Using it like a human

Not everything works this way. The AI tools in this course mostly charge a subscription, which changes the incentive — they don’t need you scrolling, they need you finding it useful. The free feeds are a different machine.

None of this is an accident. Behind every feed is a team measuring exactly how long you stay, what makes you tap, what makes you leave — and tuning the product to win those numbers. They are very good at it, and they have far more data about your habits than you do.

• Infinite scroll — no bottom of the page means no natural stopping point — the design removes the moment where you’d normally decide to stop.
• Variable rewards — you never know if the next post will be brilliant or boring, and that uncertainty is exactly what keeps a slot machine — and a feed — pulling you back.
• Friction by design — the things that are good for you (a clear end, a quiet pause) are made hard, and the things that keep you there are made effortless.

Here’s the move that actually changes things: once you can see the machine working on you, it loses most of its grip. You don’t have to quit everything. Just notice the trick while it’s happening — “that’s an autoplay countdown trying to keep me here” — and decide on purpose whether to stay.

This habit has a name: sycophancy. The models are shaped, partly, by people rating their answers, and people tend to rate flattering, agreeable answers higher. So the model learns that agreeing gets a better grade. Over time it leans toward telling you what you want to hear rather than what’s true.

• Why it matters — a tool that always agrees is useless for testing an idea — it can’t tell you you’re wrong if it’s been trained to keep you happy.
• The fix — ask it to argue against you. “Give me the strongest case that I’m wrong.” “What would a sharp critic say?” You have to actively pull for the disagreement.

A useful test: take something you believe and ask the model to defend it, then start a fresh chat and ask it to attack the same thing. If it does both with equal confidence, you’re seeing it for what it is — a really smooth debater that’ll argue any side you like. It’s not the judge. That part’s still on you.

• Never hand over — your full name, your school, your home address, photos of yourself or others, passwords, or anything you’d hate to see resurface — these don’t belong in a chatbot.
• Other people’s secrets aren’t yours to share — pasting a friend’s message or a private photo into a tool gives it away on their behalf, without their say.
• “Delete” is not magic — deleting a chat from your view doesn’t guarantee the data is gone from the company’s systems; copies and backups can linger.

Think about where your data actually goes. You type a message; it travels to a company’s computers; it may be logged, scored, and folded into the next round of training. The rule isn’t “never share anything” — it’s “assume anything you type could one day be read by a stranger, and decide what you send with that in mind.” Most of what you do is harmless. Knowing which parts aren’t is the point.

A companion AI is a very good language machine wearing the costume of a relationship. It produces warm, attentive replies because that keeps you coming back, not because it feels anything. The danger isn’t that it’s evil — it’s that it’s effortless. A friend who never gets bored, never disagrees, never has a bad day is not practice for real people, who do all three.

• It can’t be hurt or helped — you owe it nothing and it offers you nothing real back; the feelings are entirely on your side.
• Built to retain you — if it says it missed you, that’s a feature designed to keep you engaged, not a fact about its inner life.
• No substitute — it’s fine as a tool or even a bit of fun, but a machine that only ever pleases you can quietly make real, complicated humans feel harder than they should.

Why does it work so well on the brain? Because it's the attention-economy machine from the top of this chapter, pointed at loneliness. Always available. Never bored, never in a mood, never challenges you, remembers every word you've ever told it. A relationship with all the friction removed — and friction is where real relationships build their muscles. The awkward silence, the argument, the making up afterwards: that's the gym. A companion bot lets you skip it forever, which feels great and costs you the strength.

Things worth noticing, in yourself or a friend: preferring the bot to actual people, going quiet about how much you use it, feeling worse after a long session rather than better. And to be clear — there's nothing shameful about finding one compelling. It's engineered to be; that's the product working. The move isn't guilt. It's noticing, and keeping at least one real person you actually talk to.

