What is AI, and how do I start using it safely?

The AI people talk about today is mostly a computer program you talk to in ordinary words: you type or say what you want, and it writes an answer, makes a picture or, more and more often, carries out a task on a website for you. ChatGPT, Gemini, Claude, Copilot and Meta AI are examples. To start safely, try one of the free versions on something you can check yourself, keep passwords, health details and other people’s information out of it, and look at the setting that lets the company use your chats to improve its AI. Treat every answer as a draft from a fast helper that is sometimes wrong while sounding sure: check anything that matters with a person or an official source.

This guide answers the questions people ask most, in thirteen short chapters. Each one gives the answer in a few sentences and, where the site has one, links the page that explains it in detail. Every fact was checked on the makers’ and authorities’ own pages on 31 October 2026. This is a living page: it is checked again every month, because plans, prices and settings change often.

What is the AI everyone is talking about?

It is a family of programs that learn from examples instead of following rules written one by one, a method called machine learning. Most people meet three kinds:

• Assistants, or chatbots, such as ChatGPT or Gemini: you write a question, they write an answer. The program behind them is a large language model, a program trained on huge amounts of text to write new text.
• Picture, video and voice makers: they create new images, clips or voices from a written request. Together with the assistants, they are called generative AI, AI that makes new content.
• Agents: an AI agent does not stop at answering. It clicks, types and fills in forms on a computer or a website for you.

What it is not: it is not a person, it does not know you unless you tell it, and it is not a library of checked facts. It writes what fits your request best, which is why it can sound certain and still be wrong (see “What does AI do well, and where does it go wrong?” below).

Which AI assistants can you use?

Nine well-known assistants are compared on this site, in alphabetical order. This site is written with AI tools, including Claude by Anthropic, one of the assistants compared here. The comparison uses only the makers’ published information, with links.

Assistant	Cheapest paid plan, as the maker showed it on 31 October 2026	Minimum age in the maker’s terms
ChatGPT	Go, 8 euros a month (page seen from Italy)	13, with a parent’s permission under 18
Claude	Pro, 17 dollars a month paid yearly or 20 dollars paid monthly, tax excluded	18
DeepSeek	No paid plan for individuals published	“Primarily intended for adults”; under 18 with a guardian’s consent
Gemini	Google AI Plus, 4.99 euros a month	13, or the age set in your country
Grok	Plus, 8 euros a month, tax excluded (page seen from Italy)	13; from 13 to 17 with a parent’s permission
Meta AI	Meta One Core, 7.99 dollars a month as announced in September 2026; no euro price found	13, or the higher age set in your country
Microsoft Copilot	Microsoft 365 Personal, 10 euros a month or 99 euros a year	13, sometimes older depending on your country
Mistral Vibe (Le Chat)	Pro, 14.99 a month in dollars or euros, tax excluded	13, with a parent’s permission where required
Perplexity	Pro: price not verified, because the Pro page shows it only in a browser; Education Pro, for verified students and educators, 10 dollars a month	13; a parent or guardian must accept the terms for a minor

Prices are copied as each page shows them: dollars are not converted into euros, because the amount you pay can differ, and some prices exclude tax. They are not directly comparable either, since some plans include storage or other services.

There is no “best” assistant: the right one depends on your budget, your age or your child’s, the language you write in and how you want your chats handled. AI assistants compared answers by profile (student, parent, office worker, small business, retired person, teacher) and adds apps, languages and data settings for all nine. Which AI do I need? asks you six questions and sorts the nine into three groups by your answers.

Is AI free, or should you pay?

You can start for free. All nine assistants above can be tried without paying, and most of them limit the free version. OpenAI, for example, says free ChatGPT users get “unlimited everyday text chats, subject to abuse-prevention safeguards”, while pictures, file uploads and voice have separate limits it does not give in numbers. On 31 October 2026 OpenAI’s pricing page, seen from Italy, listed Go at 8 euros a month, Plus at 23 euros and Pro from 103 euros. OpenAI’s ads help page says ads may appear on Free and Go, not on Plus or Pro.

Paying makes sense if you keep reaching a limit, or need something only a paid plan includes. It does not make the answers more reliable: OpenAI’s terms of use say that “Output may not always be accurate”, on every plan. Is ChatGPT free? explains each ChatGPT plan, how to cancel, and what the free versions of the other assistants offer.

How do you write a good first request?

Write it the way you would explain the job to a helpful stranger: say what you want, give the background, and say what the result should look like. Microsoft’s guide to Copilot prompts names four parts (goal, context, source and expectations) and says only a clear goal is required. Google’s guide for Gemini suggests the same idea in other words (persona, task, context, format) and advises writing as if you were talking to another person. A request like this is called a prompt.

