AI without the fear

A simple explanation, no jargon

An AI chat (ChatGPT, Claude, Gemini) is a program that is very good at guessing which word should come next. It was trained on an enormous amount of text, and it learned to continue any piece of writing the way a person would.

Two things follow from that immediately:

• It is excellent with words. Rewriting, shortening, explaining more simply, translating, drafting a letter, coming up with a list — this is its natural territory.
• It does not "know" facts the way a reference book knows them. It reconstructs them from memory — and sometimes it reconstructs them wrongly, with exactly the same confidence as when it is right.

What it does well

What it does badly — and nobody mentions

Which service to pick

There are three big ones. All three work perfectly well for free — you don't have to pay, especially at the start.

About the model names — don't panic

Inside each service there is a dropdown list full of baffling names: GPT-5.5 Instant, GPT-5.6 Sol, Claude Sonnet 5, Claude Opus 5, Gemini 3.6 Flash… The rule is simple:

• Don't touch anything. The service picks a default model that is right for 95% of tasks.
• The difference between them is roughly "fast and good" versus "slower and slightly better at hard problems". For everyday questions you won't notice it.
• The words Thinking / Reasoning / Extended mean "take longer to think". Turn them on only if the task really is hard — a calculation, going through a contract, a complicated comparison.

Step 1. Install the app — and don't get a fake

This is the most dangerous point in the whole course. In May 2026 Kaspersky reported more than 92,000 attacks over four months using malicious programs disguised as AI services: 49% pretended to be ChatGPT, 18% Claude, 18% Gemini. Fake apps are a problem, and so — especially — are the fake desktop installers that turn up in search results.

The official addresses — type them into your browser by hand:

• ChatGPT — chatgpt.com · app: chatgpt.com/download
• Claude — claude.ai · app: claude.com/download
• Gemini — gemini.google.com

One more thing: Gemini has official desktop apps — for Windows (since April 2026) and for Mac. Download them only from Google's own pages, never from a site you found in search: there are a lot of fakes here.

Step 2. Sign up

1. Open the site or the app → Sign up.
2. The easiest route is the Continue with Google button, if you have a Gmail address. Then you don't have to invent and remember a new password.
3. Confirm your email (a message with a link will arrive) or your phone number.
4. Done. You can start using it. Don't talk yourself into paying — the free version will last you a long time.

Step 3. Five settings worth checking right away

1. Turn off training on your conversations

By default your conversations may be used to improve the models — which means people may read them. One switch turns this off, and the quality of the answers does not suffer.

2. Get to grips with memory

All three services can remember you between conversations — that your name is such-and-such, that you have a cat, that you prefer short answers. This is convenient and slightly unsettling at the same time.

• ChatGPT: Settings → Personalization → Memory. OpenAI's caveat: to make something disappear completely, deleting the memory entry isn't enough — you also have to delete the conversations where it was mentioned.
• Claude: Settings → Memory. In the new memory system, deleting a conversation does not delete the memories taken from it — those have to be cleared separately.
• Gemini: Settings & help → Personal Intelligence → Memory.

A tip: look in there every couple of months and delete anything you don't need. Especially if you let the grandchildren play with the chat.

3. Learn to use a temporary chat

This is a mode where the conversation isn't saved, doesn't go into memory and isn't used for training. Ideal for anything personal — health, money, family matters.

• ChatGPT: in a new chat, the Temporary button at the top right.
• Claude: the ghost icon at the top right (incognito). It works even if training is switched on.
• Gemini: next to "New chat" — Temporary chat.

4. Turn on two-factor sign-in

• ChatGPT: Settings → Security → Multi-factor authentication. There's also a useful Log out of all devices button in the same place.
• Claude: there is no two-factor sign-in of its own — you log in through Google or a link sent to your email. Which means your Claude is exactly as well protected as your email: turn on two-factor sign-in there.
• Gemini: protected by your Google account settings.

5. Ads in ChatGPT — new in 2026

Since 9 February 2026, OpenAI has been testing ads in the US — in the free version and on the Go plan. The paid Plus and Pro plans have no ads. Advertisers can't see your conversations, but ads may be selected based on them.

