We thought it might be a good idea to add a Glossary as this field is replete with abbreviations and terms that sound similar. If you want to learn more, the title links typically take you to the relevant article on this blog or an article on Wikipedia.
We’d also like to acknowledge the amazing Glossary of GenAI Terms created by staff at The University of British Columbia. We used this to check we hadn’t missed out any key terms, and also to improve some of our definitions.
Agents or Agentic AI
AI tools that can make decisions and perform tasks with limited or no human intervention. A commonly cited example is a tool that could book you a holiday given a destination, it will reserve and pay for flights, hotels, select suitable activities, etc. and add all the details to your calendar, setting your out of office (and applying for time off via the staff absence system). That all sounds great as long as it doesn’t go wrong.
AI Design Assistant
The name given by Anthology to the suite of AI-driven tools that they have added to the Blackboard Learn VLE. At present these are largely staff-facing.
AI Detection
This are tools that have been designed to analyse content (usually text) to determine whether it was written by a person or generated via AI. These tools apply machine learning techniques and look for patterns and inconsistencies. Many derive measures such as perplexity (the degree to which the writing follows standard patterns) and burstiness (a measure of sentence length and structure) in an attempt to spot the use of generative AI. Common examples of detection tools or tools that include a detection component are Copyleaks, GPTZero, Grammarly, Turnitin, Quillbot and OpenAI’s now withdrawn AI Text Classifier.
There is a technology “arms race” between the people developing generative AI tools and those creating detectors. The people creating the tools have larger budgets and the people writing the detectors are always playing catch-up. For this reason many people have defined this as a war that the detectors can never win.
There are significant concerns about the rate of false positives and false negatives that many detectors give, and risks of bias. Detectors may be more likely to be flag work from people writing in their second language than native speakers.
Unlike similarity detectors which compare students writing against a database of published material, there is no equivalent source for AI detectors. They convert the student’s work into numbers and base their decisions on these values.
As such we don’t recommend their use.
Annotation
Annotation is a way of labelling data, so it can be used to train, or later fine-tune models. For models that process images, these annotations might identify objects at a high level (such as “dog” or “cat”), or perhaps in more detail (“Alsatian” or “Chihuahua”) or colours, emotions, image styles, etc.
Anthropomorphism
Literally the act of assigning human-like characteristics or behaviours to an object. Although tools like chatbots may have been designed to appear human-like (starting with cheery greetings, and responding to prompts in a conversational style) the underlying models don’t have feelings or conscious understanding. The dangers of such anthropomorphism is that people start treating tools like Chatbots as humans and developing relationships with them. Generative AI should never exclude interaction with other human beings.
Artificial General Intelligence (AGI)
Sometimes referred to as Strong AI, this is an as yet theoretical type of AI that would match human capabilities across many cognitive tasks. Creating an AGI is the declared aim of some current generative AI companies. They see this as the next logical step, building on the current AI systems that are designed to perform specific tasks (sometimes referred to as Narrow AI).
Artificial Intelligence (AI)
The capability of computational systems to perform tasks that would normally be carried out by people. Tasks such as learning, reasoning, problem-solving, perception, and decision-making. This is often illustrated with the rather mundane example – “Here is what is in my fridge. Can you give me some meal suggestions?”
Artificial Super Intelligence (ASI)
An as yet theoretical type of AI that would surpass human capabilities across virtually all cognitive tasks.
Attention
This is essentially a mathematical way for models to decide what is important. In a language model, attention is used to help differentiating between potential different meanings of the same words, depending on their position in the sentence and the words around them.
For example compare the different meaning of the word “sentence” in the previous paragraph, where it refers to a collection of words, with this example, where it refers to a legal outcome in a courtroom:
“On Friday as the Sycamore Gap trial came to an end, the judge issued their sentence.”
Attention also helps to correctly process more nuanced language, such as detecting sarcasm or attempts at humour.
The transformer architecture in models such as GPT vastly improved ways to detect, process and optimise attention, leading to much quicker and better results.
Backpropagation
This is the algorithm used during the training phase of a neural network. It attempts to measure the error in the model’s predictions (the current output) and adjusts the parameters accordingly.
Bias
Somewhat confusingly, the word ‘bias’ is used in two ways when talking about generative AI.
Within a model, bias is used to refer to one of the parameters that can be used to tune the model output, to help it model complex data correctly. (The bias is combined with model weights). In this context, bias is a positive thing, as it improves the accuracy of the output.
