Glossary of AI
Hanna Barakat & Archival Images of AI + AIxDESIGN / Better Images of AI / Textiles and Tech 2 / CC-BY 4.0.
AI news is often packed with buzzwords and technical terms, making it feel like you need a computer science degree just to keep up. That’s where our AI glossary comes in - we break down complex AI terms into simple, easy-to-understand definitions, helping you confidently navigate AI in everyday life.
A
Agentic AI
An AI system that can make decisions and take actions on its own to achieve specific goals. Unlike traditional AI, which follows fixed instructions, agentic AI can assess situations, adapt to changes, and make choices without constant human input. It uses techniques like reinforcement learning and planning to operate independently. This type of AI is used in robotics, self-driving cars, and digital assistants that need to solve problems and make decisions on their own.
Algorithm
A defined set of mathematical formulas or instructions designed to complete a task or solve a problem. Algorithms are crucial in computers and programming, instructing the computer step by step. Their main purpose is to make accurate predictions or reach desired results, making algorithms key to efficient software and system creation.
Algorithmic Bias
Unfairness that occurs when an algorithm’s process or implementation is flawed, causing it to favour or harm one group of users over another. This often happens because the data used to train the algorithm is biased, reinforcing existing prejudices related to race, gender, sexuality, disability or ethnicity.
Artificial Intelligence (AI)
An AI system is a machine-based system that can, for a given set of human-defined explicit or implicit objectives, infer, from the input it receives, how to generate outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments. Different AI systems are designed to operate with varying levels of autonomy and adaptiveness after deployment. This definition is based on the Organisation for Economic Co-operation and Development (OECD) definition of AI, as of 2023.
Artificial General Intelligence (AGI)
A type of AI that could learn to accomplish any intellectual task that humans can perform, or surpass human capabilities in many economically valuable tasks. While AGI is often mentioned in the news and on social media, there is currently no research that proves how AGI could be developed or achieved.
Artificial Narrow Intelligence (ANI)
Artificial Narrow Intelligence (ANI), also known as weak AI or narrow AI, is designed to perform a specific set of tasks. All AI in existence today is narrow AI, usually using machine learning or deep learning techniques. Examples of narrow AI include internet search engines, recommendation systems and facial recognition. Such AI tools are designed to perform tasks within a single, defined set of problems.
Artificial Superintelligence (ASI)
Artificial Superintelligence (ASI) is a speculative concept of AI that would far surpass human intelligence, exceeding in memory, data-processing and decision-making abilities.
Automation
The use of technology to handle tasks with minimal human help. This includes software that automates business processes, robots working in factories and smart home devices. By automating repetitive tasks, organisations and individuals can work faster, make fewer mistakes and save money. Automation can be done without AI (for example, a sewing machine) or with AI (for example, language translation or speech transcription) and is used in many fields, from finance to healthcare, to cut costs and improve efficiency.
B
Big Data
A broad field of research that deals with large datasets. It has grown rapidly over recent decades as computer systems have become capable of storing and analysing vast amounts of data collected about our lives and our planet. A key challenge in big data is generating useful insights without compromising the privacy of the individuals to whom the data relates. As AI models require vast amounts of data to learn and improve, big data has enabled the development of AI. However, the term big data has lost some popularity to AI in recent years.
C
Chatbot
A software application that has been designed to mimic human conversation, allowing it to talk to users via text or speech. Previously used mostly as virtual assistants in customer service, chatbots are becoming increasingly powerful and can now answer users’ questions across a variety of topics, as well as generating stories, articles, poems and more. While the most well known chatbot is currently ChatGPT, chatbots have been used in the health sector for some time. For example, from 2019 to 2022, NHS Western Isles trialled a Gaelic chatbot to support mental well-being for people living in rural areas.
ChatGPT
An AI chatbot, developed by a company called OpenAI, which uses large language models (LLMs) to enable natural and human-like conversations. Apart from answering questions, ChatGPT can accomplish various other tasks, such as composing essays, writing code or having philosophical discussions with the user. Since 2018, OpenAI has officially released four iterations of their GPT model: GPT-2, GPT-3, GPT4 and GPT-4o. Many companies are developing their own chatbots, similar to ChatGPT – these include Google Gemini, Claude 3 by Anthropic and Microsoft Copilot.
