Artificial Intelligence (AI) and the Church: A Primer

View a printable PDF transcript of the video lessons

View a printable PDF of the terms glossary

1. Agent (AI Agent; Agentic AI) – A software or AI system that acts on behalf of a user or program, perceives its environment, processes information, takes actions, may learn from outcomes and adapts—essentially an autonomous (or semi-autonomous) actor in a computational system.
    
2. Alignment (AI Alignment) – The discipline and set of practices concerned with ensuring that an AI system’s goals, behaviour, and outcomes match the intentions, values and norms of the humans who deploy it—that the machine does what we want it to do, and does not deviate.
    
3. Algorithm – A clear and precise sequence of instructions or rules that a computer follows to perform a task, process data, or arrive at a decision; much like a recipe or workflow, it takes inputs, applies the prescribed steps, and produces outputs.
    
4. Artificial General Intelligence (AGI) – A hypothetical future form of AI that would possess broad human-level cognitive ability across virtually any intellectual task, not just specialised functions; it remains a theoretical concept rather than current reality.
    
5. Artificial Intelligence (AI) – The field and technology of creating computer systems or machines that can perform tasks which normally require human intelligence—such as learning from experience, reasoning, recognising patterns, adapting to new situations, or making decisions.
    
6. Attention Mechanism – A component within modern neural networks (particularly transformer models) that allows the model to dynamically focus on different parts of the input (for example which words matter most for understanding a sentence) when generating output, rather than treating all parts equally.
    
7. Automated Decision-Making – A process in which one or more decisions are made by an AI system or software with little or no human intervention; it raises issues of oversight, accountability and trust especially when decisions affect people’s rights or welfare.
    
8. Bias (in AI) – Systematic distortions or unfairness in the outputs of an AI system that arise because of unrepresentative or flawed training data, flawed assumptions in algorithm design, or unintended effects in model behaviour; bias can lead to certain groups being disadvantaged. 
    
9. Compute – The computational resources (such as processing power, memory, specialised hardware, energy) required to train, operate or deploy AI models; more advanced models often demand more compute, which affects cost, energy utilization, speed and feasibility.
    
10. Context Window – In language models, the maximum amount of input (text or tokens) the model can consider at one time when processing or generating output; it sets a limit to how much “memory” the model has of the prompt or preceding conversation. 
     
11. Data Augmentation – Techniques used during AI model training to increase the effective size and diversity of the training dataset by creating modified versions of existing data (for example slightly altered or transformed), thereby improving model robustness and generalisation. 
     
12. Deep Learning – A type of machine learning that uses structures called neural networks with many (“deep”) layers of processing units to detect very complex patterns in data (for example in images or speech), enabling more advanced recognition or generation capabilities. 
     
13. Embedding – A way of representing items (words, sentences, images) as vectors (lists of numbers) in a continuous high-dimensional space such that similar items are close together; these embeddings allow a model to reason about similarity, meaning or relationships in data.
     
14. Ethical AI – The field and practice of designing, deploying and using AI in ways that respect fairness, privacy, accountability, transparency, human dignity and societal wellbeing; it includes principles, policies and frameworks to ensure the technology serves good ends and avoids harm. 
     
15. Explainability (XAI) – The quality or practice of making AI systems’ decisions, predictions or internal workings transparent and comprehensible to humans, so that users, stakeholders or regulators can understand why a system produced a particular output and build trust in its use.
     
16. Fine-Tuning – The process of taking an AI model that has already been trained on broad data and then training it further on more specific or domain-relevant data so the model becomes better-suited for a narrower task or context. 
     
17. Foundation Model – A large-scale AI model trained on very broad, general-purpose data (for example large text corpora or combinations of modalities) and then adapted (via fine-tuning or prompting) to many downstream, specific tasks; it serves as a base layer for multiple applications. 
     
18. Generative Adversarial Network (GAN) – A special kind of deep learning architecture in which two neural networks compete: one network (the generator) tries to create realistic synthetic data, while the other network (the discriminator) tries to distinguish generated data from real data—through this competition the system improves the quality of the generated content. 
     
19. Generative AI (GenAI) – AI systems that do more than analyse or classify—they actually create new content (for example text, images, audio or video) by learning the patterns in existing data and then producing novel outputs that follow those patterns. 
     
