Why Is AI So Hard to Define?
You’ve probably noticed that the term “AI” is used to describe many different things.
It can refer to entire products, single features, model architectures, research fields, or marketing claims. In documentation work, you may be asked to “explain the AI” without a shared understanding of what that means.
The difficulty isn’t that AI is mysterious. It’s that AI is often treated as one thing when it is really a mix of goals and technical tools.
This article offers a practical mental model you can use across the rest of the series. The aim is to simplify AI without flattening important distinctions, so your documentation choices are clearer and more consistent.
A Practical Definition of Artificial Intelligence
Artificial intelligence refers to algorithmic systems that use statistical models and learned patterns to perform tasks typically requiring human cognition, such as recognizing patterns, making predictions, understanding language, or deciding on actions.
Two clarifications matter:
- AI describes a goal or capability, not a single technology.
- Algorithms and models are the means used to reach those goals, not the definition itself.
When you document AI features, you are usually describing what the system is trying to accomplish, not how it is built.
The Two Layers of AI: Goals and Tools
A useful way to understand AI is to separate what we want it to do from how we build it.
Layer 1: Human-Like Goals
These are examples of goals AI systems pursue:
- Understanding language
- Interpreting images
- Generating text, images, or other content
- Making predictions from data
- Acting within defined boundaries
These are application goals. They are not different kinds of intelligence. A single system can pursue several at once.
Layer 2: Technical Tools
These are examples of tools used to pursue those goals:
- Neural networks: layered mathematical systems that detect patterns in data by adjusting weights (numerical values that determine how strongly one input influences the next step in the model) through training.
- Transformer architectures: a type of neural network especially good at processing sequences like language by paying attention to relationships between words.
- Supervised, unsupervised, and reinforcement learning: the main learning approaches in machine learning: learning from labeled examples, finding patterns without labels, or improving through feedback and rewards.
- Generative models: models designed to produce new content, such as text or images, based on patterns learned from existing data.
The same technical tool can support many different goals. A transformer model, for example, can be used for translation, summarization, prediction, or classification.
Separating goals from tools helps prevent confusion. It keeps your documentation focused on user-visible behavior instead of internal implementation.
Where Machine Learning Fits
One of the most common technical foundations behind these tools today is machine learning.
Machine learning is a subset of AI in which systems are not explicitly programmed with fixed rules. Instead, they develop models by identifying patterns in data and refining those models as they are exposed to more data over time.
Machine learning became the dominant approach because several conditions converged: large available datasets, major increases in computing power, and advances in algorithms. Together, these made it practical to build systems that learn from data at scale.
This context matters because when you document an AI-enabled product today, you are almost certainly documenting something built on machine learning.
Why AI Categories Overlap
AI categories often overlap because modern systems pursue multiple goals at the same time.
For example:
- A language model may understand text, classify it, and generate a response.
- A multimodal system (a system that works with more than one type of input, such as text and images) may read text, interpret images, and produce new content.
Labels such as “generative AI” or “predictive AI” describe dominant behavior, not completely separate types of systems.
Overlap is normal. It reflects how flexible the underlying tools are.
Artificial Narrow Intelligence (ANI): The Reality Today
All production AI systems today are forms of Artificial Narrow Intelligence (ANI).
ANI systems are designed for specific tasks. They don’t generalize across domains in the way humans do. A system that performs well in one area doesn’t automatically transfer that ability elsewhere. For example, a model trained to detect fraudulent credit card transactions cannot automatically diagnose medical conditions without being trained separately for that task.
For documentation, this matters because it grounds expectations. AI systems are capable within defined boundaries. They are not general-purpose minds.
What AI isn’t
AI isn’t:
- A single model or architecture
- A guarantee of correctness
- Human-level understanding or intent
- Autonomous decision-making by default
When documentation blurs these distinctions, users may infer capabilities that are not present.
Clear writing starts by describing limits as well as capabilities.
Why This Mental Model Matters for Technical Writers
When AI is described loosely, documentation becomes vague.
You may see:
- Features described as “intelligent” without explaining behavior
- Overpromising about accuracy or autonomy
- Inconsistent terminology across pages
Understanding AI as a combination of goals and tools helps you:
- Choose clearer category labels
- Explain what a feature actually does
- Separate behavior from implementation
- Set realistic expectations early
This makes onboarding easier and reduces confusion later.
How This Article Fits Into the Series
This article provides a shared mental model and baseline vocabulary.
It supports the AI Product Landscape and the individual category articles. It gives you a way to interpret terms like “generative,” “predictive,” or “conversational” without treating them as separate species of AI.
This article doesn’t replace category definitions or the “Documenting X” guides, which focus on writing decisions rather than conceptual framing.
Takeaways
- AI describes goals and capabilities, not specific technologies.
- Most AI systems combine multiple goals and tools.
- Overlap between AI categories is expected.
- All current AI systems are narrow and task-specific.
- Most modern AI systems are built using machine learning, which means their behavior is shaped by data and learned patterns rather than fixed rules.
- Clear documentation begins with a stable mental model of what AI is and isn’t.