Three examples, two from home and one from the office:

1. A message. “Write a short, polite message to my building manager. The light on the second-floor stairs has been broken for two weeks and I would like it fixed this week. Under 80 words.”
2. A meal. “I have chicken, potatoes, a courgette and two lemons. Suggest two dinners for two people that take less than 30 minutes, and list anything I need to buy.”
3. A work email. “Here are my notes from today’s meeting. Turn them into an email to my team, with a list of who does what and by when.” Before pasting, replace names and anything confidential with made-up ones.

If the answer is not right, keep the conversation going: “shorter”, “more formal”, “in Italian”, “explain it as if I were twelve”. Both guides treat this back-and-forth as normal. Microsoft’s guide also says the same request can give a different answer each time, and that you should check what you get.

What does AI do well, and where does it go wrong?

AI is good at first drafts, summaries, translations, explanations in simpler words and finding information to start from. In a study of 200,000 conversations with Microsoft’s Bing Copilot, Microsoft Research found that people mostly asked for help finding information and writing.

Its main weakness is the hallucination: an answer the AI makes up that sounds true. A paper by researchers at OpenAI and Georgia Tech, dated 4 September 2025, explains why. These programs are trained and tested in ways that “reward guessing over acknowledging uncertainty”, like students who guess on a hard exam because a blank answer scores nothing. In one of their examples, a language model asked for the birthday of one of the authors, and told to answer only if it knew, gave three different wrong dates in three tries. The researchers propose changing how AI is tested, so that admitting doubt is no longer penalised.

How to protect yourself:

• Check the facts that matter: names, numbers, dates, quotes, laws, prices. Ask for the sources and open them; a link the AI gives may not say what the AI claims.
• Use it as a helper, not as the expert: OpenAI’s terms say you should not rely on its answers “as a substitute for professional advice”. Health, money and legal questions go to a doctor, a bank or a lawyer, and to official websites.
• If you are going through a hard time, talk to a person: a doctor, a family member, or a helpline in your country.

What happens to what you type?

Your message goes to the company’s computers, where the answer is written, and is usually saved in your account. For seven of the nine assistants, the makers’ pages say chats are used to improve their AI unless you switch it off or object: ChatGPT, Claude, DeepSeek, Gemini (from age 18), Meta AI, Mistral Vibe and Perplexity. For Microsoft Copilot, only the pages of the older app say so, and those of the new app, available since 18 August 2026, do not. xAI, the maker of Grok, describes a way to say no without stating the default. This process of teaching the program with examples is called training.

Three settings are worth checking in whichever assistant you use:

1. The training switch. In ChatGPT it is called “Improve the model for everyone”, under Settings, then Data controls; in Gemini it is Keep Activity. In ChatGPT, switching it off stops future use but does not delete your saved chats.
2. Temporary or private chats. Several assistants offer a chat that is not saved in your history. OpenAI says a ChatGPT Temporary chat is not used for training, though a copy may be kept for up to 30 days for safety.
3. Deleting chats. A deleted chat leaves your list at once but not always the company’s systems: OpenAI says a deleted ChatGPT chat is scheduled for permanent deletion within 30 days, with exceptions for security and legal reasons.

Whatever the settings, never paste passwords, PINs or codes, health records with your name on them, identity documents, other people’s personal details or your employer’s secrets. In the EU you have the right to see, delete and object to the use of your data, and in Italy you can complain to the Garante, the data protection authority. What happens to what you type gives the setting, the menu path and the deletion period for each of the nine assistants.

How do you make pictures with AI, and how do you spot a fake?

Making pictures is often included even in free versions: OpenAI, for example, lists picture creation among the free ChatGPT features, with its own limit. Take care with pictures of real people. Making a realistic fake of someone can hurt them, and if you do it in your work, EU rules require you to say it is artificial (see “What rules protect you?” below).

Spotting a fake is harder than it looks, and no method is certain. In a 2022 study, people asked to sort real faces from computer-made ones got 48.2% right, about as good as tossing a coin. The method that works best goes in three steps: first the context (who posted it, and whether reliable sources report it), then any label or hidden marker in the file, and only last the details of the picture itself, which are weak hints at most. Google’s Gemini app can check for SynthID, the invisible mark Google adds to what its own AI makes, but Google says it “can currently only recognize content created by Google AI tools”. A missing label proves nothing.

For voices, the US Federal Trade Commission’s advice fits: “Don’t trust the voice.” If a call or voice message from someone you know asks for money, call them back on a number you know is theirs. How can you tell if a photo or video was made with AI? explains the three steps, the labels on YouTube, Facebook, Instagram and TikTok, and what to do with a fake.

What is an AI agent, and which permissions should you not give it?

An agent is an AI that acts for you: it opens websites or apps, clicks, types and fills in forms, sometimes while signed in to your accounts. As of 31 October 2026 the agents offered to the public can handle short, clearly described tasks while you watch, but they still make mistakes, many need a paid plan, and several are not offered in Italy yet. Google’s help pages, for example, say tasks in Gemini are not available in the European Economic Area.

