A plain-language guide for healthcare and accreditation
Artificial Intelligence (AI) is a set of tools that learn patterns from data. In practice, that means computer systems doing things we usually think of as needing human smarts: spotting patterns, understanding language, making predictions, and creating new text or images. AI is not one single technology. It’s a family of methods built over many years. What’s new lately is scale—huge amounts of data, very powerful computers, and systems that can work in plain language.
At its heart, AI is about learning from data. Instead of a person writing every rule by hand, the system learns patterns from examples. That’s why the way we build and use AI matters so much: what data it learns from, who controls that data, and how we check that it’s safe and fair.
Early AI was mostly rule-based. People wrote clear instructions: “If this happens, do that.” For example: if a patient has certain symptoms, suggest a particular protocol. These systems worked for narrow tasks but were limited. Experts had to think of every situation in advance.
Machine learning (ML) changed that. Machine learning finds patterns in data and uses them to make predictions. Instead of writing every rule, we give the computer lots of data, the right answers for that data, and an algorithm that finds patterns. The system “learns” links between inputs and outputs. For instance: show it thousands of medical images labeled “benign” or “malignant,” and it can learn to predict the likelihood of malignancy on new images. The model doesn’t “understand” cancer like a doctor. It uses patterns in the data to estimate probabilities. That’s called supervised learning. Other types include unsupervised learning (finding patterns without pre-labeled answers) and reinforcement learning (improving from feedback over time).
For a technical but readable yearly overview of AI progress, see Stanford’s AI Index.
From around 2012, a branch of machine learning called deep learning started to do better than older methods in things like recognizing images and speech. Deep learning scales the pattern-finding approach using layered networks (artificial neural networks): math structures loosely inspired by how brain cells connect. Each layer turns the input into more abstract information—for example, in an image model, early layers might detect edges, middle layers shapes, and later layers whole objects.
Deep learning needs a lot of data and a lot of computing power. That power became cheaper partly because of graphics processing units (GPUs)—chips first built for video games and now used for training big models. That shift is one big reason AI has accelerated in the last decade.
Natural Language Processing (NLP) is what lets computers work with human language. Older systems used a lot of hand-written rules and smaller datasets. They could do things like basic chatbots or keyword extraction but struggled with nuance and long, context-heavy text.
The big change came with transformer models in 2017. The paper “Attention Is All You Need” introduced an architecture that could process language at scale by focusing on how words relate to each other in context. That design became the basis for today’s large language models (LLMs).
Large Language Models (LLMs) are deep learning systems trained on enormous amounts of text—books, articles, websites, and more. They learn patterns in language and use those patterns to generate text: they predict the next word in a sentence given the words that came before. That sounds small, but at huge scale it becomes very flexible. When you ask a question, the model generates an answer one word at a time, choosing what’s most likely to come next based on patterns it learned.
LLMs don’t “know” facts the way people do. They produce text based on patterns in the data they were trained on. They can be wrong or sound right when they’re not. So they are best used as tools to support people, not as final decision-makers. For more on how models like GPT work, see OpenAI’s research overview. Other well-known LLMs include Google’s Gemini, Anthropic’s Claude, and Meta’s Llama.
Because they work in plain language, LLMs can summarize documents, draft policies, translate, write code, or answer questions. That makes them useful in healthcare for tasks like drafting summaries or checklists—as long as a qualified person checks the output before it’s used in records or decisions.
Generative AI means AI that creates new content: text, images, audio, video, or code. It’s built on the same deep learning ideas but is focused on producing new things rather than only classifying or predicting. So where a traditional model might only label or score something, generative AI produces novel text, images, or other outputs. For example, a traditional model might predict a patient’s readmission risk; a generative model might draft a discharge summary. Generative AI is powerful because it can help with knowledge work—writing, summarizing, outlining—which is why it’s relevant to accreditation, quality improvement, and documentation.
Several things came together: massive data (the internet and digitized records), cheaper, stronger computers (cloud and specialized chips), better algorithms (transformers and related ideas), and big commercial investment. On top of that, chat-style interfaces made advanced models easy to use without coding—so what’s new is both scale and ease of use. The result is fast adoption in healthcare, education, government, and more. The important questions are no longer “Does it work?”—many uses do—but who governs it, who benefits, who bears the risk, and how it aligns with your quality and safety standards.
Training large AI models needs huge computing resources: data centers with tens of thousands of GPUs, heavy cooling, and fast networks. These facilities use a lot of electricity and water. As AI grows, so does that footprint. The International Energy Agency tracks data center energy use. This affects the environment, regional capacity, and questions about sovereignty and data residency—where data is stored and processed matters a lot for health and Indigenous data governance (e.g. OCAP® and First Nations control).
Bias and fairness: Models learn from past data. If that data reflects unfairness or gaps (e.g. underrepresentation of Indigenous peoples), the model can repeat or worsen those patterns.
Transparency: Deep learning is often called a “black box”—it can be hard to explain exactly why a model gave a certain answer. That’s a challenge for accountability and for accreditation when you need to show how decisions are made.
Data sovereignty: Who owns and controls the data used to train AI? Were communities asked? In First Nations contexts, OCAP® and the First Nations Information Governance Centre set out principles for ownership, control, access, and possession.
Safety and reliability: LLMs can give confident but wrong answers—“hallucination.” In health, every output should be checked by a qualified person before it’s used.
Privacy: Training data can include sensitive information. In healthcare, privacy and regulatory rules must be respected; sending patient or community data to third-party AI without clear agreements can break those rules.
Market concentration and workforce: A small number of companies run much of the world’s big AI infrastructure. That raises dependency and governance questions. AI also changes how we work—it can support people but also shifts what skills are needed.
Many governments use the OECD AI Principles as a reference for responsible AI.
LLMs predict text from patterns. They don’t have lived experience, don’t look up facts in real time by default, and can produce plausible but incorrect text. They are tools to support people, not replacements for professional judgment. In regulated areas like healthcare and accreditation, their output must be reviewed and validated by qualified staff before it becomes part of the record or a decision.
AI is already part of clinical documentation, risk scoring, scheduling, quality dashboards, and population health analytics. Accreditation is about quality, safety, and continuous improvement. So AI tools need to align with your governance, documentation standards, bias mitigation, and data security. Adopting AI is not only a technology choice—it’s a systems and governance issue. The key questions are who governs AI, who benefits, who bears risk, and how it aligns with your quality and safety standards. Document how you use AI, how you check it, and how you keep data and communities at the centre.