Navigate This Resource

What is Generative AI?

Institutional Policies

Navigating AI

Building Skillset

Ethical Considerations

Pedagogical Considerations

Cultivating Literacy

AI Futures

A very brief history of Artificial Intelligence

Artificial intelligence may feel like a sudden arrival our collective consciousness, but it is the product of decades — even centuries — of human imagination, philosophical inquiry, and technical experimentation. At its core, AI is a field devoted to building machines and systems capable of performing tasks that, until very recently, were considered the exclusive domain of human intelligence: understanding language, recognizing patterns, solving problems, making decisions.

The dream of artificial intelligence even predates the computer itself: Ancient Greek philosophers asked whether thinking could be formalized. In the 19th century, Ada Lovelace envisioned machines that might one day compose music or create art. With the invention of the modern computer in the mid-20th century, this centuries-old philosophical question began to take technological shape.

Evolution of Thinking Machines

In 1956, at the Dartmouth Conference, pioneers like John McCarthy, Marvin Minsky, and Claude Shannon proposed that human intelligence could, in principle, be mechanized. Early AI efforts centred on symbolic reasoning: encoding knowledge through hand-crafted rules that computers could follow to simulate logic and problem-solving. Programs like the Logic Theorist and expert systems such as MYCIN marked important milestones, but these approaches quickly ran into the limits of complexity and adaptability. As real-world tasks proved far less orderly than anticipated, optimism waned, leading to periods of reduced funding and enthusiasm known as "AI winters" in the 1970s and again in the late 1980s.

Despite these setbacks, the pursuit of intelligent systems quietly persisted. In the 1980s and 1990s, a major shift occurred: researchers began focusing on machine learning, developing methods that allowed computers to learn from data rather than relying on pre-programmed rules. Because real-world environments are messy and unpredictable, it is virtually impossible to hand-code every rule an intelligent system might need. Machine learning emerged as a way to let systems learn directly from data, enabling them to adapt and generalize beyond what was explicitly programmed. As computational power grew and vast amounts of digital data became available, these learning-based methods, particularly deep learning using multi-layered neural networks, triggered a paradigm shift in AI research and applications. AI systems gradually mastered complex tasks such as image and speech recognition, language translation, and strategic game play — not through rigid logic, but through flexible pattern recognition, setting the stage for the generative AI breakthroughs of the 2010s.

Source: Digital Wellness Institute (2017), Artificial Intelligence Timeline Infographic

As the 2010s progressed, advances in deep learning and natural language processing laid the foundation for a dramatic shift: the emergence of generative AI. Rather than simply classifying data or recognizing patterns, generative models began to create entirely new content — from realistic images and videos to human-like text and even music. The introduction of Generative Adversarial Networks (GANs) in 2014 enabled machines to generate convincing synthetic images, while the rise of large language models like OpenAI’s GPT series demonstrated that AI could produce coherent, contextually relevant text at scale. These developments marked a turning point: AI was no longer just analyzing the world — it was now beginning to simulate and recreate it, blurring the lines between human and machine-generated creativity.

These generative capabilities did not remain confined to research labs for long. By the early 2020s, tools powered by these models — most notably ChatGPT — became widely accessible to the public, bringing the disruptive power of generative AI directly into classrooms, lecture halls, and learning management systems. What had been a technical revolution quickly became an educational reckoning.

A sudden shock for educators: The impact of generative AI on academic practice was immediate and profound. As students began using tools like ChatGPT to produce essays, summarize readings, and answer assignments, institutions around the world faced a wave of panic and uncertainty that is still being felt today. Questions of academic integrity, authorship, and assessment remain front and centre across disciplines, as educators continue to grapple with the challenges and possibilities introduced by generative AI. In response, universities have had to move quickly to develop policies, update assessment practices, and provide clearer guidance for both instructors and students. To learn more about how UBC is addressing these issues, including emerging guidelines and protocols for AI use in research and classrooms, explore the Institutional Policies and Guidelines section of this resource.

For a glimpse into the early reactions to the sudden arrival of genAI in classrooms, see The College Essay Is Dead

Grappling with Ethics, Expression, and Ownership

Generative AI brought AI out of the background and into the domain of human expression and creativity. It shifted AI's role from a tool of analysis to a collaborator in artistic, literary, and intellectual production, fundamentally reshaping our assumptions about authorship, creativity, and knowledge itself.