Conversely, generative AI models are often accused of cultural bias in their output, i.e. that it is favours one particular view of the world, reflecting the biases that were present in the data it was trained with. For example id you entered a prompt asking for photos of doctors, early image generation tools often returned pictures of white men. Similarly if you asked for photos of people from Mexico, the results almost exclusively showed pictures of men wearing sombreros. These reflect particular cultural stereotypes, rather than the actual appearance of people in Mexico today.
Chatbot
A programme or part of a web page that uses text or spoken words to interact with people. ChatGPT is an example of a generative AI chatbot. They use natural language processing techniques to improve the relevance of their responses.
ChatGPT
A generative AI chatbot that was first released by OpenAI in November 2022 and accessed using a web browser. This release is generally acknowledged as the moment when easy to use generative AI tools first became widely available to the general public. The initial focus was generating text, later versions have added the ability to generate audio and images as well.
Cognitive offloading
The risk that if you use tools such as generative AI to undertake key parts of the learning process – e.g. to summarise articles that you never read – then you miss out on some of the activity that is required to deeply understand a concept.
Context
This word is used to refer to the information passed back and forth between you and the generative AI tool (typically a Chatbot). At first it contains your prompt, any attached files and possibly a meta prompt (some hidden instructions created by the tool designers to try and get better results, e.g. “Play the role of a knowledgable teacher”). The returned result is added to the context. If you reply again to ask for refinement or changes, this second prompt is added to the context and the whole conversation is passed back. This helps provide continuity.
Models will have a Context window – the maximum amount of information it can process when generating a response. This is usually measured in tokens. Some people think of this as the model’s working memory. When the context window limit has been reached, a new chat (with a new context) is required.
Copilot
Microsoft Copilot is a generative AI assistant that is built on top of OpenAI’s GPT-4 and GPT-5 models. It is available via a web browser, through the taskbar in the latest versions of the Windows operating system and also integrated into apps such as the Microsoft365 versions of Excel, Outlook, Word, Powerpoint, etc.
Deep Learning
Not to be confused with the concept of deep (as opposed to surface learning) from education, in machine learning, the term deep learning is usually applied to neural networks that contain many layers of artificial neurons (sometimes in the thousands) that allow complex processing of inputs. The depth term reflects the fact that each layer converts the input from the previous one into a progressively more abstract and composite forms. This depth is what allows image generation tools to recognise features such as eyes and noses, by building them up from simpler shapes (such as circles and edges) identified in earlier layers of the model. The concept of a face may them be developed later, as a circular shape that contains one nose object and two eye objects. The relationship between these objects may indicate the orientation of the head in a later layer. The depth of layers allows much more advanced analysis.
Embedding
Embedding is a key part of the conversion of text into mathematics, that generative AI models require to enable them to make predictions. Words, phrases and sometimes even documents are plotted in a multi-dimensional space. Effectively the model creates a learned representation space, so that words with similar meanings plot closely together and opposites are opposed along certain axes. This also allows the identification of relationship such as gendered terms – King vs Queen, Prince vs Princess, etc. – as the difference between the position of the two points (King vs Queen), (prince vs Princess) should be a similar vector.
This multi-dimensional space is created using algorithms that analyse usage patterns, context and relationships between words.
Encoder-Decoder Architecture
An approach that involves two neural networks, common in image processing and speech synthesis. The encoder process the input data, compressing it but attempting to maintain the core characteristics. The second neural network – the decoder, attempts to process this compressed version and recreate something as close to the original input as possible.
Ethical AI
An attempt to develop and use AI systems by prioritising ethical positions. Typically these focus on five areas: transparency, justice, fairness, non-maleficence (don’t do bad things), responsibility and privacy.
Few-shot learning
A term applied to machine learning models that can be trained and make accurate predictions from a relatively small amount of data. This technique can be useful in applications where there is not many training data available – e.g. from ancient languages. An example would be asking an AI model to provide some reviews using a specific template it did not come across during the training. If you provide a couple of examples of completed reviews, it could use this to create a new review, matching the structure and writing style (e.g. formal or conversational).
Explainable AI
This is the name for the branch of artificial intelligence that is focussed on trying to provide people with oversight over AI models, exposing the reasoning behind the decisions/predictions they make. This transparency is lacking from current models (whose inner workings are often described as “black boxes”). Providing this could address concerns about the use of such models in decision-making contexts, help expose, measure and ultimately reduce bias and give people more trust in the model output.
