AI 101 — Civic AI Canada

What is modern AI?

The term "artificial intelligence" covers an enormous range of technologies, from the spam filter in your email inbox to systems that can hold a conversation, write code, or generate photorealistic images. When people talk about AI today, they usually mean a specific type of system: a large language model, or LLM.

Here is the short version of how an LLM works. Developers collect enormous amounts of text — think hundreds of billions of web pages, books, scientific articles, court decisions, and code repositories. They then train a mathematical model on this data. During training, the model learns statistical patterns: which words tend to follow which other words, in which contexts, in response to which kinds of questions. This process runs on thousands of specialized computer chips called GPUs for weeks or months, at a cost that can reach hundreds of millions of dollars.

The result is a model with hundreds of billions of internal numerical values called parameters — think of them as billions of small dials, each calibrated to a specific setting — that together encode what the model has learned. When you type a question into ChatGPT, the system uses those parameters to predict the most statistically likely next word, then the word after that, then the word after that, until it has produced a response.

The "large" in large language model matters. These models have hundreds of billions of parameters, trained on computing infrastructure that costs enormous amounts to build and run. That scale is what makes modern AI qualitatively different from the AI systems of even ten years ago — and what makes the current moment feel so consequential.

What counts as AI right now?

Current AI systems that most Canadians will encounter include:

Chatbots and writing assistants: ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Copilot (Microsoft)
Image generators: DALL-E, Midjourney, Adobe Firefly
Code assistants: GitHub Copilot, Cursor
Voice tools: Speech-to-text, voice synthesis, AI-powered customer service systems
Recommendation engines: The systems that decide what you see on YouTube, TikTok, Spotify, and Netflix
Administrative AI: Systems used by governments and employers to screen job applications, assess immigration claims, flag benefits fraud, and make bail recommendations

That last category is less visible but in many ways more consequential — it is where AI is already making decisions about people's lives with limited transparency or accountability.

How is it different from earlier software?

Traditional software follows explicit rules written by programmers. A weather forecasting program contains equations describing how air masses behave. A chess engine evaluates possible moves according to a scoring function. A tax return program follows the Income Tax Act, rule by rule. The programmer decides what the software does by writing instructions.

AI systems like LLMs work differently. No programmer explicitly wrote the rules for how ChatGPT answers a question about molecular biology, summarizes a legal document, or writes a sonnet. Instead, the system learned those capabilities from patterns in data, through a process called machine learning. This has several important and sometimes unsettling implications:

Emergent capabilities

Large AI models can do things their creators didn't explicitly program — and sometimes didn't anticipate. Researchers have observed capabilities that seem to appear suddenly as models get larger, rather than smoothly improving. No one programmed ChatGPT to pass the bar exam; it became capable of doing so as a side effect of being trained on enormous amounts of text.

Unpredictable failures

Because the system learned from patterns in data rather than explicit rules, its failure modes are difficult to anticipate. It can produce very confident-sounding wrong answers. It can behave differently on inputs that seem superficially similar. It can replicate biases embedded in its training data in subtle and hard-to-detect ways.

No ground truth

Unlike a calculator, an LLM doesn't "know" when it's wrong. It generates plausible-sounding text based on statistical patterns. This is the root cause of hallucination — the tendency to confidently produce false information. An LLM trained on text that contains errors will learn those errors. An LLM asked about something outside its training data will often make something up rather than say it doesn't know.

Opacity

We cannot easily inspect an AI model and understand why it gave a particular answer. The parameters encode information in a form that is not human-readable. This "black box" quality makes it hard to audit AI systems for bias or errors — a significant challenge for regulators and courts.

What can AI do — and what can't it do?

What current AI does well

Drafting and editing text: emails, reports, essays, code, scripts, legal documents
Translating between languages
Summarizing long documents
Answering factual questions (with significant risk of error)
Generating realistic images, audio, and increasingly video from text descriptions
Writing and debugging software code
Identifying patterns in large datasets
Passing standardized professional exams (bar, medical licensing, CPA)

Important limitations

Hallucination: LLMs can generate false information presented with complete confidence. They don't know what they don't know, and they don't distinguish between things they're sure of and things they're guessing at.
No persistent memory: Most LLMs don't remember previous conversations. Each interaction starts fresh unless the system has been specifically designed otherwise.
Training cutoffs: LLMs are trained on data collected before a certain date. They don't know about recent events, updated laws, or new scientific findings.
Reasoning failures: Despite impressive performance on many tasks, LLMs can fail at surprisingly simple logical reasoning, multi-step arithmetic, or problems that require spatial thinking.
Susceptibility to manipulation: AI systems can be "jailbroken" — tricked by clever inputs into bypassing their safety guidelines and producing harmful content.