20. Hallucination (in AI) – When an AI system generates content that appears plausible and credible but is actually incorrect, misleading or fabricated; the model essentially “makes things up” even though the output looks like an informed answer. 
     
21. Human-in-the-Loop (HITL) – An approach in system design where humans remain actively involved in overseeing, reviewing or intervening in the operation of AI systems—ensuring that critical decisions are not solely automated and that human judgment and moral agency remain central. 
     
22. Large Language Model (LLM) – A kind of AI model trained on vast volumes of text so that it can generate language, answer questions, summarise content, translate, or otherwise understand and produce human-style text by predicting what comes next in sequences. 
     
23. Machine Learning (ML) – A subset of AI in which computers use large amounts of data and statistical techniques to “learn” how to improve at a task over time, without being explicitly programmed for every possible scenario. 
     
24. Model – A trained AI system (or mathematical representation) that has “learned” from data how to perform a given task—once training is done, the model is what is deployed to make predictions, generate outputs, or take actions. 
     
25. Multimodal Model – An AI model designed to handle and integrate more than one type of data or “modality” (such as text and images and/or audio) so it can process input from several channels and/or produce output in several forms. 
     
26. Natural Language Processing (NLP) – The branch of AI concerned with enabling computers to understand, interpret, generate, and respond to human language (spoken or written), bridging the gap between human communication and machine processing. 
     
27. Neural Network – A computational structure inspired by the human brain, composed of interconnected nodes (“neurons”) organised in layers, where each connection has a weight; during training the network adjusts these weights to recognise patterns or relationships in data and produce appropriate outputs. 
     
28. Overfitting – A situation during model training where the model learns the training data—including noise or irrelevant details—so precisely that it fails to generalise well to new, unseen data; essentially it memorises instead of learning the underlying pattern. 
     
29. Parameters – The internal numerical values (weights, biases, etc) within an AI model that are adjusted during training; these values determine how the model transforms input into output, and the total number of parameters is a rough measure of model complexity. 
     
30. Prompt Engineering – The process of carefully designing, refining and controlling the input prompt (text or instructions) given to a language model or generative AI so as to steer its output in desired directions, maximize accuracy and minimise unwanted behaviour. 
     
31. Reinforcement Learning (RL) – A learning paradigm in which an AI “agent” interacts with an environment, takes actions, receives rewards or penalties depending on the outcomes, and uses that feedback over time to learn which actions lead to better long-term results.
     
32. Robustness (in AI) – The capacity of an AI system to maintain reliable, stable and acceptable performance even when faced with varied, unexpected or difficult inputs, changes in environment, noise, or adversarial attempts to degrade it; robustness is vital for safe, real-world deployment. 
     
33. Self-Supervised Learning – A learning method in which the model uses raw, unlabeled data and generates its own supervision or “labels” from that data (for example by hiding part of the input and asking the model to predict it) in order to learn useful representations, without needing extensive human-labeled examples. 
     
34. Semi-Supervised Learning – A hybrid training approach combining both labelled data (with correct outputs) and unlabelled data, enabling the model to leverage the labelled portion for guidance and the unlabelled portion for discovering additional structure when labels are limited or expensive to obtain. 
     
35. Supervised Learning – A training method for machine learning in which the model is provided with labeled input-output pairs (examples where the correct answer is known) and the model learns to map inputs to the correct outputs. 
     
36. Tokenisation – The process of breaking up text (or other sequential input) into smaller units called “tokens” (which might be words, sub-words or characters) so that an AI language model can process the input efficiently, treat each token as a unit, and track sequences of tokens for meaning. 
     
37. Transformer – A powerful neural-network architecture that uses attention mechanisms and processes input data in parallel instead of sequentially, which makes it extremely effective at tasks like natural language understanding, translation, and generation; it is a foundation of many current language and multimodal models. 
     
38. Transfer Learning – A method in machine learning where knowledge acquired while solving one task or using one dataset is reused or transferred to help solve a different but related task—this saves time, data and performance overhead compared to training from scratch. 
     
39. Underfitting – The opposite of overfitting: when a model is too simple or inadequately trained, it fails to capture the underlying pattern in the data and thus performs poorly both on the training data and on new data—it hasn’t learned enough.
     
40. Unsupervised Learning – A training method in which the model is provided with input data that lacks explicit labels or outputs, and the model learns to find patterns, structures or groupings in the data on its own.