Large Language Models as auto-completes

In this section, we are going to take a deep dive into Large Language Models (LLMs). We will explore their auto-complete nature, and the term “hallucinations” often used to describe their behaviour.

LLMs as Auto-Completes

With the four key ingredients of AI previously discussed, these systems can accomplish extraordinary feats. Specifically, we will focus on text generation as our illustrative example.

TASK: Open a chosen LLM and type in three or four words that may begin a story. For example, ‘The owl hooted…’, or, ‘The light flickered…” See what you get.

*Note, sometimes this works, and sometimes it doesn’t. This illustrates the non-deterministic nature of AI systems.

At their core, LLMs function as highly sophisticated auto-complete systems similar to the predictive text feature on a smartphone, but much much more powerful. Despite their apparent simplicity, the ability to predict the next word in a sequence underpins the astonishing range of tasks these models can perform. 

TASK: Ask your chosen LLM “How is it that by simply predicting the next word in a stream of text, an algorithm like an LLM can do all the things it can do?”

*Note, the above is an example of reinforcement-learning with human feedback (RLHF).

A key reason (answering the above question) lies in how AI companies have framed our interaction with these models. Rather than presenting them explicitly as advanced auto-complete systems, they are packaged within a chat interface – an intuitive format that fundamentally shifts our perception of these systems, and potentially leads us to anthropomorphise them (addressed in greater detail in the AI and Existential section of the course, Week 2). This idea of the contextual implications of our interactions with AI is aptly explored in the podcast, Philosophize This!, Episode 183, “Is ChatGPT really intelligent“.

With the four key ingredients of modern AI previously discussed, LLMs have become the powerful tools many of us are now familiar with, and the framing of these LLMs within a chat window significantly influences how we engage with them and what we expect from them.

LLMs and Hallucinations

The true marvel of LLMs lies in how convincingly this probability-driven process can simulate thoughtful and coherent responses. LLMs do not think, reason, or understand in the way humans do. When asked a question, an LLM does not engage in introspection or logical deduction. Instead, it simply identifies the most statistically probable sequence of words based on its training data. 

These models lack basic truth, a structured understanding of the world, or intrinsic rules of logic. Their words do not refer to real-world entities; they are merely patterns of statistically likely text. When an LLM generates incorrect information, it is not attempting to deceive as it has no conception of truth or falsehood. It is simply predicting words that are plausible based on prior patterns in its dataset.

This is why LLMs often generate confidently incorrect statements. When an LLM fabricates a falsehood, it is not malfunctioning; it is doing precisely what it was designed to do: produce text that appears authoritative, regardless of factual accuracy. In computer science, this phenomenon is termed “hallucination”: when an AI confidently generates incorrect, misleading, or entirely fictional information. 

It is worth noting here, that even the use of the word “hallucination” is contested, because it implies some kind of intention behind LLM auto-complete. In other words, by using the word “hallucination” we are doing exactly what those designing LLMs desire, we are anthropomorphising AI systems (addressed in greater detail in the AI and Existential section of the course, Week 2).

Catching an AI in the act of Hallucinating 

On some LLM platforms, you can show the internal “reasoning” that the AI is going through. Taking us back to the task of asking an AI to auto-complete a story, using the first few words as a prompt. In my example, I don’t remember mentioning Elara, can the AI explain its “reasoning” for its particular auto-complete? What you are seeing below is a fresh conversation…

An interesting method of exploring how these auto-completes work is by playing with them. https://deepinfra.com/chat is a useful resource that allows you to play with some of the “black box” settings that define an LLM’s output. 

TASK: Explore working with an LLM with non-default settings and see just how fragile these systems are. Note, no matter what you do, this is still an advanced auto-complete system working with the same model. The settings we are allowed to change are very surface level.

* There is no specific, expected learning outcome associated with this task. Just play around, explore what happens and what you learn about the auto-complete and “hallucinatory” nature of LLMs.

Auto-completing an Image

A liar knows the truth but actively misleads. An “hallucinator”, on the other hand, does not know or care about the truth, it merely aims to produce text that sounds convincing. This leads to a significant challenge: misinformation generated by AI is effortless to produce but difficult to correct. This problem is best captured by Brandolini’s Law, which states:

“The amount of energy needed to refute [a hallucination] is an order of magnitude larger than that needed to produce it.”

This misinformation isn’t restricted to text generation either. The following images, published by Our World in Data in 2022, highlight the development of AI image generation over the past decade. All these images were generated by an AI system.

Image from Our World in Data

Detecting AI-Generated Content

We have clearly moved on from the world of 6-fingered AI-generated images, and clearly identifiable AI-generated text. AI-prepared content may consist of text, an image, or something more devious such as a full, autonomous AI agent acting on the internet. Definitively detecting AI-generated content is now extremely challenging given the non-deterministic nature of AI systems. AI purporting misinformation, has been identified as one of the primary risks associated with AI systems today, including in the notable paper by Weidinger et al., “Taxonomy of Risks posed by Large Language Models” in FAccT in 2022. 

This idea of the risks of AI-generated misinformation will be covered in greater detail in the AI and Governance sections of the course (Week 2). 

Another excellent resource published by Bergstrom and West in 2025, “Modern-day Oracles or Bullshit Machines?” has partly inspired (and also compliments) this section of the course and is well worth exploring in more detail.

A further, interesting, and somewhat technical concept around identifying AI “hallucinations” in images is the idea of “A Watermark for Large Language Models“, published in 2024 by Kirchenbauer and Geiping, University of Maryland.