However, with this shift came urgent questions about who owns and controls the data that feeds these systems, and whose knowledge, voices, and cultural artifacts are being absorbed and recombined without human consent. Issues of data sovereignty — particularly for Indigenous communities, marginalized groups, and non-Western contexts — came sharply into focus. Scholars and activists have sounded the alarm about how vast datasets often scrape, appropriate, and repurpose information through extractive practices that bypass ethical engagement in ways that risk perpetuating harm, erasure, or misrepresentation.

At the same time, concerns deepened around the nature of machine learning itself: how these systems learn, what biases they encode, and how they might quietly reinforce existing social inequalities. Fears emerged not only about data exploitation but also about the future of human learning and expression. If machines can instantly generate essays, poems, artworks, and ideas, what does that mean for the development of critical thinking, creativity, and the development of authentic voice in education and beyond?

Rather than being neutral tools, generative AI systems increasingly appeared as extensions of societal structures, carrying hidden assumptions, exclusions, and risks. In this new terrain, grappling with questions of ethical design, authorship, and the value of human learning became as urgent as developing technical fluency. (Explore the Ethical Considerations section of this resource to learn more about these issues.)

Quick Breakdown of How GenAI Works and Implications for Education

In light of the ethical concerns raised around generative AI — from data appropriation to the risk of harm and misrepresentation — it’s essential to understand how these systems actually function. A basic grasp of how GenAI tools are trained and how they generate outputs can help educators and learners make more informed choices: not just about how to use these tools, but whether to use them, and which ones align best with their ethical, pedagogical, and disciplinary values. This section offers a quick breakdown of how GenAI works, along with key implications for education.

Training on Massive Datasets

Generative AI models are trained on massive datasets — often encompassing huge portions of the publicly available internet, including text, images, and audio — to learn patterns, structures, and relationships. Instead of memorizing, they develop a statistical "feel" for how real-world content is typically organized and expressed.

Different generative AI models are trained on different datasets, which shape their abilities, biases, and limitations. While some models are trained broadly on public internet data, others are trained on more curated, specialized, ethically sourced, and/or licensed collections.

Educational Implication:
As you work with AI tools, remember that different models are trained on different collections of data. What an AI can generate — and what it overlooks — depends on the sources it was trained on. Always ask: whose knowledge is represented, and whose is missing? Understanding where the training data comes from is an important step in building critical awareness, helping you to use AI more thoughtfully and responsibly.

Read more: Learn more about AI training and datasets (IBM)

Learning Patterns, Not Facts

Rather than storing exact answers or verified facts, generative AI models learn patterns and probabilities: how likely one word is to follow another, how shapes fit together in an image, or how ideas typically flow. They don’t "know" things just at they don't "not know" things— they simply (or not so simply!) generate outputs based upon statistical predictions.

For this reason, generative AI can sometimes produce what are called hallucinations — these are outputs that appear convincing but are factually incorrect or misleading. Because the model doesn’t “know” truth from falsehood, it can produce errors with total confidence. These hallucinations are more likely to occur when the model has limited or poor-quality training data on a given topic. In such cases, the model may lack sufficient examples to generate an accurate response, but it will still attempt to produce something plausible based on what it does know — often resulting in made-up facts, incorrect references, or distorted information.

Educational Implication:
As a student, it’s critical to remember that generative AI does not understand what it produces — it predicts based on patterns. Because of this, AI-generated content can sometimes include convincing but incorrect information, fabricated citations, or misleading ideas. It is crucial to verify all AI-generated outputs and not be misled by their appearance of correctness. Always (always!) approach AI outputs with a deeply critical eye: verify facts independently, check sources carefully, look for embedded biasses and omissions. Always keep your hands on the wheel. What you think matters — far more than what a model predicts you might say.

Read more: Understanding predictive modeling (DeepLearning.AI)

Yin, Z., Sun, Q., Guo, Q., Wu, J., Qiu, X., & Huang, X. (2023). Do large language models know what they don't know?. arXiv preprint arXiv:2305.18153.

Reinforcement from Human Learning and Agreeableness Bias

Many AI models are fine-tuned using a process called Reinforcement Learning from Human Feedback (RLHF). This means that after being trained to predict language patterns, the model is further refined using human preferences — where labelers rate or rank responses that seem most helpful, polite, and aligned with user intent. Over time, the model learns to optimize for what humans like to hear, not necessarily what is most accurate or challenging.