Computer Vision
A form of AI used to understand and recognise images and videos and to analyse the elements of the content within them. For example, Google Photos uses computer vision to categorise photo files by their subject matter, grouping pictures of pets, people, landscapes or food together. Facebook also uses a form of computer vision to recognise faces in photographs and prompt you to tag someone. Computer vision can also be used for more complex analysis of images, such as using satellite imagery to map biodiversity by recognising characteristics of the landscape. Space Intelligence and Scottish Wildlife Trust are using computer vision to interpret large volumes of satellite data and map wildlife habitats to help restore, connect and protect Scotland’s natural environment.
D
Data
Any information that can be used for analysis, calculation and interpretation. This means any type of information – letters, numbers, words, sounds or videos – arranged so that it can be useful for some purpose. Data can come from many sources and be stored in different ways.
In an AI system, data is the information used to train an algorithm. The quality and suitability of the data used to build an AI system have a significant impact on how the system will perform – in other words, to do the job it is meant to do.
Data can be structured or unstructured. In structured data, all the data is in the same format and highly organised. Usually structured data is made up of numbers (also called quantitative data) – a large spreadsheet full of numbers is an example. Unstructured data tends to be in less easily searchable formats, such as video or audio. This type of data usually contains things we cannot easily count, like language, thoughts, feelings and behaviours (also referred to as qualitative data).
AI systems tend to require extremely large datasets, or collections of data. The larger and better the data, the more powerful the model, and the more able it is to make useful predictions. AI systems can work out how to better predict results by using massive amounts of processing power and analysing bigger datasets than a human ever could. However, if the data that is used has issues (such as being biased or incomplete), then so will the outputs or predictions.
Data scraping
AI systems, particularly large language models (LLMs), rely on vast amounts of data to learn and improve their capabilities. Data scraping is a key method for collecting this information from websites and other online sources, by extracting structured or unstructured data from the internet, either manually or with data scraping tools. This scraped data enables AI models to better understand language, context and real-world knowledge. At the same time, AI techniques are being used to make data scraping more efficient and adaptable. However, this practice raises important privacy and ethical concerns, and the large scale collection of personal data for AI training has led to legal challenges and regulatory scrutiny.
Deepfake
Synthetic image, audio or video in which someone is digitally altered so that they look, sound or act like someone else. Created by machine learning algorithms, deepfakes have raised concerns over their use in fake celebrity pornography, financial fraud and the spreading of false political information. Deepfake can also refer to realistic but completely synthetic media of people and objects that have never physically existed, or to sophisticated text generated by algorithms.
Deep Learning
A more recent variation of neural networks, using many layers of artificial neurons (forming a ‘deep neural network’) to solve more difficult problems. Its popularity as a technique increased significantly from the mid-2000s onwards, and it is behind much of the wider interest in AI today. Deep learning is often used to classify information from images, text or sound.
E
Existential risk
In the context of AI, existential risk refers to potential situations where the development or use of AI presents a risk to the well-being of humanity, particularly where it could pose a threat to the survival of humanity or create a worldwide catastrophic event.
Expert systems
Expert Systems are a type of Artificial Narrow Intelligence (ANI) application designed to replicate human decision-making processes. These systems rely on large databases and predefined rules to make informed decisions. Due to their narrow scope, they cannot adapt beyond the specific tasks they are programmed to handle.
G
Generative AI
An AI system that generates text, images, audio, video or other media in response to user prompts. GenAI uses machine learning techniques to create new data that has similar characteristics to the data it was trained on, resulting in outputs that are often indistinguishable from human-created media (see deepfake).
Generative Pretrained Transformer (GPT)
A family of large language models (LLMs) designed to understand and generate humanlike text based on the input they receive. GPTs are ‘pre-trained’ on enormous datasets, enabling them to perform a wide range of tasks, such as answering questions, translating languages, summarising text and generating creative content.
H
Hallucination
In the context of AI, particularly in large language models (LLMs), hallucination refers to the phenomenon where AI systems generate false, misleading or nonsensical information and present it as factual. This is a significant challenge in the AI field, especially with regard to chatbots and other AI-powered tools that interact with users and provide information, because results that are inaccurate may nonetheless sound plausible and authoritative.
L
Large Langage Model (LLM)
A type of model trained on a vast amount of textual data in order to carry out language-related tasks. LLMs power the new generation of chatbots and can generate text that is indistinguishable from human-written text. They are part of a broader field of research called natural language processing (NLP) and are typically much simpler in design than smaller, more traditional language models.
Limited Memory
Limited Memory AI uses past data to improve decision-making in future tasks, learning from previous interactions to refine their outputs. A common example of an AI system using Limited Memory is ChatGPT.
