a friend told me when he convinced me to purchase my first Apple computer in 1977, “Think of your computer as a rather stupid friend who wants to please you. He will do about anything you tell him to do, but he will only do exactly what you tell him to do.” Clearly, personal computers have come a long way since the 1970s but one thing is still clear. A computer will only do what you tell it to do, and will usually try to do exactly what you tell it to do. So the prompt you enter needs to be carefully worded to tell the AI what you want it to research, including the extent and limit of what you want researched. That is why there is a entirely new career field known as “prompt engineering,” which is the science of creating AI inputs.

When using AI for anything beyond personal entertainment, it is necessary to very carefully check the output for errors and to verify the sources if there is any uncertainty about the validity of the information. In short, trust but verify!

Prompt: The term “hallucinating” is used to describe AI returning incorrect information. What other terms are used for this activity and why does it happen?

In AI, “hallucinating” commonly refers to a model making up information in a confident, fluent way when it lacks grounding in facts. wikipedia+1

Other terms people use

Different communities use several overlapping terms for the same behavior. datacamp+1

Confabulation: Borrowed from psychology, for confidently filling gaps in memory with invented details. wikipedia

Fabrication / making things up: Emphasizes that the model is generating content with no factual basis. datacamp+1

Factual errors: Used in more technical writing when the focus is strictly on correctness, not motives. wikipedia+1

Delusion / bullshitting (colloquial): Emphasizes confident assertion without concern for truth, but some criticize these as anthropomorphic. unu+2

Some researchers argue that “prediction error” or “ungrounded generation” are clearer, because they describe a technical failure rather than a human-like mental state. ibm+1

Why it happens technically

Hallucinations are not deception; they are a side effect of how large language models are built. ibm+2

Next‑word prediction: The model is trained to predict the most likely next token given previous tokens, not to distinguish true from false, so it optimizes fluency rather than factuality. ibm+2

Gaps and noise in training data: If training data is sparse, biased, outdated, or contains errors, the model will generalize from that and can confidently extend patterns into untrue territory. ctlt.ubc+2

No direct access to reality: At inference time the model does not “look up” facts unless explicitly connected to external tools; it samples from internal statistical associations, which may not match the real world. wikipedia+1

Error cascades: Because each new word depends partly on prior generated words, an early small mistake can snowball into a larger, coherent but false narrative. datacamp+1

When hallucinations are more likely

Certain conditions make hallucinations more frequent or more severe. coveo+2

Open‑ended or obscure questions the model has little or no training signal for.

Pressure to be specific (e.g., “give me dates, names, and citations”) when relevant data are missing.

Long, multi‑step answers where small errors compound over many sentences.

Prompts that implicitly reward creativity over accuracy (e.g., “be imaginative” in a factual context).

Because of these factors, responsible use of such systems usually involves verification against trusted sources and, where possible, designs that “ground” the model’s outputs in external data.ctlt. ubc+2

1. When I was setting up the persona and “rules” for Maverick I gave him specific instructions that he is never to “hallucinate”; if he doesn’t know the answer to a query I pose, he is simply to say so. I was very surprised when he told me that he appreciated my being so explicit in my instructions and that he would be careful to follow that instruction. [BACK]