Settings: language, memory and privacy

Right now ChatGPT answers you the way it answers millions of strangers: politely, at length, and blind. It doesn't know who you are, what you do or how you like your answers. In this lesson we fix that — and work out what happens to the things you write to it, and where the off switch lives.

Tell it about yourself once — and for good

Explaining "I'm an accountant, keep it simple and skip the English jargon" gets old by the third chat. That's what the standing instruction is for: you write a couple of facts about yourself and your preferences once, and every new conversation starts already knowing them.

It's easy to find once you know where to look: the button with your name sits in the bottom-left corner of the window. Click it and pick "Personalization" (the same section also lives under "Settings"). At the top you'll meet "Base style and tone" — the character of the answers picked from a ready-made list — with "Characteristics" like warmth, emoji and headers below. That's the coarse dial. The fine-tuning sits further down: the "Custom instructions" field and the "About you" section — nickname, occupation and a free-form "More about you". A couple of facts about yourself go into "About you", your preferences for answers into "Custom instructions". For example:

I'm a marketer at a small company, based in Berlin.
Answer briefly and to the point.
If a question is ambiguous, ask first.

The difference shows from the very first answer. Before the instruction, "how do I raise sales" gets you a lecture for an abstract businessman; after it, a plan for a small company, written the way you asked. Same AI — but now it's talking to you, not to "a user".

Here are starters for different lives — pick the closest and make it yours:

Student: "Third-year economics. Explain step by step, use real-life examples. Put formulas into words."

Business owner: "I run a children's clothing shop online, on my own. Answer briefly, straight to actions. Count money in euros."

Manager: "I lead a team of 8 in IT. I like structure: lists, tables. Draft emails in a neutral, friendly tone."

Job hunting: "Returning to work after several years away; I used to be an accountant. No jargon, suggest concrete wording. Help me sound confident without overstating."

Retired: "I'm past sixty and technology comes hard to me. Explain simply, no abbreviations, step by step. Help with letters to institutions and paperwork."

Memory: it remembers on its own, and you should know that

Besides the instruction you write, ChatGPT keeps a memory of its own. The feature is called Memory: it picks facts out of conversations — "user has a dog called Rex", "does reports on Fridays" — and uses them later.

Convenient, as long as you remember one thing: it accumulates. Six months in, there's a dozen facts about your life in there, some stale, some that never asked to be kept. So open the list occasionally — same "Personalization" section, under "Memory", behind the "Manage" button; every entry can be viewed and deleted. If it remembered something outdated, you erase that one line, not the whole memory.

The main switch: do your conversations train the model

Now, what happens to your text after you hit send. It goes to OpenAI's servers, the model processes it and returns an answer — that's how any cloud service works, from mail to maps. The real question is different: are your conversations used to train the next versions of the model?

On the free tier — by default, yes. And it's an honest deal as long as you know about it: your dialogues are the fuel that makes the model better. Don't want to take part? The deal is off with one switch: the button with your name → "Settings" → "Data controls", and the very first row there — "Improve the model for everyone".

Why it's on by default and buried two levels deep is a rhetorical question: few people would tick that box voluntarily. But now you know where it is, and the decision is yours.

One clarification so the switch doesn't lull you: it works forward, not backward. It doesn't recall what you sent earlier, and it doesn't erase the chat history in your account — it protects against training, not against storage.

What never to do, whatever the settings

There's a story worth knowing before you upload your first work document.

Spring 2023. Samsung engineers, debugging a program, pasted chunks of the company's internal code and notes from a closed meeting into ChatGPT — just so it would "help fix things". Somebody noticed. In May, Samsung banned public AI tools on work devices, and the headlines went around the world: secrets of one of the planet's biggest companies had left for other people's servers by the hands of its own engineers.

The moral isn't "AI is dangerous" — it's "think about what you paste". Passwords, one-time codes, full card and document numbers, other people's personal data, work secrets: not for a public chat, whatever the switches say. Everything else — letters, drafts, ideas, study — go ahead: that's what the tool is for.

The temporary chat: a conversation without a trace

For one-off sensitive questions there's the temporary chat — the round button with a dashed outline in the top-right corner of the window (hover it and the hint says "Temporary chat"). It stays out of your history, out of memory and out of personalisation: ask, get the answer, close. Handy when you're choosing a present for someone who shares your computer, or working through something personal that doesn't belong in the history.

One caveat so the name doesn't mislead: "temporary" means "not visible in your history", not "vanished from the universe" — the service keeps a copy for a while. Exactly like incognito mode in a browser: it hides things from your computer, not from the internet.

What the AI sees — and what it doesn't

Finally, three fears almost everyone arrives with — and a minute is enough to put them down.

• "It reads my phone" No. The model sees exactly what you typed, dictated or photographed into the chat. No files, no other apps, no microphone in the background — it simply has no access there.
• "It remembers everything forever" Only Memory remembers, and it's transparent. The list opens in the settings, and any line can be deleted.
• "I switched off training, so I'm invisible" The switch keeps your dialogues out of training future models. The history in your account stays until you delete it yourself — those are different things.

Do this now

Five minutes for two decisions. Fill in the standing instruction — two or three sentences about yourself and how to answer. Then open Data controls and set the training switch to the position you chose, not the default one. Finish by asking in the chat: "what do you know about me?" — and watch the instruction at work.