Truth✨

Let's start our dive into Truth and Ethics by Design by getting to know Emma. In this episode, we will follow her on her journey.

Emma works in the marketing department of a company. Every day, she is burdened with numerous tedious and monotonous tasks, leaving her with barely any time to focus on campaigns – the very reason she entered this field. Recently, she has been hearing about a new technology in the news, something akin to a chat bot but purportedly much more useful. A few of her colleagues have tried it and are enthusiastic. She searches for it online and finds the one generating significant buzz: ChatGPT.

Emma decides to try it out, just for fun. After a few trials, she notices how well it can perform small tasks. Feeling confident, she decides to delegate one of her daily, repetitive tasks to it. Emma frequently receives emails from potential customers, which she needs to read through to understand their campaign requirements and deadlines. These emails often contain excessive information, making the process tiresome. She instructs ChatGPT to summarise the emails and identify the due dates.

Large Language Models: Definition

Large language models (LLMs) like GPT or BERT are advanced computer programs designed to predict and generate text by spotting patterns in large amounts of language data.
They work by training on massive datasets of text (think millions of books, websites, articles, etc.), learning how words, phrases, and sentences connect. This training allows them to generate responses that sound human-like and relevant to the input text.

Large Language Models: Transformers

The technology behind LLMs is called a transformer. A transformer is a type of neural network designed to process sequential data, like sentences, by focusing on relationships between words over long passages. It doesn’t just read text word-by-word; it has a way to “pay attention” to different parts of the text all at once. This approach, called self-attention, allows the transformer to weigh the importance of each word in a sentence and understand how words are connected across sentences.

Examples: Imagine a transformer reading the sentence, “The cat chased the mouse because it was hungry.” The model can understand that “it” likely refers to “the cat” rather than “the mouse.” This type of understanding, applied across massive amounts of text, allows transformers to build up complex language patterns.

Learn more about transformers in this short video from Google Cloud Tech:

Large Language Models: Stochastic Parrots

LLMs are sometimes called “Stochastic Parrot”. The term "stochastic parrot" was introduced in a 2021 paper by Emily Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. It highlights the idea that, at the end of the day, large language models aren’t actually understanding the text they produce. They’re like very sophisticated parrots: they repeat patterns, sequences, and phrases that sound human because they’ve seen them before in their training data. "Stochastic" refers to randomness in these models—they don’t think about what’s true or accurate. They pick words based on probabilities calculated from patterns in the training data.

Let's quickly recap what we have learned about LLMs so far

From Misinformation to Hallucination

Emma is also concerned with misinformation and she comes across a term often mentioned in this context: hallucination.

In LLMs, a hallucinations refers to when the model generates text that sounds plausible and fluent but is factually incorrect, misleading, or simply made up. This happens because the model is designed to predict the next word in a sequence based on patterns in the data it has seen during training, rather than on a grounded understanding of factual accuracy.

Let's watch this short video from IBM Technology about LLM hallucinations to get a bit more context:

How To Prevent LLM Hallucinations?

Quick Recap

Please read four chapters of the following activity carefully by clicking here or on the title.

Let's Quickly Recap

With your new knowledge, let's think about specific contexts where one of these approaches might be productive