A key implication for policy: AI systems are simultaneously more capable than they appear in simple tests and less reliable than they appear in polished demonstrations. This asymmetry makes calibrating AI governance genuinely difficult — and makes it important to test AI systems in realistic conditions before deploying them in high-stakes settings.

AI safety: What are the risks?

The phrase "AI safety" covers a wide range of concerns — from immediate, concrete harms to longer-term risks that are harder to quantify. It helps to organize them into four categories, roughly ordered from the near-term to the speculative.

These risks are taken seriously not just by activists but by the scientists building these systems. Geoffrey Hinton, often called the "godfather of deep learning" and a 2024 Nobel laureate in Physics, resigned from Google partly to speak more freely about AI risks. The heads of the major AI labs — Anthropic, OpenAI, Google DeepMind — have all publicly stated that they believe the technology they are building poses serious risks if not developed carefully.

Misuse Risks

Humans deliberately using AI to cause harm: fraud, disinformation, cyberattacks, weapons development.

Accident Risks

AI systems working as designed but causing unintended harm: bias, hallucination, system failures in high-stakes settings.

Structural Risks

Longer-term changes to society: labour displacement, concentration of power, erosion of democratic institutions.

Frontier Risks

Speculative but serious risks from much more capable future AI systems: misalignment, loss of human control.

Misuse risks

Misuse risks arise when people deliberately use AI to cause harm. These are the most immediate and, in many ways, the most tractable — they are the result of human choices, not system failures.

Fraud, scams, and identity crimes. AI makes it substantially cheaper and easier to generate convincing phishing emails, create fake identities, and produce voice and video impersonations. Voice cloning — replicating someone's voice from a short audio sample — is already being used in fraud schemes. The "grandparent scam," in which fraudsters impersonate a relative in distress, has become vastly more convincing with AI-synthesized voices.

Disinformation at scale. Synthetic text and images can now be produced faster than any fact-checking system can respond. AI-generated "deepfake" videos can put words in public figures' mouths. Entire networks of fake social media accounts can be operated with minimal human effort. This threatens not just individual reputations but the shared epistemic environment that democratic deliberation depends on.

Weapons and mass-casualty threats. This is the most alarming near-term misuse concern: AI systems that can provide meaningful assistance to people attempting to synthesize dangerous biological, chemical, or radiological agents. Major AI labs conduct extensive testing — called red-teaming — specifically to prevent this, and it remains a central concern of AI safety researchers and biosecurity experts.

Cyberattacks. AI significantly lowers the skill threshold for conducting sophisticated cyberattacks, including attacks on critical infrastructure like power grids, water systems, and financial networks.

Accident risks

Accident risks arise from AI systems working more or less as designed, but producing harmful outcomes that were not intended.

Bias and discrimination. AI systems trained on historical data learn the patterns in that data — including its biases. A hiring algorithm trained on a company's past hiring decisions will tend to replicate the criteria those past decisions implicitly applied, including criteria that discriminate based on gender, race, or postal code. A facial recognition system trained predominantly on images of white faces will identify non-white faces less accurately, with real consequences for policing and border security. Importantly, these biases are often invisible without careful auditing.

Hallucination in high-stakes settings. A lawyer using AI to research case law who doesn't verify the AI's output may submit briefs citing cases that don't exist — this has already happened in multiple Canadian and US court cases. A doctor using AI for diagnostic assistance who over-relies on AI output may miss a diagnosis the AI got wrong. The concern is not that AI is useless — it is often useful — but that its failure modes are not well understood by people who rely on it.