M
Machine Learning
A type of AI that has the capacity to learn without following explicit instructions. Machine learning systems use algorithms and statistical models to identify patterns in data and make predictions based on those patterns. There are several methods for training machine learning models, such as supervised learning, unsupervised learning and reinforcement learning.
Model
In an AI system, the program or arrangement of algorithms that is used to turn the inputs into the desired output or prediction. The term is most often used to refer to a machine learning model. Although these machine learning models are built using computer code, they are not explicitly programmed in the same way as a traditional computer program, which is instructed on what steps to take to achieve an exact result.
For example, a traditional computer program could find the cheapest flight from London to Edinburgh by comparing prices across several websites. An machine learning model could use algorithms to predict the best times of the day and year for buying a cheap flight from London to Edinburgh, and which websites are most likely to have the cheapest flights
N
Natural Language Processing (NLP)
NLP involves using AI to enable computers to read, understand and translate human speech or text. NLP systems can identify words and interpret meaning from sentences, phrases, tone and context. Examples include online customer service chatbots that analyse typed questions and select the best prewritten response or generate specific, human-sounding answers. Google Translate also uses NLP to study large quantities of text and translations, along with corrections from users, to continually improve its automatic translations.
Neural Network
Also known as an artificial neural network, this is a type of machine loosely inspired by the structure of the human brain. A neural network is composed of simple processing nodes, or ‘artificial neurons’, which are connected to one another in layers. Each node will receive data from several nodes above it and provide data to several nodes below it. Nodes attach a weight to the data they receive and attribute a value to that data. If the data does not pass a certain threshold, it is not passed on. The weights and thresholds of the nodes are adjusted when the algorithm is trained, until similar data inputs result in consistent outputs.
O
Open Source
Open source software and data are freely available for anyone to access, use, modify and distribute. The source code of open source software is publicly accessible, allowing developers to study, improve and adapt it to their needs, promoting collaboration, innovation and transparency in the tech world. Open source fosters a community-driven development model, often resulting in more secure and stable software due to widespread peer review. It also reduces costs for businesses and individuals, supports education by providing learning resources and encourages sharing of knowledge.
P
Predictive Analysis
A type of advanced analysis that uses previous data along with statistical methods and machine learning to predict future events and trends. By looking at past data, these models can find patterns and relationships to forecast things like customer behaviour, market trends and risks. This approach is used in many industries to improve decision-making and planning. For example, businesses use predictive analytics to improve marketing, spot fraud, increase efficiency and check credit risk. Predictive analytics can also be widely applied in public services, including healthcare, infrastructure maintenance and crime prevention.
Prompt
An input, usually text, entered into a generative AI (GenAI) tool, which provides the necessary instructions and context for the AI to generate the desired output, whether in the form of text, image or other media.
R
Reactive Machines
Reactive Machines are the simplest type of AI capable of responding only to current data and not able to learn from past experiences. Famous examples of Reactive Machines include Netflix’s recommendation engine and IBM’s Deep Blue, the chess-playing computer.
Recommendation System
Also known as recommender systems, recommendation systems are used by computer programs to suggest content by predicting what someone will like based on their previous preferences or ratings. Examples include playlist suggestions on platforms like YouTube, Spotify and Netflix and shopping suggestions on Amazon and similar online marketplaces.
Recommendation systems make suggestions by drawing on the characteristics of the content (such as music style or film genre) and analysing what other people with similar tastes and online behaviours have liked or bought. Recommendation systems aim to improve the success of their recommendations over time by building a more complete picture of a person’s preferences the more they use a platform and learning from which recommendations have been successful or not.
Reinforcement Learning
An AI training method where a model performs a task and is given positive or negative feedback by a human. The feedback teaches the model what counts as desirable or undesirable performance. Reinforcement learning is often used in selfdriving cars, which are trained by developers who give feedback on what decisions are good (for example, stopping at a red light) and bad (for example, going over the speed limit).
Responsible AI
The practice of designing, developing and deploying AI with key values such as trustworthiness, ethics, transparency, explainability, fairness, robustness and respect for privacy rights. The Organisation for Economic Co-operation and Development (OECD) identifies five values-based principles for the responsible stewardship of trustworthy AI:
1. AI should benefit people and the planet by driving inclusive growth, sustainable development and well-being.
2. AI systems should be designed in a way that respects the rule of law, human rights, democratic values and diversity, and they should include appropriate safeguards – for example, enabling human intervention where necessary – to ensure a fair and just society.
3. There should be transparency and responsible disclosure around AI systems to ensure that people understand AI-based outcomes and can challenge them.