Fine tuning
This is the name for the process used to create specialist models (e.g. an AI tool for lawyers or to help radiologists analyse x-ray images). It involves taking an existing pre-trained model (e.g. ChatGPT) which learnt from a very broad dataset. This model is then provided with a set of specialist data (e.g. English case law, or x-ray images of both tumours and healthy tissue) and the parameters are adjusted to obtain better results. This should produce a model optimised for the selected subject.
Gemini
Gemini (originally known as Bard) is Google’s generative AI chatbot and virtual assistant. The underlying model is also called Gemini.
Generative AI (genAI)
Artificial intelligence tools/systems that create new content in response to a prompt. This content can be text, poems, images, video, computer code, etc. The models have been previously trained on large sets of data. Common examples are tools like ChatGPT, Gemini, DALL-E, Midjourney.
Generative Pre-trained Transformers (GPT)
This is a type of large language model that uses the transformer architecture to create a model from an enormous set of training data. The resulting pre-trained model is widely used in chatbots and have been surprisingly effective in a wide range of applications. Although originally developed to process text, GPT models have now been applied to process images and audio. The most well know models are the various flavours of GPT that are used in OpenAI’s ChatGPT.
Hallucination
A term used to label errors or mistakes in the model output. These could be fabrications – such as citing published works that were never written, a result of poor coding, or or just collections of words that don’t make sense. The term has been criticised for humanising these failures (anthropomorphism) in an attempt to minimise their significance.
Knowledge Cutoff
This refers to the date after which new information was not added to the model’s training data. A classic example of this was asking early ChatGPT models the first name of the current UK monarch. Although at the time the correct answer was Charles, the answer deemed most likely from the training data was Elizabeth.
Large Language Model (LLM)
Artificial intelligence systems that are designed to process, create and respond to human languages. They are typically trained on enormous data sets, in an attempt to grasp the complexity and nuance of a language in their mathematical models.
Local Models
Most generative AI models run in the cloud – i.e. in large data centres with racks of servers. This provides rapid responses and ensures they can process large numbers of requests at once.
Local models are an alternative approach, using software that takes the same model (i.e. the same parameters, with their weights and biases) and runs it locally on your computer (if it has sufficient power). Standard models can be tuned for running locally (e.g. reducing some of the least significant parameters) to make them run faster or on lower hardware specifications.
Using a local model gives users more control and offers the potential to analyse sensitive data because the processing can be run offline and occurs on your device. The trade off is lower performance and less access to advanced third party features – such as image generation.
Machine Learning (ML)
This is the name given to the part of artificial intelligence research that looks at ways that algorithms can learn from data and then generalise these relationships to new, previously unseen data, and so perform tasks without specific instructions. Natural language processing is an example of a machine learning techniques.
Model
The term model is applied to the algorithms and computational structures (code) that allows a generative AI tool to process data and generate output. Models comprise a series of parameters, which are configured during the training process to set the relevant weight and bias values.
Models differ in their complexity, in the number of parameters they use, the data and the target outputs used in their training.
Natural Language Processing (NLP)
A set of approaches that draw from artificial intelligence and computational linguistics. NLP is designed to help machines to correctly process instructions that are provided in “every day language” and generate outputs that are also easy for people to read.
Neural Network
Inspired by the human brain, but actually a mathematical system, a neural network is a computer system made up of layers. Each layer process inputs, passing on outputs to the next one. Each layer contains many nodes referred to as artificial neurons – basically algorithms that have weights and biases that determine how to respond. In combination and through repeated training, these networks can identify statistical patterns in data (such as eyes in photographs, or concepts such as gendered words like ‘King’ and ‘Queen’) which can be used within generative AI models.
Open Weight Models
A type of generative AI where the parameters to the underlying model have been released to the public. You are able to download these models and run them as local models on your own computer (depending upon its processor and memory). You could also fine tune the model, to better suit your own applications.
Parameters
The core internal variables of a model that define it and determine how it processes information to create output. These parameters are defined and refined during the training process. In a neural network, the parameters include both weights and biases. The number of such parameters is one way to compare models.
Prompt
The input provided to a model. The text you type into the chatbot. This acts as instructions for the generative AI model (and may be combined with a hidden meta prompt crafted by the tool authors). This will become part of the Context.
For a long time there has been a lot of discussion and posts about different approaches to prompt crafting (sometimes over-hyped as prompt engineering, but no sums are involved folks). As models and natural language processing improve, the importance of the exact wording used in your prompt may well decline.
