Machine learning, natural language processing, and computer vision are just some of the many systems that make AI work. Knowing what these systems are and how they work is essential to critically thinking about what AI can and cannot do.
How Does AI Actually Work?
The Systems That Make AI Work
Machine Learning (ML) — Learning From Examples
Machine learning is the foundation of most modern AI. Instead of being programmed with a set of rules, a machine learning system is trained on large amounts of examples. It identifies patterns in those examples and uses them to make predictions or decisions about new data.
Supervised learning: A type of machine learning where the system is trained on data that already has the correct answers attached. A human labels the examples ("this email is spam," "this application is eligible"), and the system learns to recognize the patterns that led to each label.
Unsupervised learning: Works on data that has no labels attached. The system explores the data on its own, grouping similar items and flagging unusual patterns without being told what to look for. It's useful for discovering things you didn't know to ask about.
Reinforcement learning: Teaches AI through trial and error. The system tries different approaches, receives a reward when things go well and a penalty when they don't, and gradually learns the best strategy. It's similar to the way animals and people learn through experience.
Deep learning: A more advanced form of machine learning that uses neural networks — systems loosely inspired by the structure of the human brain — with many layers. These layers process data in increasingly abstract ways, making deep learning particularly good at tasks like recognizing speech, images, and language.
You don't need to be an engineer to make decisions about AI. But understanding the basics of how different systems are built helps you ask better questions and spot when someone is oversimplifying or overclaiming.
Machine learning and AI are increasingly used as tools for Indigenous language revitalization, helping to document, translate, and teach languages with few remaining speakers.
- Automatic speech recognition (ASR): AI models, such as those developed by Te Hiku Media, are trained on audio recordings to transcribe spoken language into text, achieving high accuracy.
- Predictive text and keyboards: Open-source platforms like Keyman are used to add Indigenous language predictive text to smartphone keyboards.
- Language models: Small language models are fine-tuned on limited data to create spell-checkers, next-word predictors, and chatbots.
- Educational tools: Projects like use AI to create "Language in a Box" systems for home practice, while SkoBots (interactive robots) are used in classrooms to make learning interactive.
Key challenges and considerations
- Data sovereignty: A critical concern is ensuring communities maintain ownership over their language data and digital resources.
- Data scarcity: Most Indigenous languages are considered "low-resource" in the context of AI, meaning there is not enough text or audio data to train robust, traditional models.
- Quality of output: AI models, especially when trained on limited or low-quality data, can produce inaccurate text or "gibberish," which may negatively impact efforts.
- Cultural sensitivity: AI may fail to understand cultural nuances, such as when specific stories should be told, necessitating community oversight to avoid misusing sensitive knowledge.
Natural Language Processing (NLP) — Reading and Writing Like a Human
Natural Language Processing (NLP) enables computers to work with human language: reading, summarising, translating, searching, and generating text. It's what powers chatbots, search engines, translation tools, and writing assistants. NLP is now embedded in tools most of us use every day.
Computer Vision — Seeing the Land
Computer vision enables computers to interpret and understand images and video by identifying objects, detecting changes, and recognizing patterns. It's what lets your phone unlock with your face and lets cars detect pedestrians. It can also be a powerful tool for land and environmental stewardship.
Expert Systems — Rules Set by People
Expert systems are different from the machine learning approaches above. They don't learn from data; they follow rules written by human experts. Think of them as a digital decision tree: if this, then that. They're predictable, transparent, and easy to audit.
Ask Questions About the AI Tools You Use
Every time you use an AI tool, you are making decisions about data: what goes in, where it's stored, who can see it, and how it might be used. These decisions connect directly to data sovereignty.
Here are five questions to ask before using any AI tool:
Where is my data stored: on my device, in North America, or overseas?
Does the tool use my inputs to train or improve its AI models?
Who (besides me) can access what I've entered: staff at the company, third parties, government agencies?
What happens to my data if I stop using the tool or close my account?
Is this tool appropriate for sensitive community, health, or cultural information?
The answers to these questions should be clearly explained in the company’s privacy policy, data use policy, or terms of service. If you cannot find the answer, if the wording is vague, or if the company says it may use your inputs to train its systems, treat that as a red flag. This means you should pause, avoid entering sensitive or community-related information. When in doubt, do not upload the information.