Artificial Intelligence

Machines That Seem to Think

Artificial intelligence (AI) is the broad name for computer systems that perform tasks normally thought to need human intelligence: recognising faces in photos, understanding spoken language, recommending a song you'll like, translating between languages, playing chess at a superhuman level. AI is not one single technology — it is a goal, pursued by many different techniques, and today's most talked-about approach is machine learning.

Two Very Different Ways to Build a Solution

Every algorithm so far in this course has been built the traditional way: a programmer works out the exact rules and writes them, step by step, as pseudocode or code. IF age >= 13 THEN allow — a human decided that rule, completely, in advance.

Machine learning flips this around. Instead of a human writing the rules, the system is shown a huge number of examples — and it searches for patterns that predict the right answer, adjusting itself as it goes. Want to recognise cats in photos? The traditional approach would try to hand-write rules about whiskers, ears, and shapes — famously difficult, because "cat-ness" resists being pinned down in rigid IF statements. The machine-learning approach instead shows the system thousands of photos labelled "cat" or "not cat", and lets it discover, statistically, what patterns of pixels tend to separate the two groups. Nobody writes "IF whiskers THEN cat" — the system arrives at something that works, through exposure to data, not explicit instruction.

This is why the field is often summarised as learning from data rather than being told rules — a genuinely different relationship between programmer, program, and problem than anything else in this course.

Data Is the Fuel — and the Risk

Because machine learning learns from examples, its behaviour is shaped entirely by the data it was trained on. This has a serious consequence: a system only ever learns what its training data actually shows it.

If a face-recognition system is trained mostly on photos of one group of people, it can perform noticeably worse on people it saw few examples of — not because of any deliberate intent, but simply because the patterns it learned were drawn from an unrepresentative sample. This documented problem is called bias, and it is one of the most important, actively studied issues in AI today — a direct consequence of the "learn from data" approach, and a reason organisations building these systems must think hard about what data they use, and who it represents. It connects directly to the wider questions in ethics and technology: who is affected by a system, and who was consulted while building it?

What AI Is Good At — and Where It Struggles

AI built this way is genuinely powerful at pattern recognition at scale: spotting patterns across millions of examples faster and more consistently than any human could. Voice assistants, translation tools, spam filters, and recommendation systems all lean on this strength daily.

But two limitations deserve equal weight, not an afterthought:

• It optimises for "plausible", not "true". Text-generating AI systems are built to produce statistically likely continuations of text — which is not the same guarantee as producing verified facts. This is why such tools can state confident-sounding falsehoods (sometimes called "hallucinations") — the system isn't lying on purpose, it is doing exactly what it was built to do, which is not the same thing as fact-checking.
• Whether it truly "understands" is genuinely disputed. Finding statistical patterns extremely well is not obviously the same thing as human reasoning or comprehension — researchers and philosophers actively debate this, and a healthy habit is treating confident-sounding AI claims the way you would treat an unverified source: useful, but worth checking, especially for anything important.

Using AI Well

None of this means avoiding AI — it means using it like any other powerful tool, with clear eyes about its shape. Treat its outputs the way testing a program treats any result: worth verifying, not worth blind trust, especially on the boundaries — unusual cases, underrepresented groups, and high-stakes decisions are exactly where machine-learning systems are most likely to show their limits.

Try It Yourself

Pick an AI tool you have used (a voice assistant, a recommendation feed, a chatbot, a translation app). For it: what examples do you think it was likely trained on? Where might its training data have gaps? And — the crucial habit — how would you check whether one of its outputs was actually correct, rather than just plausible?

Worked Example — Traditional Programming vs Machine Learning on the Same Task

Comparing both approaches on one concrete task makes the difference from earlier in this lesson far more vivid. Task: decide whether an email is spam.

Traditional approach — a programmer writes explicit rules by hand:

IF email contains "WIN FREE MONEY" THEN mark as spam
IF sender is not in contacts AND contains a link THEN mark as spam
IF ALL CAPS SUBJECT LINE THEN increase suspicion score

This works reasonably well at first, but spammers constantly invent new wording, and the programmer must keep manually adding new rules to keep up — an endless, reactive arms race, similar to a test plan that can never quite cover every case in advance.

Machine learning approach — instead of hand-writing rules, the system is shown a huge number of real emails, each already labelled "spam" or "not spam" by humans, and it learns statistical patterns from the examples: certain word combinations, sender patterns, and structures that tend to correlate with spam — without a programmer ever writing "IF contains X THEN spam" directly.

The genuine trade-off this comparison reveals: the traditional approach is fully transparent (you can read every rule and know exactly why an email was flagged) but brittle against new tricks; the machine-learning approach adapts more readily to spam styles nobody anticipated, but its internal reasoning is far harder for a human to inspect directly — you can observe that it flags an email, but not always easily why, in the same clear-cut way an IF statement explains itself. Neither approach is simply "better" — they represent genuinely different trade-offs between transparency and adaptability.

Key Words

• Artificial intelligence (AI) — systems performing tasks normally requiring human intelligence
• Machine learning — a system learning patterns from example data, rather than following explicit hand-written rules
• Training data — the examples a machine-learning system learns from
• Bias — systematically worse performance caused by gaps or imbalance in training data
• Hallucination — confident-sounding but false output from a text-generating AI system