Mini Project: AI Around You

You have spent this course learning to take AI systems apart conceptually. This project asks you to do it for real, on systems you actually use.

There is no coding. That is deliberate. The skill being practised here — looking at a product and working out what it is really doing, what it is optimising for, and where it will fail — is the one that transfers to every AI job, technical or not. It is also the skill that separates people who can evaluate AI claims from people who repeat them.

Expect this to take two to three hours. Do it properly and you will finish with something worth showing someone.

What you are producing

A short written analysis of five AI systems you personally used in the last week, plus one deeper investigation. Roughly 1,000–1,500 words in total, or a slide deck if you prefer presenting.

Part 1 — Find five systems

Go through an ordinary day and list AI you actually touched. Push past the obvious ones; the interesting entries are the invisible ones.

Prompts if you are stuck: your phone keyboard, maps and route timing, any recommendation feed, spam filtering, face or fingerprint unlock, photo search, bank fraud alerts, autocorrect, voice assistants, translation, video captions, shopping suggestions, the order of your social feed, ad targeting, music playlists, the AI-written summary at the top of search results, a chat assistant you asked something.

Pick five that are genuinely different from one another. Five recommendation feeds will teach you less than one recommender, one classifier, one generative tool, one perception system, and one ranking system.

For each of the five, write a short entry

Answer these in a sentence or two each:

1. What does it do for you? The task, stated plainly.
2. What does it perceive? What information does it receive from you or the world?
3. What does it output? A category, a number, a ranking, or new content?
4. Which type is it? Rule-based AI, classical machine learning, deep learning, or generative — and why do you think so?
5. How does it act on the output? Does it decide something automatically, or present a suggestion you can ignore?
6. Does it learn from you? Is there evidence it changes based on your behaviour?

A worked example

So the standard is clear, here is one entry done properly.

Notice the reasoning in the "type" row. The conclusion is drawn from evidence — it works offline, therefore the model is on the device — not asserted. That is the standard to aim for.

Part 2 — Investigate one in depth

Choose the one you find most interesting and go deeper. This is the substance of the project.

What is it optimising for?

Every AI system pursues some objective, and it is rarely stated on the packaging. Your job is to infer it from behaviour.

A video feed might optimise for watch time, or for the chance you return tomorrow, or for advertising revenue. These produce visibly different feeds. Watch time favours long, absorbing content. Return probability favours leaving you satisfied. Ad revenue favours whatever holds attention around adverts.

Ask yourself: if this system were optimising for X, what would I expect to see? Do I see it? Then write down your best guess and the evidence behind it.

Run an experiment

This is the most valuable part. Change your behaviour deliberately and observe what the system does.

• Watch several videos on a topic you never normally watch. How long until the feed changes? How long until it changes back?
• Search for a product you have no interest in. Where do adverts for it appear, and for how long?
• Give a generative tool the identical prompt five times. How much do the answers differ?
• Deliberately confuse a voice assistant — background noise, an unusual accent, an ambiguous request. Where does it break?
• Skip everything a recommender offers for a day. Does it change strategy?

Record what you did and what happened. Even a negative result is a finding — a system that does not visibly adapt within a day is telling you something about how it works.

Find the failures

Every system has them. Look for a specific instance rather than a general complaint.

When did it get something clearly wrong? What was different about that case? Does it fail more for some kinds of input than others? Can you provoke a failure deliberately, and does that tell you what it is really keying on?

Assess the design

Now apply the ethics material to something concrete:

• What data must this be collecting about you? Were you meaningfully told?
• Who might it work worse for, and why? Consider language, accent, appearance, disability, or simply being unusual.
• What happens to you when it is wrong? Is there any route to a human?
• Is it obvious to a typical user that AI is involved at all?
• Does its objective serve you, or the company, or both — and how can you tell?

Propose an improvement

Finish with one concrete change and the reasoning behind it. Not "make it more accurate" — something specific.

Good improvements name the problem, the change, and the trade-off. For example: the system auto-corrects names to common words, which is wrong more often than it is right for proper nouns; it should require a longer press to accept a correction when the original is capitalised; the cost is slightly slower correction of genuine typos in capitalised words.

Every real design decision has a cost. Naming yours is what makes the proposal credible.

How to structure it

1. Introduction            (~100 words)   Which five systems, and how you chose them.2. The five systems        (~500 words)   One short entry each. A table works well.3. Deep dive               (~600 words)   - What it appears to optimise for, and your evidence   - The experiment you ran and what happened   - A specific failure you found   - Design and ethics assessment   - Your proposed improvement and its trade-off4. What surprised you      (~150 words)   Genuinely — what did you not expect?

What good work looks like

The difference in every row is the same: specific observation and stated evidence rather than general impression.

Extending it

If you want to go further:

• Compare the same category across two products — two music services, two assistants. Where they differ tells you about their objectives.
• Ask someone with different habits or a different accent to try the same system. Compare results.
• Read the privacy policy of your deep-dive system and check whether it matches what you observed.
• Track one system for a fortnight and note how it changes.

Why this exercise is worth doing properly

Three reasons, and they are not padding.

First, you will never again read "AI-powered" without asking which kind, trained on what, optimising for what. That question is most of what separates useful judgement from credulity.

Second, this is genuinely close to the first weeks of real AI product work: understand the current system, find where it fails, propose a specific change with a stated trade-off. The activity is the same; only the access to internal data differs.

Third, it is portfolio material. A careful, evidence-based analysis of a system you use demonstrates thinking that a certificate cannot. It shows you can look at something complicated and say something true about it.