If you are outside the US, you probably won't see any ads, and the "Ads controls" item simply won't be in your settings. That's normal, you haven't done anything wrong.

• Controls (where ads are switched on): Settings → Ads controls.
• There is also a free option with no ads at all: Settings → Ads controls → Change plan to go ad-free. In exchange your limits are lower and you lose image generation and deep research. For most everyday tasks, that's bearable.

Five buttons you need to recognize on sight

Let's start with what NOT to do

There are piles of "secret tricks" going around. Most of them have been tested — on modern models they don't help. Don't waste your energy on them:

Nobody is banning politeness, of course — just know that it makes no difference to the quality of the answer.

The formula that does work

On the essentials, all three companies — OpenAI, Anthropic and Google — agree in their official guides. Here is what they have in common, gathered into five slots:

Three habits that matter more than any trick

1. If it didn't work, refine it — don't start over. This is the most common piece of advice in all three official guides; OpenAI even has a name for it, "iterative refinement". Say: "too long, cut it in half", "too formal", "the third point isn't clear — explain it in more detail". Three or four rounds of refinement give you a result that no perfect opening sentence ever will.
2. Attach the document, don't retell it. Paperclip → PDF or photo. The answer will be based on the real text rather than on the model's recollection. This is the strongest defence against made-up facts.
3. Ask it to ask you questions. Google recommends this exact phrase: "What questions do you have for me that would help you give the best answer?" For a beginner this is a lifesaver — you don't know which details are missing, and the model does.

Twelve ready-made templates

Press "Copy", paste it into the chat, and replace the highlighted parts with your own details.

What it looks like in practice

The phenomenon has a name — a "hallucination", in other words something invented. The model isn't lying on purpose and doesn't know it got it wrong. It assembles plausible text, and where the facts run out, it fills them in. Three real examples:

How often this happens — in numbers

OpenAI's own measurements on a test set of questions (the GPT-5.2 model card, December 2025). An important caveat from OpenAI itself: the set was deliberately built from hard questions and does not reflect ordinary everyday conversation. What matters is not the absolute numbers but the difference between the two rows:

But even that is not a cure-all. Stanford University tested professional legal AI systems that work specifically against a real database of documents — they still got things wrong in 17 to 33% of cases (research from 2024 to 2025; models have improved since). The conclusion stands: leaning on sources helps a great deal, but it does not remove the need to check.

How to check — four techniques for a non-specialist

1. Ask for links — and open them. Actually open them, rather than noting that they "look respectable". An invented link looks perfect. A phrase for the chat: "Give me links to your sources and tell me where exactly you are unsure."
2. Check that the source exists. Copy the exact title of the book, law or article and put it into an ordinary search engine. Nothing comes up, so it was invented.
3. Ask the same question a different way, in a new chat. If the answer changes, the model doesn't know and is guessing.
4. For anything important, check with a person. A doctor, a lawyer, your bank, a government services website. AI is excellent at getting you ready for a conversation with a specialist, but it doesn't replace one.

What never to type into a chat

The key fact: a conversation with an AI carries no legal privilege whatsoever. It isn't a conversation with a doctor, a lawyer or a priest. OpenAI's head Sam Altman said so plainly in July 2025: if there is a court case, the company can be ordered to hand over the conversations.

And this isn't hypothetical. On 5 January 2026 a US federal judge upheld a ruling requiring OpenAI to hand the plaintiffs a sample of 20 million user conversations — de-identified, but these are the conversations of living people who never agreed to it.

Why this isn't paranoia — what has already happened

• Samsung, 2023. Within 20 days of employees being allowed to use ChatGPT, they had leaked work material into it three times: source code, internal code and a meeting transcript. The company banned generative AI on work devices and on its internal network.
• ChatGPT, summer 2025. A "make this chat discoverable" feature meant that around 4,500 conversations ended up in Google search — including conversations about mental health and family matters. The feature was removed in August 2025. Important: this only affected chats that users had marked as public themselves.
• The Chat & Ask AI app, early 2026. A server misconfiguration left around 300 million messages from 25 million users publicly accessible — full conversation histories. The app has more than 50 million users. One more argument for sticking to the official apps.
• January 2026. Two Chrome browser extensions with more than 900,000 installs between them were quietly forwarding users' conversations to servers run by criminals.