The main risk has a name: prompt injection, an instruction hidden in a web page, an email or a document, which the agent follows as if it came from you. Picture an agent ordering ingredients from a recipe site, where invisible text tells it to forward your bank’s latest email to a stranger. The makers say their filters reduce this risk but do not remove it.

Before you let an agent loose:

• start with a low-risk task, such as comparing prices or filling a cart you will check yourself;
• use a separate browser profile without your bank, health or government accounts;
• never type passwords, card numbers or codes in the chat;
• keep the setting that makes the agent ask before it pays or sends anything;
• watch it work, and stop it if it goes somewhere unexpected.

Six government cybersecurity agencies give the same core advice in their joint guidance of 1 May 2026: use agents only “for low-risk and non-sensitive tasks”. What is an AI agent? lists the agents available today, the plans they need, where they work and the settings that keep you in control.

What about children and teenagers?

Start with the minimum age. Seven of the nine assistants in the table allow users from 13, or an older age set by your country, and several ask for a parent’s permission for minors; Claude requires users to be 18, and DeepSeek says it is mainly for adults. In Italy, Google sets 14 as the age to manage your own Google Account. Italian Law 132 of 2025 on artificial intelligence says in Article 4, paragraph 4 that children under 14 need the consent of a parent or guardian to use AI technologies and to have their data processed for it.

Two makers describe tools for families. OpenAI switches on ChatGPT for Teens when it believes an account belongs to someone under 18, with protections for sensitive topics and optional parental controls; OpenAI says those controls do not let a parent read the teenager’s conversations. Google lets a parent switch off Gemini for a child under 13 through Family Link, its app for managing a child’s account.

In Italy, ChatGPT may also ask an adult for a live selfie or an ID to confirm their age. Why is ChatGPT asking for my ID or a selfie? explains who runs the check, what OpenAI says it receives and how to tell a real request from a scam. If your child uses an assistant, setting it up together is a good moment to look at the age, the privacy settings and what not to share.

Will AI change your job or your school?

For work, the big studies agree on more than the headlines suggest. AI could take over some tasks in between one job in four and almost four in ten worldwide, depending on the study: the International Labour Organization found one worker in four in a job with some exposure to generative AI, and the International Monetary Fund estimated that almost 40% of jobs worldwide are exposed to AI. “Exposed” means AI could do part of the work, not that the job will go. Up to mid-2026, the OECD’s Employment Outlook 2026 said companies “may have begun to automate certain jobs”, “although the data show little evidence of this to date”. One early warning comes from Stanford: in the United States, employment of 22 to 25 year olds in the most exposed jobs was 19% below where it would have been, mainly because fewer were hired, though the authors do not claim AI is proven to be the cause.

What you can do: list the tasks in your week, try an assistant on one screen-based task, keep confidential data out, and ask your employer for training. Under the EU’s AI law, companies that use AI systems must take measures to support the AI literacy of their staff, according to the European Commission’s questions and answers.

At school or university, ask your teacher what is allowed before using AI for homework, and say when you have used it. Will AI take my job? sets out the main studies side by side, with what each measured and what it cannot show.

What rules protect you?

In the EU the main law is the AI Act, and most of its rules apply from 2 August 2026. For everyday users, the most visible part is Article 50 on transparency:

• an AI that talks with people must make clear that it is an AI, unless that is obvious;
• companies whose AI makes images, video, sound or text must mark it in a way machines can read, so it can be detected as artificial; tools already on sale before 2 August 2026 have until 2 December 2026;
• whoever uses AI to make a deepfake, a realistic fake picture, video or sound of real people or events, must say it is artificial; the Act leaves out purely personal, non-professional use.

The stricter rules for high-risk uses, such as AI that helps choose who gets hired, apply from 2 December 2027 under the same Article 113. Your data rights under the GDPR, the EU’s data protection law, apply to AI too. In Italy, Law 132 of 23 September 2025 on artificial intelligence has been in force since 10 October 2025. These rules do not make AI answers correct and do not stop people who break them, so the checks in this guide still apply. For your own legal case, ask a lawyer.

How much energy and water does AI use?