Reinforcement learning
A process used to train models where it receives feedback to reward good answers. This approach is designed to encourage the correct behaviour and reduce the chance of unwanted outputs.
Retrieval Augmented Generation (RAG)
Often abbreviated to RAG, this approach combines the strengths of generative AI with more traditional “2020 Google Search” type retrieval-based models. This allows the model to obtain information from existing datasets (or knowledge bases) and then use this information to provide extra context with which the model generates a response. When used effectively, this method should produce better results, complete with attributions.
Singularity
The name for the hypothetical point in time, where Artificial Super Intelligence is achieved and the technology outperforms humanity. It gained prominence in Ray Kurzweil’s 2005 book The Singularity is Near (possibly because he predicted AI would match human intelligence in 2029 and that the singularity would occur in 2045) when human and machine intelligence would merge.
Supervised Learning
Learning that uses annotations (or other forms of labelled data) is known as supervised learning. These are used to guide the model (essentially by supplying the right answers).
Temperature
This is a numerical parameter in a model that controls randomness and is usually assigned a value between 0.0 and 1.0. Adjusting the temperature is often used as a way to provide more creative (less likely, but still valid) responses. The higher the model temperature, the greater the degree of variability and use of less-likely terms in the output, the lower the value the more deterministic it is. This can be used to great effect when creating stories or poems for example.
Token
The smallest unit of data that AI models process. There are different ways to spilt up words (tokenisation), so for some models tokens may closely represent whole words, in others they may look more like syllables, or letters that indicate a plural (such as “s” in English). Interactions may be limited by the number of tokens that the model is able to process.
Training
This is the name for the process where a mathematical model (often involving a neural network) learns to perform a certain task – generate a valid output, given the input. During training, models are typically provided with an enormous amount of data (the training dataset) and this is used to adjust the model’s parameters, which influence the output. The output is compared with the desired output (either mathematically, or also involving human input – sometimes given the rather grand title of Reinforcement Learning from Human Feedback – RLHF). The parameters are then adjusted and the input re-run until a suitably high quality output is obtained. This can involve very large numbers of cycles.
Transformer
Transformers are one of the most common models currently used in generative AI tools. Unlike past models that processed data sequentially, transformers work much faster and better, because they process the entire sequence of data simultaneously (your input, then the output). They use self-attention to detect complex relationships and generate better output. The most well known transformer model is GPT (as in ChatGPT), where the letter T stands for Transformer.
Trojan Horse
A term sometimes applied to attempts by staff to detect student use of generative AI in assessments where this is prohibited, by adding a hidden prompt to the materials. For example this may be written in small white text on a white background, or concealed in alt text, it is designed to be invisible to most readers. If a student uses a generative AI tool to read the document, the model will detect the hidden prompt and use it to help shape the output.
Well-publicised examples include requesting the output is written from a Marxist perspective (when this is not relevant or mentioned in the “visible” instructions), to refer to solutions involving mermaids and unicorns, or to incorporate words such as “Frankenstein” and “banana” in the answer. In the US it is sometimes called adding an “AI banana”, for this reason.
This approach is not recommended. It is entrapment and risks souring the crucial bond of trust between staff and students. It also will affect some students more. For example, the “hidden” message will be exposed to students using screen-readers and other assistive technologies, which may cause confusion and lead learners down the wrong path, wasting their time.
Vocabulary
This is the name for the set of unique tokens (words, parts of words and characters) that a model recognises from it’s training and so can use when predicting a response. Modern generative AI models have a large vocabulary (over 50,000 tokens).
Weights
Weights determine how a neural network processes input data and so generates outputs. Effectively they define the strength of connections between the artificial neurons in a neural network. Weights can be thought of as knobs that control the relative importance of a parameter, how much influence it has on the final output.
Zero Data Retention
The deliberate design to data processing by a model, where the input data is not stored after it has served it’s purpose. For many, this is a key part of meeting privacy, ethical and data security requirements.
Zero-shot learning
Describes a model that is able to learn to perform tasks that it was not previously trained to do. It attempts to apply existing relationships and apply them to the new context. Wikipedia provides an example of an AI model that had been trained to recognise horses being shown a picture of an animal it had never seen before (a zebra) and asked to classify it. If the model has access to other sources of information – notably that zebras look like striped horses – it can apply this and so infer that the animal image it has been given is probably a zebra.