AI in Plain English
A practical mental model for modern AI — what it can do, what changed, and what still goes wrong.
AI is an umbrella, not one thing
Artificial intelligence is a broad label for systems that perform tasks associated with perception, language, prediction, pattern recognition or decision support. Machine learning is one way to build such systems: instead of hand-writing every rule, developers train models on examples.
Predictive AI
Classifies, ranks or forecasts: spam filters, fraud detection, recommendations and many traditional ML systems.
Generative AI
Creates new text, images, audio, video or code based on patterns learned during training.
Multimodal AI
Works across several kinds of input or output — for example a model that can read a PDF, inspect an image and answer in voice.
Agentic AI
Adds tools and actions: the system can search, call software, update files or carry out parts of a workflow.
What changed after the first chatbot wave?
The useful change is not that AI suddenly “became human.” It is that model systems became more capable around the model itself. Some can spend more computation on hard tasks, accept far larger and multimodal inputs, retrieve fresh information, run code, call tools and coordinate multi-step work.
A better mental model
Think of a modern AI product as model + context + tools + permissions + product rules. The underlying model matters, but the surrounding system often matters just as much.
How generative AI produces an answer
Your input is converted into model-readable units
Text is split into tokens; images, audio and other inputs are converted into numerical representations.
The model predicts useful continuations
It uses patterns learned in training plus the current context to generate an output. This can look like understanding even though the mechanism is not human thought.
The product may add retrieval or tools
A system can fetch documents, browse, run code or call another service before producing the final response.
Reasoning models: stronger does not mean infallible
Reasoning-focused models are designed to use extra computation on harder problems. They can be markedly better on tasks such as coding, planning and technical analysis, but they still make factual, logical and tool-use mistakes. Treat them as more capable assistants, not automatic authorities.
Five limits worth remembering
- Hallucinations: a model can invent or distort a fact while sounding certain.
- Context is finite: large context windows help, but important details can still be overlooked.
- Freshness varies: a model may not have live information unless the product connects it to current sources.
- Outputs are probabilistic: the same prompt can produce different results.
- Capability is not authority: impressive writing does not confer medical, legal, financial or factual certainty.
Useful default
Use AI freely for brainstorming, transformation and drafts. Raise the verification standard as the consequence of being wrong rises.
A simple prompt that works across models
Goal: [what I need] Context: [what the model should know] Evidence: [documents, data or sources to use] Constraints: [what to avoid, limits, deadlines] Output: [format, length, audience] Quality check: flag assumptions and anything that needs verification.
That structure is deliberately boring. It works because it reduces ambiguity, not because it contains secret words.