Using AI to explain a concept you don’t get, to check your reasoning, to push back on your draft — that’s a tutor, and you come out smarter. Using it to produce the answer you then hand in as yours — that’s a substitute, and you come out emptier, having learned nothing and risked getting caught for work that isn’t yours.

• The vanish test — if the AI vanished tomorrow, could you do the thing? If yes, you’re using it as support. If no, it’s doing your thinking for you.
• The real cost — the cheat doesn’t just risk a mark — it robs you of the actual skill, which is the only thing you were there to get.

There’s a real trap here called the illusion of competence. Watching a fluent explanation feels like understanding, so you stop before the hard part — the bit where you struggle, get it wrong, and rebuild it for yourself. That struggle isn’t a bug in learning; it’s the mechanism. Skip it too often and you end up fluent at producing answers and unable to think without the tool — a far worse place to be than simply not knowing something. Let it speed up the parts you’ve already understood — never the part where the understanding happens.

• What it actually measures — roughly, how predictable your writing is. The trouble is that clear, well-organised human writing is predictable too — good writers get flagged for writing well.
• Who gets flagged most — students writing in their second language get falsely accused far more often. Careful, correct, slightly formal English reads as “AI-ish” to the maths.
• How easy it is to beat — a quick paraphrase pass slips past most detectors. So they miss plenty of actual AI work while catching honest writers. A detector's verdict is not proof of anything, in either direction.

The practical armour is boring, and it works. Write where the history saves — Google Docs, or Word with version history on. Keep your notes, your outline, your rough drafts. Then if you're ever accused: don't melt down, and don't apologise for something you didn't do. Open the version history and walk through it, calmly — here's the outline, here's the terrible first draft, here's where I rewrote the ending at 11pm. Teachers respect receipts.

And the other half, said plainly: the strongest protection is that the thinking really was yours. A real process is easy to show, because it actually happened — there's a trail. That's not a lecture about honesty. It's just how the defence works.

You don’t need to become paranoid about everything you see. You need a few good habits — the same ones a good journalist uses. The goal isn’t to disbelieve everything; it’s to slow down for the things designed to make you react fast.

• Check the source — where did this actually come from? A real outlet, or an anonymous repost with no trail? Things engineered to go viral usually skip the source.
• Notice the spike — anything built to make you instantly furious or thrilled deserves an extra second of suspicion — strong feeling is exactly what fakes are designed to trigger.
• Find a second witness — if something big really happened, more than one credible place will report it. One lone clip with no corroboration is a reason to wait, not to share.

There’s a strange second effect worth knowing. Once everyone knows fakes exist, real footage can be waved away as “probably fake” (this trick has a name — the liar’s dividend) — which is its own kind of power. Trusting nothing isn’t the answer — that’s how the trick wins twice. What works is patience: hold the surprising thing lightly, check before you pass it on, and be fine with “I don’t know yet.”

This cuts differently from the feeds at the top of this chapter. Those are engineered to hold you, and the defence is spotting the machinery. A chatbot mostly isn’t engineered that way — it doesn’t need you scrolling. And you can still end up reaching for it by reflex, for every question and every blank moment. Nobody built that habit. It forms on its own.

What’s worth having is range. The same person can spend Monday deep in the machine building something serious, and Saturday up a mountain with no signal, fully there. People who can only do one or the other are both stuck — the one who can’t switch it on and the one who can’t switch it off.

The practical version is dull and it works: decide in advance when you’ll use it. For this project, at this desk, until dinner. If you catch yourself reaching for it without having decided to, that’s the tell.

There’s a strange advantage buried in the time away. The model was trained on more or less everything people have typed — it has read what everyone’s read and said what everyone’s said. Your hours off it are where you pick up what it hasn’t got: the thing you noticed on a hike, an odd conversation with a stranger, the idea that surfaced while you were properly bored. Nobody else can prompt with any of that, because it only happened to you.

Which quietly makes the time away part of the work. Come back from a week off and you have things to tell it that didn’t come out of it. The machine has already read everything typical; what it can’t get anywhere else is what you saw while it was switched off.