Automation of consequential decisions. AI is already being used in Canada to screen immigration applications, flag benefit fraud, assess child welfare risks, and make parole recommendations. When these systems are wrong — and all systems are sometimes wrong — people can lose access to benefits they're entitled to, face deportation, or remain imprisoned because an algorithm misjudged their risk. Unlike a human decision-maker, an algorithm can apply errors consistently to every person who shares a characteristic, at scale.

Structural risks

Structural risks are longer-term changes to how economies and societies are organized — risks that may not be visible as they accumulate, but that could be very difficult to reverse.

Labour displacement. This is the most discussed structural risk. AI is already capable of automating significant portions of knowledge work: drafting, translation, data analysis, customer service, basic legal research, accounting, and more. Economic history suggests that technological displacement is eventually absorbed into new kinds of work — but "eventually" can mean decades of disruption for real people. The current transition is unusual because AI is capable across a much wider range of cognitive tasks than previous automation technologies.

Concentration of power. Training frontier AI models requires enormous computing resources, proprietary datasets, and specialized talent — resources concentrated in a handful of large technology companies, most of them American. This raises questions about who benefits from AI, who controls it, and whether governments can meaningfully regulate it. The risk is not just economic inequality but a fundamental shift in the balance of power between private technology companies and public institutions.

Erosion of epistemic institutions. If AI-enabled disinformation makes it impossible to agree on basic facts — about elections, public health, or climate change — it becomes much harder for democratic institutions to function. The challenge is not just that false information spreads, but that the volume of synthetic content may overwhelm people's capacity to distinguish real from fabricated, leading to generalized distrust.

Frontier risks

Frontier risks are speculative in that they concern AI systems that are substantially more capable than anything that exists today. They are taken seriously because the people who are building toward those systems believe we may reach them within years to decades — and because some of the risks, if they materialize, could be extremely difficult to reverse.

Alignment. The alignment problem asks: how do we ensure that an AI system actually pursues the goals we want it to pursue, rather than proxy goals we inadvertently specified? A system optimizing for "user engagement" may learn that outrage is more engaging than accuracy. A system tasked with "reducing hospital wait times" might find that turning away patients achieves that goal. A much more capable system pursuing a misaligned objective at scale could cause harm that is hard to detect and hard to stop.

Autonomous AI agents. AI systems are increasingly being given the ability to take actions in the world — browsing the web, writing and executing code, booking appointments, controlling other software systems. As this capability expands, and as AI systems become more capable of pursuing multi-step plans with limited human oversight, the opportunities for unintended consequences grow significantly.

Loss of meaningful human control. A recurring concern in AI safety research is that sufficiently capable AI systems might not remain under meaningful human oversight — either because they pursue misaligned goals at a speed and scale that humans cannot monitor, or because they are deployed in competitive contexts where the pressure to automate oversight away is strong. This concern motivates much of the research into AI interpretability, evaluation, and governance frameworks.

The Bletchley Declaration, signed by Canada and 28 other countries in November 2023, specifically acknowledges these frontier risks — noting that the most capable AI models "present a range of both deliberate and unintentional risks" including potential misuse for "attacks on critical infrastructure" and the development of weapons capable of mass casualties. This is not science fiction rhetoric; it is the stated position of the governments of Canada, the United Kingdom, the United States, China, the European Union, and Japan.

Why this matters for Canada specifically

Canada is not a passive bystander in the AI story. It is home to three of the world's leading AI research institutes — Mila in Montreal, the Vector Institute in Toronto, and Amii in Edmonton. It has produced some of the field's most influential researchers, including Geoffrey Hinton (Nobel Prize in Physics, 2024) and Yoshua Bengio (Turing Award, 2018). Canadian companies like Cohere are building AI systems used globally.

And Canadian governments are making decisions about AI right now — decisions about procurement, regulation, investment, and international cooperation that will shape how AI affects healthcare, education, immigration, policing, and democratic life in Canada. Quebec has already passed one of the most significant algorithmic accountability laws in North America. Ontario has regulated AI in hiring. The federal government committed $2.4 billion to AI infrastructure in 2024.

Those decisions are being made. The question is whether they are made with or without an informed public.

Glossary

Every technical term used on this site. Terms are linked throughout the site to this glossary.