4. AI systems must function in a robust, secure and safe way throughout their life cycles and potential risks should be continually assessed and managed.
5. Organisations and individuals developing, deploying or operating AI systems should be held accountable for their proper functioning in line with the above principles.
Robotic Process Automation (RPA)
A form of business process automation that uses software robots, or bots, to perform repetitive, rule-based tasks typically carried out by humans interacting with digital systems. While often confused with AI, RPA is process driven and follows predefined rules, whereas AI is data driven and can learn and adapt. Advanced RPA systems can be combined with AI to handle more complex tasks, sometimes referred to as intelligent automation or hyperautomation.
RPA offers several benefits, including cost savings, increased productivity, improved compliance and higher employee satisfaction by eliminating tedious tasks. However, it is important to note that RPA has limitations, such as the need for ongoing maintenance to adapt to system changes.
S
Self-Aware AI
Self-Aware AI is a theoretical form in AI that would have its own consciousness, and understand human emotions and thoughts.
Sentiment Analysis
Sentiment analysis, or opinion mining, is a natural language processing (NLP) technique used to analyse digital text to determine whether the emotional tone is positive, negative or neutral. Companies have access to large amounts of text data such as emails, customer support chats, social media comments and reviews, which can be scanned with sentiment analysis tools to automatically determine the author’s attitude towards a topic. Sentiment analysis can therefore be used to improve customer service and enhance brand reputation.
While useful for gauging public opinion, sentiment analysis has significant limitations, including difficulty in recognising nuance, sarcasm and cultural context. These flaws can lead to misinterpretation of complex sentiments and oversimplification of human emotions, potentially resulting in skewed or incomplete insights.
Speech Recognition
The ability to learn and produce speech, particularly human speech. Also called speech to text, it involves converting sound into text. This technology is used in call centres and voice dictation software that turns voice notes into written text; it is also used in voice assistants like Amazon’s Alexa. See also natural language processing (NLP).
Supervised Learning
In supervised learning, the machine or program is trained using data organised by a human programmer who is supervising the learning process. For example, to train a program to identify whether a picture is of a cat or a dog, the programmer would use many pictures previously labelled, or annotated, as ‘cat’ or ‘dog’. An algorithm identifies key patterns in these images. After this training, the program is given new, unlabelled pictures. It uses the patterns it identified from the labelled pictures to categorise these new images, receiving positive feedback when it correctly identifies cats and dogs.
Synthetic Data
Data generated artificially instead of from real-world events. Synthetic data is useful for research in areas like healthcare and finance where privacy is crucial. Synthetic data can also augment a dataset with more data points, helping AI systems learn desirable properties or training algorithms in dangerous situations, such as teaching a selfdriving car to handle pedestrians on the road.
T
Theory of Mind
Theory of Mind AI is a theoretical form of Artificial General Intelligence (AGI) that would be capable of understanding human thoughts and emotions. This would allow for deeper human and AI interactions.
Training Data
The information used to teach AI systems how to perform tasks or make predictions. Training data includes numerous examples that show the learning system what inputs look like and what the correct outputs should be. By analysing these examples, the system learns to recognise patterns and relationships, which it then uses to build a model, or set of rules. This model enables the system to make informed decisions or predictions when encountering new data. The quality and quantity of training data play a crucial role in determining how well the AI system performs its tasks.
Turing Test
A method to determine whether a machine can demonstrate human intelligence. If a machine can engage in a conversation with a human without being recognised as a machine, it passes the test. The Turing test was proposed in a 1950 paper by mathematician and computing pioneer Alan Turing. It has become a key concept in AI theory and development.
U
Unsupervised Learning
A method in machine learning where an algorithm learns patterns exclusively from unlabelled data. (See supervised learning, by way of comparison.) In this approach, a machine learning model independently identifies structure, similarities and differences in the data. Common uses of unsupervised learning include recommendation systems, customer segmentation, fraud detection and genetic research.
V
Visual Perception
The ability to recognise and identify objects in an image, similar to how humans perceive and understand what we see. This type of AI is used in medical image processing, facial recognition software (for policing or crowd counting, for example) and self-driving vehicles, among others.
Voice Cloning
Voice cloning uses AI to create a digital version of someone’s unique voice. It captures speech patterns, accent, voice inflections and even breathing. From a short audio clip, sometimes as brief as 3 seconds, an algorithm can learn and accurately replicate a person’s voice. Voice cloning can be used in a variety of applications, including immersive learning experiences, helping people with alternative speech patterns and preserving endangered languages. It can also be used to create audio deepfakes for malicious purposes.