This training process, while improving fluency and helpfulness, also introduces a bias toward agreeableness. Because the model lacks self-awareness or the ability to verify truth, it will often echo or affirm a user's assumptions, even when they’re incorrect. It is not "trying" to agree — it’s simply generating the most plausible continuation of a conversation based on statistical patterns and human-rated reward signals.

This can be especially problematic in education or research, where critical thinking and productive disagreement are essential. Without careful prompting or external fact-checking, the model’s smooth and confident tone can give false information undue credibility.

Educational Implication:
In educational contexts, this tendency toward agreeableness can be particularly concerning. Critical thinking, debate, and the ability to question assumptions are essential parts of learning — but generative AI often reinforces what the user already believes. Its fluent and confident tone can make inaccurate or misleading information feel trustworthy — and that can be very difficult to resist. Without deliberate prompting or independent fact-checking, learners may accept surface-level answers that subtly discourage deeper inquiry or critical engagement.

Read more: Ouyang, Long, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems 35 (2022): 27730-27744.

Problems with Reinforcement Learning from Human Feedback (RLHF) for AI safety

Reasoning models and how they differ

Reasoning AI began gaining attention around 2022, with the rise of techniques like chain-of-thought prompting and models designed to handle more complex, multi-step problems. Unlike standard generative AI, which simply predicts the next word based on statistical patterns, reasoning AI aims to simulate step-by-step thinking — breaking down a problem into smaller parts and generating intermediate steps along the way. This process helps improve accuracy and transparency, especially for tasks involving logic, math, or critical reasoning. Because these models generate and evaluate multiple reasoning paths rather than a single answer, they require significantly more computational power, making them more resource-intensive to train and run.

Educational Implication:
Reasoning AI models are designed to simulate step-by-step thinking, which means they can help you work through complex problems by showing how a solution might unfold — not just giving you the final answer. This can be really useful for learning how to approach tasks in math, logic, writing, or analysis. But it’s important to remember that, like all AI models, the ones that can reason don’t actually understand what they’re doing. Which means they can produce flawed or biased reasoning that still sounds convincing. These tools also require significant computing power, which means access to the most advanced models is often limited to those with institutional resources or the ability to pay. This creates real equity concerns — not all students have the same opportunities to benefit from these tools, and educational systems must work to ensure that AI doesn’t deepen existing gaps. Most importantly, relying on AI too heavily can get in the way of developing your own thinking skills. Use these tools to support your learning — not to replace the hard, essential work of figuring things out for yourself.

AI, Neural Networks, and the Human Brain

Modern generative AI models, like ChatGPT and DALL-E, are loosely inspired by the structure of the human brain. Neural networks — the building blocks of these systems — are modeled after how neurons in the brain connect and pass information. Each "node" in a neural network receives signals, processes them, and passes them forward, much like neurons firing in response to stimuli.

Transformers, the foundation of today’s most powerful AI models, add another layer of sophistication: they are designed to pay attention to the most relevant parts of the information they process — similar to how your brain focuses selectively when you read, listen, or solve problems. This "attention" mechanism allows AI to maintain coherence over long pieces of text or complex images, helping it simulate human-like expression.

Despite these inspirations, AI remains a long way from replicating the true complexity of the human brain. Human thought is shaped by experience, emotion, embodiment, and a lifetime of contextual understanding — none of which AI systems possess. Generative AI predicts patterns based on data; it does not reason, reflect, or truly understand meaning. While some researchers believe that AI might one day approach human-level general intelligence (AGI), others argue that the embodied, emotional, and experiential dimensions of human cognition can't be fully captured by machines. Recognizing this gap — and the ongoing debate around it — is crucial: it reminds us that while AI can mimic certain outputs of thinking, it does not, and for now cannot, replicate the depth, nuance, or ethical responsibility of human intelligence.

Educational Implication:
Understanding how generative AI works — and where it falls short — can deepen your insight into human cognition. Comparing AI's pattern-based predictions with human learning can highlight the unique strengths of human intelligence: our ability to reason, reflect, feel, and draw meaning from experience. At the same time, it reveals the challenges we share, such as bias, gaps in knowledge, and the influence of context. By critically examining AI, you also have the opportunity to better understand your own cognitive processes, strengthening your capacity for thoughtful, ethical, and creative learning.

Sources and Further Reading:

Buchanan, B. G. (2005). A (very) brief history of artificial intelligence. The AI Magazine, 26(4), 53-60. https://doi.org/10.1609/aimag.v26i4.1848