Alignment: The challenge of ensuring that an AI system pursues the goals its designers intended, rather than unintended objectives. A misaligned AI might optimize for a proxy goal — like user engagement or minimizing measured errors — in ways that cause broader harm. Alignment research tries to develop techniques for reliably specifying and verifying AI goals.
Algorithm: A set of rules or instructions that a computer follows to solve a problem or make a decision. In modern AI, "the algorithm" often refers to a learned statistical model rather than a set of hand-coded rules.
AI agent / Autonomous system: An AI system that can take actions in the world — browsing the web, sending emails, writing and executing code, controlling other software — with limited or no human oversight of each individual action. Current AI systems are increasingly being deployed as agents, raising new questions about accountability and control.
Compute: The computing power required to train or run AI models. A key constraint on AI development. The most powerful AI models require enormous (and expensive) computing infrastructure — primarily specialized chips called GPUs or TPUs. Access to compute is a major factor in which organizations and countries can develop frontier AI.
Deep learning: A family of machine learning techniques that uses multi-layered neural networks. The dominant approach behind most modern AI capabilities, including large language models and image generation systems.
Deepfake: Synthetic media — video, audio, or images — generated by AI to make it appear that a person said or did something they did not. Deepfake technology has become increasingly accessible and realistic, raising serious concerns for fraud, disinformation, and non-consensual intimate imagery.
Foundation model: A large AI model trained on broad data that can be adapted for many tasks. Large language models like GPT-4 and Claude are foundation models. The term emphasizes that these models form a "foundation" for many downstream applications.
Frontier model: An AI model at or near the leading edge of current capabilities — the most powerful publicly available systems. "Frontier AI" is a governance term used to distinguish the most capable systems (which require more stringent oversight) from more limited AI applications.
Generative AI: AI systems that can produce new content — text, images, audio, video, code — rather than simply classifying or analyzing existing content. ChatGPT is a generative AI. Older AI systems (like spam filters) are not.
GPU (Graphics Processing Unit): A specialized computer chip, originally designed for rendering video games, that turns out to be well-suited for the mathematical operations involved in training AI models. Access to large numbers of GPUs is a key constraint on AI development.
Hallucination: When an AI system generates false information presented as fact. Large language models can "hallucinate" — confidently make up statistics, citations, quotes, court cases, or events that never happened. This is not a bug that can simply be fixed; it is a fundamental property of how these systems work, arising from the fact that they predict plausible text rather than look up verified facts.
Large language model (LLM): An AI system trained on large amounts of text, capable of generating, translating, summarizing, and reasoning about language. Current examples include GPT-4 (OpenAI), Claude (Anthropic), Gemini (Google), and Llama (Meta). The "large" refers to the enormous number of parameters and the scale of training data.
Machine learning: A subfield of AI in which systems learn to perform tasks from data, rather than following explicit hand-coded rules. Modern AI systems are almost entirely based on machine learning rather than hand-written rules.
Neural network: A type of mathematical model loosely inspired by the structure of the human brain, consisting of layers of interconnected nodes. Modern AI systems are typically very large neural networks with billions of parameters.
Parameters: The numerical values inside an AI model that encode what it has learned during training. A model with more parameters can, in principle, represent more complex patterns. Large language models have hundreds of billions of parameters.
RLHF (Reinforcement Learning from Human Feedback): A technique used to fine-tune AI models based on human ratings of their outputs. After initial training, human raters assess which responses are more helpful, more accurate, or safer. The model is then adjusted to produce more responses like the highly-rated ones. RLHF is widely used to make AI chatbots more helpful and to reduce harmful outputs.
Red-teaming: Adversarial testing of an AI system — deliberately trying to make it produce harmful, biased, or otherwise problematic outputs — in order to identify and address safety vulnerabilities before deployment. Named for the "red team" in military exercises, which plays the role of the adversary.
Training: The process of adjusting the parameters of an AI model by exposing it to large amounts of data. During training, the model learns to predict patterns in the data. The result of training is a model that has encoded statistical regularities from that data — including both useful knowledge and any biases or errors in the training data.
Transparency: In AI governance, the principle that organizations should be clear about when and how AI systems are being used, what data they were trained on, and how they reach their outputs. Transparency requirements are a central feature of Quebec's Law 25 and most proposed AI regulations.