Telling Real AI From Marketing Hype

Artificial Narrow Intelligence (ANI) is every AI system that exists today. It does one thing - or a cluster of related things - extremely well, and nothing else.

A facial recognition system cannot play chess. An AI that plays chess cannot write emails. An AI that writes emails cannot detect fraud. Each is trained for its specific domain and has no ability to transfer skills to a new task without retraining.

The tools this course spends most of its time on - ChatGPT, Claude, Gemini - are one family inside narrow AI, called generative AI, or GenAI when someone is in a hurry. They produce something new, usually text or images or code, rather than sorting or scoring something that already exists. Their cluster of related things is an unusually wide one, which is why they can feel general. They are not. The same tool that drafts your board paper will not read an X-ray or price an insurance policy.

You already rely on narrow AI. The spam filter sorting your inbox and the image recognition tagging faces in your photo library are both narrow AI, and that is exactly why narrow AI carries high business relevance: it is the only kind you can actually buy, deploy and budget for. Two other terms get used in the same breath and neither is a product. Artificial General Intelligence (AGI) would match a person on any cognitive task and transfer knowledge across domains the way people do. Artificial Superintelligence (ASI) would surpass all human abilities in every domain, and is purely theoretical.

This matters more than most people realise. When a vendor says their product "uses AI," they mean ANI. The question is always: what specific task did it learn from what specific data?

The most important distinction in AI is not technical - it is conceptual. Traditional software follows explicit rules written by a programmer: if X, then Y. If a new situation arises that the programmer did not anticipate, the software fails or ignores it.

AI systems learn the rules from data. Instead of being told "spam emails often contain the word 'free'," a spam filter is shown millions of examples of spam and non-spam and discovers its own patterns - including ones no programmer thought to specify.

The diagnostic question for any tool claiming to use AI: does it learn from data, or is it following rules someone wrote? If someone had to explicitly code every decision, it is traditional software. That is not an insult - most of what runs your organisation is exactly that, and for most jobs it is the better tool. The thing worth naming out loud is narrower: the same rule-following software, sold to you as AI, at an AI price.

The AI/rules distinction has three practical consequences for everyone who works with these systems:

• Explainability - traditional software can always explain its decision (follow the code). AI systems often cannot - the pattern exists in millions of numerical weights, not in readable rules.
• Failure modes - traditional software fails in predictable ways. AI systems can fail in unpredictable ways on inputs outside their training distribution.
• Improvement - traditional software improves when engineers rewrite rules. AI systems improve when given more or better data.

This shapes how you audit, maintain, and govern AI tools - very differently from traditional software procurement.

When a vendor, colleague, or news article makes a claim about an AI system, run it through five questions:

1. What specific task was it trained on? "AI-powered" without specifics is not a claim.
2. What data was it trained on? Training data determines what the system can and cannot do.
3. How was it evaluated? Benchmark results mean little without knowing the benchmark's conditions.
4. What are its known failure modes? Every AI system has them. If the vendor can't name any, that's a red flag.
5. Who is responsible for errors? In high-stakes decisions, "the AI got it wrong" is not an acceptable answer.

Certain phrases consistently appear in AI claims that do not survive scrutiny. Treat these as prompts to ask the five questions immediately:

• "General AI" - no general AI exists. This usually means "we have not defined what it does"
• "Learns continuously from every interaction" - worth drawing a careful line here, because half of it is true. A recommender or a spam filter genuinely does keep up with changing behaviour, because someone retrains it on newer examples on a schedule. What almost nothing does is learn from the message you sent thirty seconds ago while you are still typing. Ask which of the two they mean, and how often the retraining actually runs.
• "100% accurate" - no AI system is. Either the benchmark is deceptive or the claim refers to a narrow test case.
• "Like having an extra employee" - AI tools fail silently and have no accountability. This framing dangerously misrepresents the risk.
• "Understands your business" - pattern matching on business data is not understanding. It will fail on situations outside its training distribution.