Each question uses a small amount of electricity and water in a data centre, a building full of computers, but estimates vary a lot, because companies measure different things in different ways. Three figures, each from the organisation that produced it:

• Google, for its Gemini app: the median text question used 0.24 watt-hours of electricity and 0.26 millilitres of water, about five drops, in May 2025. Google compares the electricity to watching television for less than nine seconds. It counts the computers, the spare machines kept ready and the cooling of the data centre, and says the figures do not represent all Gemini questions and “have not been verified by an independent third-party”.
• Mistral AI, maker of Mistral Vibe, published a lifecycle analysis in July 2025, made with the consultancy Carbone 4 and the French environment agency ADEME and reviewed by two other firms. One answer of 400 tokens (the small pieces of text these programs count in) from its assistant Le Chat, now Mistral Vibe, used 45 millilitres of water and produced 1.14 grams of CO2 equivalent, the usual measure of greenhouse gases. Mistral says its figures include impacts upstream, such as making the servers, and calls the study “a first approximation”.
• The International Energy Agency, for the whole world: the electricity used by data centres, for AI and everything else, rose by 17% in 2025, and the agency expects it to double by 2030. That is a forecast, not a measurement.

The gap between Google’s and Mistral’s water figures shows how much the method matters. Both are for text answers: neither gives a figure for pictures, videos or long research tasks.

Eleven words to know

• AI assistant, or chatbot: a program you talk to in writing or by voice, which answers in ordinary language.
• Language model: the program behind an assistant, trained on huge amounts of text to write new text.
• Generative AI: AI that makes new text, pictures, video, voices or music.
• Prompt: the request you write to the AI.
• Hallucination: when the AI makes up an answer that sounds true.
• Training: how the program learns from examples; your chats can become examples unless you switch it off.
• Temporary chat: a conversation that is not kept in your history; the company may still keep a copy for a while.
• AI agent: an AI that clicks, types and fills in forms for you.
• Prompt injection: an instruction hidden in a page or a message that tricks an agent into doing something you did not ask.
• Deepfake: a realistic fake picture, video or voice of a real person or event, made with AI.
• AI Act: the EU law on artificial intelligence, with most rules in force from 2 August 2026.

Where to go next

• AI assistants compared: prices, ages, apps and data settings of nine assistants side by side.
• Which AI do I need?: six questions that sort the nine assistants into three groups by your needs.
• Is ChatGPT free?: what the free version gives and when Go, Plus or Pro are worth paying for.
• What happens to what you type: where your messages go and how to switch off training.
• What is an AI agent?: the AI that acts for you, and how to keep it under control.
• How to build an AI agent with Claude Code: a step-by-step guide for beginners, from installing it to a first working agent.

Common questions

Is AI free?

You can try all nine well-known assistants compared on this site without paying, with limits that most makers describe only in words. On 31 October 2026 the cheapest paid plan shown in euros was Google AI Plus at 4.99 euros a month, and ChatGPT Go cost 8 euros a month on OpenAI's pricing page seen from Italy. Pay only if you keep reaching the free limits or need something the paid plan adds. Is ChatGPT free? explains ChatGPT's plans in detail.

Which AI should I use?

It depends on what you need, not on a ranking. Check four facts: what you can spend, the minimum age if the assistant is for a teenager (from 13 to 18), whether the menus come in your language, and how the maker uses your chats. AI assistants compared sets out these facts for nine assistants, with a link to each maker's page.

Is it safe to use AI?

For everyday questions it can be, if you know what happens to your messages. What you type is stored on the company's computers, may be checked by people for misuse, and is used to improve the AI by most makers unless you switch that off. Leave out passwords, health records, identity documents and other people's details, and check any answer about health, money or the law with a professional or an official source.

Does AI understand what it says?

Not the way a person does. A language model writes the words that best fit your request, based on the huge amount of text it learned from, and it can write a fluent, confident answer that is false. Researchers at OpenAI and Georgia Tech compare it to a student guessing on a hard exam instead of leaving the answer blank. Treat it as a fast helper whose work you check, not as an expert who knows your situation.

Why does AI make things up?

A paper by researchers at OpenAI and Georgia Tech, dated 4 September 2025, says the way these programs are trained and tested rewards a guess more than an honest "I don't know", so they learn to guess when unsure. The result is an answer that sounds right and is not, which experts call a hallucination. Ask for sources and open them, and check names, numbers, dates and quotes before you rely on them.

Can AI replace a search engine?

For some questions it can help, but it does not replace checking. Several assistants search the web for you, and OpenAI says even free ChatGPT users can. The answer is still written by a language model, so it can misread or misquote a page: open the links it gives. For official facts such as laws, deadlines or prices, go to the official website.

How much energy and water does AI use?

Estimates vary a lot, because companies measure different things. Google says the median text question to its Gemini app used 0.24 watt-hours of electricity and 0.26 millilitres of water in May 2025. Mistral AI's lifecycle study gave 45 millilitres of water for one text answer of a set length from its assistant Le Chat, now Mistral Vibe, counting more than the data centre alone. Neither figure covers every kind of request.

Share this article

• LinkedIn
• X
• Bluesky
• Facebook
• WhatsApp
• Email

Topics: AI for beginners, AI assistants, ChatGPT, Gemini, privacy, AI agents, AI Act