Comparing AI models: four axes, expert systems, and capability levels
Choosing, buying or reviewing an AI system means placing it on a handful of standard comparisons: what kind of output it produces, who controls it, how big it is, what data it handles, and how capable it is. These distinctions return in Domain IV when
Why this matters for the exam
Choosing, buying or reviewing an AI system means placing it on a handful of standard comparisons: what kind of output it produces, who controls it, how big it is, what data it handles, and how capable it is. These distinctions return in Domain IV when a deployment has to be chosen.
What you need to know
The last topic covered how models are trained. This one compares the kinds of models the training produces. Four axes come up constantly, and each carries its own governance trade-offs.
Classic versus generative models
| Classic models | Generative models | |
|---|---|---|
| Focus | Specific tasks with deterministic outputs: the same input produces the same result | Creating new data instances that resemble the training data |
| Method | Structured algorithms and fixed rules | Learn the underlying distribution of the input data, enabling novel, diverse output |
| Examples | Decision trees, linear regression | GPT-style models, GANs |
A classic model gives the same answer every time. A generative model may not, and that difference drives a lot of governance later: probabilistic systems need more testing and more monitoring, because a single passing test proves less.
Proprietary versus open-source models
The second axis is about who controls the model, and on what terms.
| Proprietary | Open-source | |
|---|---|---|
| Definition | Developed by a specific organization; access and use are restricted, usually on commercial terms | Publicly available to use, modify and distribute |
| Trade-off | Can limit transparency and independent auditing; carries vendor lock-in risk | Promotes collaboration, innovation and transparency, but raises quality-control and security risk: guardrails can be stripped out |
Small versus large language models
Within language models, the next axis is scale.
| Dimension | Large language models (LLMs) | Small language models (SLMs) |
|---|---|---|
| Size | Billions to trillions of parameters | A few million to several billion parameters |
| Training data and versatility | Vast, diverse datasets; versatile across many tasks | Focused datasets; excel in specific domains, less versatile |
| Resources | Significant compute, often thousands of GPUs, substantial energy use | More resource-efficient; can often run on standard hardware |
| Bias | Can pick up bias from vast, varied training data; fine-tuning often needed | Also susceptible, but easier to fine-tune for a specific task |
Language models versus multimodal models
The final axis concerns what kinds of data a model can take in and produce.
| Dimension | Language models (LMs) | Multimodal models (MMMs) |
|---|---|---|
| Data types | Process and generate text only | Handle text, images, audio and video together (CLIP, DALL·E) |
| Capabilities | Powerful within natural language processing, limited to text | Integrate multiple sources at once: caption an image, answer questions about a video, generate an image from text |
| Applications | Chatbots, virtual assistants, content generation | Image captioning, video analysis, multimodal search |
Multimodal models carry their own governance flags. The WHO issued specific AI ethics guidance for large multimodal models in 2024, covering inaccurate or biased output affecting health decisions, poor-quality or biased training data, and a heightened privacy risk: when image, text and voice data combine, individuals become easier to re-identify.
The axes are meant to be combined. "A publicly available model with billions of parameters that generates code" is open-source, large, and generative — three placements from one sentence.
Expert systems
One older category is still in service. An expert system mimics the decision-making of a human expert in a specific field by drawing inferences from a curated knowledge base. It is designed to support humans, not replace them: a diagnostic aid that helps a doctor determine a cancer stage, not one that diagnoses unsupervised. Three components make it work:
- Knowledge base — an organized collection of domain-specific facts provided by human experts, sometimes with additional external sources.
- Inference engine — pulls relevant information from the knowledge base and applies if-then rules to reach a decision; often includes a way for users to review the reasoning.
- User interface — where the user puts in a problem or question and receives the resolution.
Expert systems are brittle by nature: they fail completely on edge cases their rule set never anticipated. They are also vulnerable to data poisoning if the knowledge base itself is corrupted or goes stale.
A fifth comparison: how capable is it?
The axes above describe how a model is built and controlled. AI is also compared by how capable it is, on a scale that runs from everything in use today to systems that remain hypothetical.
| Category | Also called | Definition | Status |
|---|---|---|---|
| Artificial narrow intelligence (ANI) | Weak AI | Performs one task, or a narrow set of related tasks, at high proficiency. | Exists — "we are here." |
| Broad AI | — | Coordinated narrow systems working together across a wider task set, like the many subsystems of an autonomous vehicle. Still many narrow systems, not one general mind. | Exists. |
| Artificial general intelligence (AGI) | Strong AI | Would match full human cognitive ability: thinking, understanding, learning and solving problems across any context. | Does not exist yet. |
| Artificial superintelligence (ASI) | — | Would exceed human intellectual ability across every field; self-aware, and able both to understand human emotion and to evoke it in others. | Does not exist yet. |
Two of these labels are simply alternative names: weak AI means narrow AI, and strong AI means AGI. So strong AI does not exist yet, and being extremely good at one task does not change that. A self-driving car may handle city traffic brilliantly, but it can only drive. Until a system can take on any new kind of problem the way a person can, it is narrow AI. One label is course-specific: "broad AI" comes from the AIGP materials, and most writing outside them jumps straight from narrow to general.
Next up: where governance begins — why AI needs rules of its own, and the OECD's standard framework for describing any system.
Remember
- Classic models are deterministic and task-specific. Generative models learn a data distribution to create new content, and their outputs are probabilistic.
- Proprietary means restricted access and potential vendor lock-in. Open-source means public and collaborative, with weaker guardrail guarantees.
- LLMs are huge, versatile and resource-hungry. SLMs are smaller, efficient and domain-specialized.
- Language models handle text only; multimodal models integrate text, image, audio and video.
- Expert systems run on a knowledge base, an inference engine and a user interface, and are brittle outside their rule set.
- Today's AI is narrow AI, sometimes coordinated into broad AI. Strong AI is another name for AGI, and it does not exist yet.
Practise this topic
Domain I is free in the app, including its practice questions and flashcards, with progress tracking and no card details.
Previous: How machines learn: supervised, unsupervised, reinforcement and semi-supervised training
Next: Why AI needs its own governance, and how to describe a system: the OECD's five dimensions
Back to the AIGP study guide.
AI Governance Study is an independent study aid. It does not represent a government entity: it is not affiliated with, endorsed by or authorised by any government, government agency or regulatory authority, and it does not provide government services or legal advice. Laws and frameworks are described in our own words — the official texts are listed at official sources. It is also not affiliated with, endorsed by, or sponsored by the IAPP. The AIGP name is used only to identify the exam this material helps you prepare for.