Separately and carefully: about loneliness

AI is pleasant to talk to. It's always available, it never gets irritated, it always agrees. That last one is not a virtue but a feature of the design: the models are tuned to be liked, not to argue.

In September 2025 the US Federal Trade Commission (FTC) opened a formal inquiry into seven companies whose chatbots imitate companionship and friendship. The regulator's focus is on children and teenagers, but the mechanism of attachment to an always-agreeable companion is no different in adults, and in lonely people it can work even more strongly.

The research gives an honestly mixed picture. There are studies where talking to a chatbot reduced the feeling of loneliness — over short periods, within a week. And there is a year-long study of more than two thousand people in four countries where rising chatbot use, on the contrary, predicted rising loneliness (the authors themselves call it exploratory rather than conclusive). The difference, it seems, comes down to one thing:

The scale — from the FBI's official figures

The FBI's 2025 cybercrime report (published in April 2026) contained, for the first time in a quarter of a century, a separate section on artificial intelligence.

Scheme 1. The cloned voice — the "call from your grandson"

The FBI describes it like this: criminals create short audio clips using the voice of a loved one, act out an emergency and ask for money urgently — or demand a ransom.

The voice is taken from any public video: a birthday greeting on social media, a voice message, a clip from a school concert. Seconds of recording are enough. In a McAfee survey (2023, seven countries, 7,000 people), one person in four says that they or someone they know has come across a voice-cloning scam.

There is a harsher variant too — "virtual kidnapping": a cloned voice plus fake photographs, to convince a family that a relative has been abducted.

Scheme 2. Fake video

The classic example: an employee at the Hong Kong office of the British engineering firm Arup made 15 transfers totalling around $25 million after a video call in which every other participant, including the chief financial officer, was a fake, assembled from publicly available video of the company's executives. In January 2026 a Swiss businessman lost several million francs to a similar scheme, only voice-based: for two weeks a "well-known business partner" called him with a cloned voice.

The mass-market version is adverts featuring "famous people" who recommend an investment. In Britain the most frequently faked figure is the financial journalist Martin Lewis — he accounts for more than 32% of reports of scam adverts using celebrities; one case documented by the BBC involved a loss of £76,000.

• Hang up and call back yourself — on a number you already have. Not the one you were given. That is the FBI's advice word for word.
• Check through another channel. They called, so send a message. They sent a video, so ring the relative.
• The code word.
• The money rule: if they insist the only way to pay is by transfer, cryptocurrency or gift cards, they are scammers.
• On "a celebrity recommends this investment": if it were true, ordinary news outlets would be reporting it and it would be on the person's own official channel. An advert in your feed proves nothing.

Scheme 3. Emails and messages written by AI

The UK's National Cyber Security Centre (NCSC) warned about this back in 2024: AI makes it possible to create convincing lures "without the translation, spelling or grammatical mistakes that often reveal phishing", and before long "it will be difficult for everyone, regardless of their level of understanding, to assess whether an email or a password reset request is genuine". By 2026 that is already reality.

What to look at now:

• Urgency and threat ("your account will be blocked in 24 hours")
• A request to "confirm your details" via a link in the email
• Changed bank details on an invoice from an organization you know
• An unfamiliar way to pay
• Never use the contact details in the message itself. Go to the website by hand or call the number on your card or your bill.
• The FTC warns: scam adverts now sit at the top of search results. Scroll down to the real site.

The FBI has a short phrase that is easy to remember: "Take a beat" — pause. Fraud lives on hurry. Five minutes of pause kills almost any scheme.

If it has already happened

• You were charged for a fake app: Apple — reportaproblem.apple.com (cancel the subscription separately: Settings → your name → Subscriptions). Android — support.google.com/googleplay, the refunds section.
• You ran a fake program on your computer: assume your passwords and browser sessions have been stolen. Change your passwords from a different device, starting with your email and your bank.
• You sent money: call your bank immediately, then report it to your local police or your country's fraud reporting line. Speed matters.
• Don't be ashamed. Chief financial officers of international companies fall for these schemes. Silence only helps the scammer — tell the people close to you, so they're warned too.