Pick up any newspaper, scroll through any tech feed, and you’ll encounter the word “AI” used to describe everything from a spam filter to a system that reportedly passes the bar exam. The term has become almost meaninglessly broad in popular usage — which is a problem, because the differences between types of AI are not cosmetic. They have direct consequences for how we build, deploy, and test AI-based systems.

The ISTQB Certified Tester AI Testing (CT-AI) syllabus v2.0 draws a clear line between three points on the AI capability spectrum: Narrow AI, General AI, and Super AI. Understanding where a system sits on that spectrum is not just academic. It shapes your test strategy, your risk assessment, and your understanding of what can — and cannot — go wrong.

Narrow AI: Where We Actually Live

Every AI-based system in production today is an example of Narrow AI, also called weak AI. The name is slightly misleading — these systems can be extraordinarily capable within their domain. A chess engine beats grandmasters. An image classifier identifies tumors with radiologist-level accuracy. A large language model writes code, summarizes legal documents, and holds convincing conversations.

But here’s the key constraint: a Narrow AI system can only do what it was trained to do. A face recognition model cannot translate languages. A recommendation engine cannot diagnose disease. Each system operates within a tightly defined domain, and it cannot generalize beyond the functions it has learned without being explicitly retrained.

This constraint has a direct testing implication: the operational domain boundary is one of the most important things to test. What happens when the system encounters input that falls outside its training distribution? What happens at the edges? Narrow AI systems can fail in unexpected and sometimes dramatic ways when they encounter data they were never trained on — a phenomenon the CT-AI syllabus addresses through concepts like data representativeness testing and adversarial testing.

Within Narrow AI, the syllabus also introduces the concept of Frontier AI — the most advanced subset of narrow systems, exemplified by today’s most capable large language models. Frontier AI systems push the boundaries of what narrow AI can do: they exhibit highly autonomous decision-making, operate across multiple tasks, and have capabilities that emerge in ways their developers did not explicitly design. They remain task-specific at their core, but the boundary between “very capable narrow AI” and “something approaching general AI” is becoming harder to define — which creates its own testing challenges around scope, specification, and safety.

General AI: The Benchmark That Doesn’t Exist Yet

General AI — also called strong AI or Artificial General Intelligence (AGI) — refers to a system that can perform most intellectual tasks that a human can perform, across an unrestricted range of domains, without needing to be retrained for each new task.

A general AI system would understand, learn, and apply knowledge flexibly. It would solve unfamiliar problems in domains it had never encountered before, much as a human expert brought into a new field can leverage existing reasoning skills to get up to speed quickly.

The critical fact for testers: no such system exists today. Despite the impressive capabilities of frontier AI systems and the hype that surrounds them, we remain firmly in the world of Narrow AI. No AI system currently deployed demonstrates genuine general intelligence — the ability to flexibly transfer knowledge across fundamentally different domains without explicit retraining.

Why does this matter for testing? Because it defines the limits of what current AI systems can reasonably be expected to do. A tester who understands that today’s systems are narrow will not be surprised when a highly capable language model fails at a task that seems trivially simple to a human. They will also be better equipped to define realistic acceptance criteria, scope appropriate test cases, and communicate risk to stakeholders.

Super AI: The Theoretical Horizon

Super AI — artificial superintelligence — is the form of AI in which a system continuously improves itself without human intervention or control, surpassing human intelligence across all domains. This is the scenario associated with the concept of the technological singularity: the point at which AI-based systems transition from general to super AI, triggering a cascade of self-improvement beyond human capacity to predict or control.

It’s important to be precise here: Super AI does not currently exist, and there is no scientific consensus on whether or when it might. The CT-AI syllabus includes it as part of the definitional landscape, not as an imminent engineering challenge.

That said, the concept is not irrelevant to testers. It anchors the far end of a risk spectrum that begins with the very real challenges of testing today’s narrow AI systems. Many of the regulatory frameworks now emerging — including the EU AI Act — are motivated in part by concerns about the trajectory of AI development, not only its current state. Testers who understand this spectrum are better placed to engage with risk-based conversations about AI governance, safety, and ethics.

The Testing Takeaway

For practical purposes, the most important thing a tester can take from this section of the CT-AI syllabus is a clear-eyed understanding of what today’s AI systems actually are: powerful, narrow, brittle at their edges, and incapable of the kind of flexible generalization that humans take for granted.

That understanding should inform every aspect of your test approach:

• Scope your tests to the operational domain — and deliberately test what happens outside it.
• Don’t assume capability transfers. A model that performs brilliantly on one task gives you no guarantee about adjacent tasks.
• Treat impressive performance as a reason for more rigorous testing, not less. Frontier AI systems fail in ways that are hard to predict precisely because their capabilities are so broad.
• Keep the regulatory context in mind. The distinction between narrow and general AI is embedded in emerging regulations, and testers will increasingly be asked to provide evidence that systems stay within their intended scope.

The AI landscape is moving fast. The spectrum from Narrow to Super AI is the map you need to navigate it.