A Realistic View of AI

Narrow and general AI

Everything discussed in this course is narrow AI: systems that do one task, or a set of related tasks. A model that reads chest X-rays cannot route your support tickets. A remarkable language model cannot drive a car.

General AI — a system with the broad, flexible competence of a human across arbitrary tasks — does not exist. Serious researchers disagree profoundly about whether and when it might, and anyone offering you a confident date is telling you about their convictions rather than about evidence.

The distinction matters for one practical reason: almost all public discussion of AI risk is about general AI, and almost all actual AI risk in your organisation is from narrow AI. The realistic risks are a biased hiring filter, a model that quietly stops working, a chatbot confidently telling a customer something false, a system nobody can explain to a regulator. These are unglamorous, they are happening now, and they are the ones you can do something about.

The pattern of hype and winter

AI has run through this cycle repeatedly since the 1950s.

📅 Timeline

1950s–60sEarly successes with logic and games produce confident predictions that human-level machine intelligence is roughly two decades away.

1970sProgress stalls against problems far harder than expected. Funding is withdrawn. The first AI winter.

1980sExpert systems — hand-coded rules capturing specialist knowledge — become a commercial boom.

Late 1980s–90sExpert systems prove brittle and expensive to maintain. The second AI winter.

2012 onwardsDeep learning delivers genuine, repeated breakthroughs in vision, speech and language. Real capability, real products, and a fresh cycle of extraordinary claims.

The current wave is different in one important respect: it is producing things that visibly work and that ordinary people use daily. This is not the 1980s.

It is also the same in one important respect: the gap between demonstrated capability and claimed capability is enormous, and it is where most disappointment comes from.

Reading claims sensibly

A few habits that hold up well.

Separate the demonstration from the deployment. "Researchers achieved X" and "X is available to buy" are years apart, and many things in the first category never reach the second.

Ask what the benchmark was. "Outperforms doctors at diagnosis" almost always means on a specific dataset, on a specific condition, under conditions unlike a real clinic. That can still be genuinely valuable; it is not the same claim.

Notice who benefits from the claim. Vendors, researchers seeking funding, and journalists competing for attention all face pressure toward overstatement. This does not make claims false, but it should set your prior.

Watch for the shift from "can" to "will". "This technique can reduce errors in controlled tests" quietly becomes "AI will eliminate errors". The first is a finding; the second is a forecast dressed as one.

Discount confident timelines heavily. The field's own track record on predicting its progress is poor in both directions — some things arrived far sooner than expected, others are still twenty years away and have been for fifty years.

Two postures dominate organisational conversations about AI, and both damage decisions.

The uncritical optimist believes AI will transform everything shortly, so the priority is to adopt fast and broadly. In practice this produces a portfolio of pilots, none integrated, chosen for how impressive they sound rather than whether the data exists. When they do not deliver, the response is usually to blame the vendor and try another. Money is spent, capability is not built.

The dismissive sceptic has seen technology waves before and treats this as another. Their scepticism is often locally correct — the specific vendor claim in front of them is usually overstated. But applied uniformly it means missing the genuine change: some tasks that were expensive are now cheap, and competitors acting on that get a real advantage.

The productive position is neither, and it is not a compromise between them. It is specific. Instead of asking "how transformative is AI", ask about one task at a time: could a person do this in about a second, do labelled examples exist, what would change if it were automated, and what happens when it is wrong. Those questions produce different answers for different tasks, which is exactly right, because the technology genuinely is transformative for some tasks and useless for others.

The person who can hold both of those thoughts at once — this is real, and most claims about it are overstated — is the person you want deciding your AI investments.

❓ Knowledge Check

Why does the narrow-versus-general distinction matter for an organisation?

📚 Flashcards1 / 5

Term

Narrow AI

Click to flip

Definition

Systems that perform one task or a set of related tasks. Everything in commercial use today, and the source of essentially all real organisational AI risk.

Click to flip back

💡Key Takeaway

Everything in use today is narrow AI, and that is where your real risks live — not in the general-intelligence debate that dominates coverage. The field has cycled through hype and winter since the 1950s, and while this wave is genuinely producing working products, the gap between demonstrated and claimed capability remains where disappointment comes from. Neither enthusiasm nor dismissal helps; specificity does.