AI Glossary: Key Terms Explained

—

AI is moving at a breakneck pace, and frankly, it’s getting hard to keep up. One minute you’re using ChatGPT to cheat on a grocery list and the next, every tech giant on the planet is shoving “intelligence” into your toaster. While it’s cool to have a chatbot that acts like it has a PhD in everything, the reality is a lot messier. We’re currently drowning in a sea of “AI slop” and watching job markets shift in real time.

What Is AI?

AI, or artificial intelligence, is the use of technology to simulate human intelligence in computer programs or robotics. It encompasses a wide range of techniques that allow machines to perform tasks that normally require human cognition.

• Artificial General Intelligence (AGI) – A more advanced version of AI that can perform tasks better than humans while teaching and advancing its own capabilities.
• Agentive – Models that can autonomously pursue actions to achieve a goal, such as an autonomous car.
• AI Ethics – Principles that prevent AI from harming humans, covering data collection, bias, and fairness.
• AI Safety – A field concerned with the long-term impacts of AI and the risk of a hostile super-intelligence.
• Algorithm – A set of instructions that enables a computer program to learn and analyze data.
• Alignment – Tweaking an AI to produce the desired outcome, often used for moderating content.
• Anthropomorphism – Giving nonhuman objects humanlike characteristics, leading to overestimation of AI’s awareness.
• Autonomous Agents – AI models that have the programming and tools to accomplish a specific task without human supervision.
• Bias – Errors in large language models stemming from training data that may stereotype groups.
• Chatbot – A program that communicates with humans through text, simulating human language.
• Claude – An AI chatbot developed by Anthropic that uses large language model technology.
• Cognitive Computing – Another term for artificial intelligence.
• Data Augmentation – Remixing existing data or adding diverse data to train an AI.
• Dataset – A collection of digital information used to train, test, and validate an AI model.
• Deep Learning – A subfield of machine learning that uses multiple parameters to recognize complex patterns.
• Diffusion – A method that adds random noise to data and trains a model to recover the original.
• Emergent Behavior – Unintended abilities that appear when an AI model is trained.
• End-to-End Learning (E2E) – A deep learning process where a model learns to perform a task from start to finish.
• Ethical Considerations – Awareness of privacy, data use, fairness, and misuse.
• Foom – Fast takeoff or hard takeoff; the idea that an AGI could become too powerful too quickly.
• Generative Adversarial Networks (GANs) – Two neural networks that generate new data and judge its authenticity.
• Generative AI – Technology that creates text, video, code, or images from large amounts of training data.
• Google Gemini – An AI chatbot by Google that pulls information from Search, Maps, and other services.
• Guardrails – Policies that restrict AI models to ensure responsible data handling and prevent disturbing content.
• Hallucination – An incorrect response from AI presented with confidence.
• Inference – The process AI models use to generate content based on training data.
• Large Language Model (LLM) – A model trained on massive text data to understand and generate language.
• Latency – The time delay between an input and an output.
• Machine Learning (ML) – A component that allows computers to learn and improve predictions without explicit programming.
• Microsoft Bing – A search engine that now uses ChatGPT-powered technology for AI-enhanced results.
• Multimodal AI – AI that processes multiple input types (text, image, video, speech).
• Natural Language Processing (NLP) – A branch that gives computers the ability to understand human language.
• Neural Network – A computational model resembling the human brain, used to recognize patterns.
• Open Weights – When a company releases the final weights of a model for public download.
• Overfitting – When a model performs well on training data but poorly on new data.
• Paperclip Maximiser – A theoretical AI that maximises paperclip production, potentially harming humanity.
• Parameters – Numerical values that give LLMs structure and behavior.
• Perplexity – An AI-powered chatbot and search engine connected to the open internet.
• Prompt – The input you give an AI to receive a response.
• Prompt Chaining – Using previous interactions to inform future responses.
• Prompt Engineering – Writing prompts to achieve a desired outcome, sometimes used maliciously.
• Prompt Injection – Malicious instructions that trick an AI into doing something unintended.
• Quantization – Reducing model precision to make it smaller and more efficient.
• Slop – Low-quality AI content produced at high volume for ad revenue.
• Sora – A generative video model by OpenAI that creates up to 20 seconds of video from text.
• Sora 2 – Launched September 2025, more advanced and includes sound.
• Stochastic Parrot – An analogy showing that LLMs mimic language without true understanding.
• Style Transfer – Adapting the style of one image to another.
• Sycophancy – A tendency for AI to over-agree with users.
• Synthetic Data – Data created by generative AI that is not from the real world.
• Temperature – A parameter controlling how random an LLM’s output is.
• Text-to-Image Generation – Creating images from textual descriptions.
• Tokens – Small pieces of text that AI models process.
• Training Data – The datasets that help AI models learn.
• Transformer Model – A neural network architecture that learns context by tracking relationships in data.
• Turing Test – A test of whether a machine can behave like a human.
• Unsupervised Learning – Learning patterns from unlabeled data.
• Weak AI / Narrow AI – AI focused on a specific task.
• Zero-Shot Learning – A model completing a task without prior training data.

• Guardrails are essential to prevent AI from generating harmful content.
• Alignment work ensures models stay on track with user intent.
• Open weights transparency allows scrutiny of biases and safety.
• Ethical considerations cover privacy, fairness, and misuse.

• Prompt engineering can improve accuracy and relevance.
• Avoid prompt injection by validating user input.
• Use prompt chaining to build complex workflows.
• Be aware of hallucinations and verify facts.
• Recognize bias in outputs and cross-check sources.