2.2 Narrow AI, General AI and Super AI

2.2 Narrow AI, General AI and Super AI

Marketing copy loves the word intelligence. Testers cannot afford that blur. The CT-AI picture uses three capability classes—narrow AI, general AI, and super AI—plus a modern label, frontier AI, that sits inside narrow AI. If you mix those classes, you will write the wrong risks, the wrong oracles, and the wrong stop-ship criteria.

Narrow AI, also called weak AI

Narrow AI (weak AI) is designed to perform specific tasks. It is not a polite way of saying the product is bad. It is the name for every AI-based system actually deployed today. A warehouse vision model that finds damaged boxes, a speech-to-text engine, a machine-translation service, a fraud scorer, and a coding assistant are all narrow. Each can look spectacular inside its job and still be helpless one job to the left.

Narrow systems operate inside a limited domain. They can be extremely efficient at that domain. They do not freely generalize beyond the functions they have learned. A face-recognition service that is excellent at matching badge photos does not wake up one morning able to translate a maintenance manual. To take on translation, the organization must retrain, swap in another model, or assemble a new pipeline. That retraining is a project with data, labels, evaluation, and release gates—not a side effect of being good at faces.

The nickname weak is historically unfortunate. Weak here means narrow scope, not low accuracy. A model that detects tumors on a defined imaging protocol can outperform a tired human on that protocol and still be narrow. Testers should describe capability in task language: what input, what output, what population, what environment. A claim such as the system is intelligent is not a testable requirement.

Everyday pictures testers already know

• Image recognition that sorts photos or flags defects on a line
• Speech processing that turns audio into text or detects a wake word
• Language translation between a trained pair of languages and registers
• Recommendation that ranks films, products, or next best actions
• Predictive maintenance that scores likely failure for a known asset class

Each of those systems can fail outside its training envelope: a new camera, a new accent, a new product catalog, a new failure mode. That is not a plot twist. That is what narrow means in the field.

Frontier AI is still narrow

Frontier AI is a subset of narrow AI. The term points at the most advanced systems currently fielded, especially large generative models and other large-scale systems that can take highly autonomous actions (tool use, multi-step workflows, limited unsupervised operation). Frontier systems push today's capability ceiling. They still have not achieved the versatility of general AI.

This is the sentence the exam wants you to keep when a vendor demo looks magical: impressive autonomy is not the same as general intelligence. A model that can draft email, write a function, and summarize a PDF is still doing language-shaped tasks it was trained and instructed to do. Give it a genuinely new sensorimotor skill, a new regulatory regime, or a new scientific instrument with no relevant training or tools, and it does not become a general mind. It becomes a narrow system off its map.

Testers working on frontier systems should still inventory task boundaries. What tools may the agent call? What data stores may it read? What actions are irreversible? High autonomy raises the cost of a wrong action, but the class of system remains narrow until it can perform most human intellectual tasks without retraining.

General AI, also called strong AI

General AI (strong AI) would possess the ability to perform most intellectual tasks that a human can. It would understand, learn, and apply knowledge across a broad range of tasks without needing to be retrained for each new task. It would exhibit human-like reasoning and adaptability, including solving unfamiliar problems in various domains the way people do when they transfer skill from one job to another.

No AI-based system today possesses general intelligence. That statement is easy to undercut emotionally because chat interfaces feel general: you can ask about cooking, calculus, and contract clauses in one sitting. Feeling general is not being general. The model is still a narrow (often frontier) system whose training and interface happen to span many text tasks. It does not acquire a new professional competence the way a human transfers from nursing into a related clinic role. It does not form durable, independently verified world models across arbitrary domains. If the team wants a new specialty, they still fine-tune, retrieve, tool-up, or retrain.

For testers, the practical rule is: do not accept a general-intelligence claim as an exemption from task-based testing. If the product is sold as a medical scribe, test it as a medical scribe. If it is sold as a general assistant, still enumerate the assisted tasks and test each one, plus the dangerous combinations (the assistant that both books travel and sends money).

Super AI and the technological singularity

Super AI (artificial superintelligence) is the hypothesized form in which an AI-based system continuously improves itself without needing human intervention or control, and in doing so surpasses both human intelligence and general AI. Many commentators treat that prospect as an existential risk. Testers are not asked to settle that debate. Testers are asked to know the definition and not to apply it to today's recommenders and chatbots.

Access to the Internet is not strictly required for super AI to be conceivable. A superintelligent system isolated from the public network could still, in the hypothesis, self-improve inside its environment. Internet access would significantly expand its capabilities and influence: more data, more actuators via connected systems, more people and institutions it could affect. Isolation is therefore a containment idea, not a proof that super AI is impossible.

The hypothesized point at which AI-based systems would transition from general AI to super AI is commonly called the technological singularity. It is a forecasted discontinuity, not a feature flag in a current release plan. If a stakeholder says we have hit the singularity because a model writes better ads, they are using the term as hype. Super AI presupposes general AI first in this framing, and general AI is not here.

The exam trap: chatbots are still narrow AI

A fluent chatbot is the favorite trick question. Conversation looks like the kitchen-sink of human intellect. It is still narrow AI:

• The core skill is generating and interpreting language (and, for multimodal models, related media) under learned patterns
• Breadth of topics is not the same as independent skill transfer without retraining
• Tool-using agents remain bounded by tools, permissions, and trained policies
• Failures cluster in the usual narrow ways: hallucination, missing context, domain shift, prompt injection, outdated knowledge

If the chatbot is connected to the Internet, that still does not make it super AI or even general AI. It makes it a narrow system with a larger information hose, which changes security and freshness testing, not the capability class.

flowchart TB
  N[Narrow / weak AI: all deployed systems]
  F[Frontier AI: advanced subset of narrow]
  G[General / strong AI: not achieved]
  S[Super AI / ASI: hypothesized]
  N --> F
  N -.-> G
  G --> S

How this changes test strategy without waiting for science fiction

You do not need general AI on the lab bench to use this taxonomy. You need it to stop over-claiming and stop under-scoping:

1. Write capabilities as tasks. Recognizes pedestrian traffic lights at night in rain is testable. Is generally smart is not.
2. Budget retraining as a release. A new country, language, or product line is not a free generalization; it is a new model or a new data job.
3. Treat frontier autonomy as a hazard class. An agent that can email vendors or change prices needs stronger controls than a classifier that only writes a log.
4. Keep science-fiction terms off defect tickets. If a chatbot invents a citation, that is a narrow-system quality failure (grounding, truthfulness, retrieval), not evidence that super AI is misbehaving.

On the exam, if two answers both sound true, prefer the one that keeps today's systems inside narrow AI, treats frontier as advanced narrow, and reserves general and super for systems that do not yet exist. That is not pessimism. It is the vocabulary that lets you pick the right tests.