Appendix D: Glossary

AI

Artificial intelligence leverages computers and machines to attempt to mimic the problem-solving and decision-making capabilities of the human mind.

Tasks may require human abilities such as perception, reasoning, problem solving, and understanding natural language. Large collections of data as well as new experiences are used by algorithms to find patterns and use them to take actions or make predictions/provide insights (IBM, 2023).

AI Forensics

This refers to the use of forensic techniques to identify if text was AI-generated and then the source of the AI product.  Once the source is known, it can be checked for accuracy and credits. This can ultimately reveal the bias in the training data set (Martineau, 2023).

Algorithm

A set of step-by-step directions for solving a problem or accomplishing a specific task (Berkman Klein Center, 2019).

As an example, here’s a simple computer algorithm for finding the highest number in a list:

1. Start with the first number in the list and remember it as the current, most significant number.
2. Compare the current largest number with the next number in the list.
3. If the next number is larger than the current most significant number, update the current largest number to be the next number.
4. Repeat steps 2 and 3 for all the numbers in the list.
5. When you reach the end of the list, the current largest number will be the largest number in the list.

Deep Learning

A subset of machine learning using a neural network with at least three layers. Deep learning distinguishes itself from classical machine learning by the type of data that it works with and the methods in which it learns.

Machine learning algorithms leverage structured, labeled data to make predictions—meaning that specific features are defined from the input data for the model and organized into tables; there is generally some pre-processing to organize the data into a structured format.

Deep learning eliminates some of data pre-processing that is typically involved with machine learning. These algorithms can ingest and process unstructured data, like text and images, and it automates feature extraction, removing some of the dependency on human experts. For example, let’s say that we had a set of photos of different pets, and we wanted to categorize them by “cat”, “dog”, “hamster”, et cetera. Deep learning algorithms can determine which features (e.g. ears) are most important to distinguish each animal from another. In machine learning, this hierarchy of features is established manually by a human expert” (IBM, n.d.).

Deepfake

Deepfakes are digital forgeries that use artificial intelligence to create believable but misleading images, audio, and videos (Personal and Security Research Center, 2022).

Generative AI

Generative AI is a type of AI system capable of generating text, images, or other media in response to prompts. It uses its collection of data and experiences to generate new content. Generative AI is different from General AI (see below) (Benefits and Limitations, 2023).

General AI / Artificial General Intelligence (AGI)

General AI refers to the development of AI systems that possess human-level intelligence across a broad range of tasks and domains. AGI aims to create machines that can understand, learn, and perform complex cognitive functions that mimic human intelligence. This is in comparison to the specific, task-focused output of Generative AI (Mock, 2023).

Hallucinations

Since generative AI is based on statistical patterns, it may not always produce accurate or meaningful results. “Hallucinations” refers to computer-generated information that does not correspond to objective reality (Mair, 2023; Alkasissi & McFarlane, 2023).

Large Language Model (LLM)

A deep learning algorithm that can recognize, summarize, translate, predict and generate text and other forms of content based on knowledge gained from massive datasets (Lee, 2023).

Machine Learning

A subfield of AI where a computer imitates human learning using data and algorithms to gradually improve its accuracy without additional programming changes or corrections (IBM, n.d.; Brown, 2021).

Model

A program that is trained on a dataset to identify patterns and, possibly, make decisions (IBM, n.d.). In machine learning, algorithms that go through training develop an understanding of a topic, or their own “model” of the world.

Multimodality

The ability of generative AI to provide responses based on multiple types of prompting material that can include images, audio, and possible others, as well as text (Infusion, 2023).

Natural Language Processing (NLP)

A branch of artificial intelligence that enables understanding and production of human language by a computer (IBM, n.d.). It became popular in smart speakers like Siri and Alexa.

Neural Network

Mathematical models for programming inspired by the human brain, primarily for problem solving and pattern recognition. These can be fairly simple or include multiple internal layers meant to increase learning capacity, efficiency, and accuracy (Zwass, 2023). Data fed to a neural network is reduced into smaller pieces and analyzed for underlying patterns, often from thousands to millions of times depending on the complexity of the network. A deep neural network is when the output of one neural network is fed into the input of another, chaining them together as layers. Various types of neural networks include:

• Convolutional Neural Network (CNN): A neural network with the ability to process dense data, such as millions of pixels or tens of thousands of audio samples. These networks are primarily used to recognize images, video, and audio data. They are effective for learning with less parameters, however are comparatively slow and complex to maintain.
• Deep Neural Network: A neural network that contains more than one hidden layer, i.e., layers between the input and output layers.
• Generative Adversarial Network (GAN): A system in which two neural networks, one that generates an output and another that checks the quality of that output against what it should look like. For example, when attempting to generate a picture of a cat, the generator will create an image, and the other network (called a discriminator) will make the generator try again if it fails to recognize a cat in the image.
• Recurrent Neural Network (RNN): A neural network frequently employed for natural language processing. It analyzes data cyclically and sequentially, i.e., it can process data such as sentences or words while maintaining their order and context in a sentence.
• Long Short-Term Memory Network (LSTM): A variation of a recurrent neural network that is intended to retain structured information based on data. An RNN could recognize all the nouns and adjectives in a sentence and if they’re used correctly, but an LSTM could remember its placement in a book.

Prompt

Prompts are the requests/information we provide to AI to let it know what we’re looking for. They may be snippets of text, streams of speech, or blocks of pixels in a still image or video. The importance of an effective prompt has generated a new job – Prompt Engineer. (Martineau, 2023; Popli, 2023; Shieh, 2023).

Prompt Injection Attack

A prompt injection attack crafts a prompt that causes the AI tool to provide output that has been forbidden by its training (Selvi, 2022).

Puppeteering

Puppeteering refers to the manipulation of full-body images to perform actions and behaviors determined by AI (like a puppeteer). It is also known as full body deepfakes. For example, the image of someone who has two left feet when it comes to dancing could be made to perform as if they were a talented dancer (Jaiman, 2022).

Training

The process of supplying data (usually data sets) to an algorithm for it to learn. It can apply to various machine architectures, or models, including neural networks. Three styles of machine learning training include:

• Supervised Learning: The data fed to the algorithm to process is already organized and labeled. The goal of most supervised learning is to predict output based on known input. For example, if you are training a network with supervised learning to identify cars by providing the model with thousands of images labeled “cars”. The approach helps minimize error between the outputs and the actual images. Common techniques of supervised learning include: linear regression, logical regression and decision trees.
• Unsupervised Learning: The algorithm is not provided any information about how it should categorize the data it is given, but instead it must find relationships and classify the unlabeled data. The goal of employing this style of learning is to identify valuable relationships between input data points. These relationships can then be applied to new input. This approach finds spatters, anomalies, and similarities within the data. Common techniques include: clustering, probability density, and associated rule learning.
• Self-Supervised Learning: The algorithm uses the structure within the data, its own inputs or modifications, to generate labels. Self-supervised learning typically identifies a secondary task where labels can be automatically obtained, and then trains the network on the secondary task. While the targets are missing in both unsupervised learning and self-supervised learning, the latter uses the data itself to generate the supervisory signals.