Start here: a parent's guide to AI

1. What AI actually is

Four words that get muddled

Have you noticed that "AI" now seems to mean everything from your phone sorting photos to a chatbot writing a poem? It's one word doing a lot of jobs, so it helps to pull four of them apart.

AI is the broad term. It's software doing things that used to need a person: spotting faces in your camera roll, filtering spam, suggesting what to watch next. It's been quietly in your phone for years.

Generative AI is the newer kind, and it makes something new: a paragraph, a picture, a plan, a summary. When people say "AI" now, this is usually what they mean.

A chatbot is one way of using generative AI. It's the box you type into and get a reply from. ChatGPT, Claude, Gemini and Copilot are all chatbots.

You'll also hear agentic AI. That's when the tool doesn't just answer you but goes and does something: fills in a form, sorts a folder, books a slot, working through the steps on its own. Most chatbots now have a mode like this, and section 9 covers when it's worth using and what to watch.

So AI is the whole field, generative AI is one part of it, a chatbot is one door into that part, and agentic AI is what you get when the chatbot is allowed to act as well as talk.

How it comes up with an answer

Ever wondered why it can write a perfectly good thank-you note but confidently tell you a school holiday date that's wrong? Both come from the same place.

Behind the chat box is a language model. To build one, software was fed a huge amount of text and learned, statistically, which words tend to follow which. When you type a message, it writes the reply one small piece at a time, each time choosing a likely next word given everything so far.

The way I think of it is an assistant who has skimmed most of the internet, took no notes, and can't check anything before they speak. Ask them a question and you'll get a fluent, confident answer in the shape of a good answer, because they've seen thousands of good answers.

That's also where the analogy stops. A person usually knows when they're guessing. A language model doesn't, in any real sense. It writes plausible text, and plausible text is usually right and sometimes wrong in exactly the same tone. So the wording tells you nothing about whether it's true. That one idea is worth more than most of the rest of this guide.

What it tends to be good at

Anything where the shape of the answer matters more than a precise fact. Drafting the email to the school you've been putting off. Turning a rambling list into a plan. Explaining a topic at the level you ask for. Coming up with ten ideas so you can pick one. Summarising a long document you've pasted in. And, which people don't expect, asking you questions back when you're trying to think something through.

Where it struggles

It can invent facts, dates, quotes and references. Unless it searches the web, it may be working from information that's months or even years old. It's unreliable at exact arithmetic unless it uses a calculator tool. And it doesn't know your family, your children's school or what happened yesterday unless you've told it in that conversation.

Why you get a different answer the second time

If you've asked the same thing twice and got two different replies, that's not a fault. Because it's choosing likely words rather than looking up one fixed answer, the same question can come back differently. A little randomness is built in on purpose so it doesn't sound robotic, and small changes in how you phrase things matter too. It's a good reason to ask twice on anything that matters.

2. Which tool should I start with?

If you've been putting this off because you don't know which one to pick, I'd take some comfort from this: the main tools overlap far more than they differ. All of them can draft, summarise, explain, plan and chat, and all of them have a free way in. So the choice is less important than it looks. If your life is in Gmail and Google Docs, Gemini is already there. If you're a Microsoft household, so is Copilot. If you've tried ChatGPT and it felt fine, there's no reason to move.

Company, app and model

Has a friend told you to "use the latest model" and you couldn't find it? This is why.

There are three things that often share a name. The company makes the product: OpenAI, Anthropic, Google, Microsoft, Perplexity. The app is what you open: ChatGPT, Claude, Gemini, Copilot, Perplexity. And the model is the engine inside, with names like GPT-5.6, Claude Opus 5 or Gemini 3.6 Flash.

The apps swap models in and out as new ones arrive, sometimes several times a year, and some let you choose between them. One app can even run another company's engine: Copilot uses OpenAI's models, and Perplexity lets paying users pick from several companies' models.

You don't need to follow any of this. Choose an app, get to know its settings, and let the engine updates happen underneath.

The main ones at a glance

Sources: ChatGPT free tier, Claude plans, Gemini plans (UK), Copilot plans (UK), Perplexity pricing.

Other things you'll hear about

Image generators are one, though the chatbots above all make images now, so you don't need a separate tool for that. AI browsers, like Perplexity's Comet or Copilot inside Edge, can read and act on web pages for you. And there are tools for working with your own documents, where you upload files and ask questions about them, which all five above can do too. None of these needs to be where you begin.

3. AI you didn't choose

Have you noticed a new coloured circle in WhatsApp, or a "help me write" button appearing in Gmail? Most of us don't meet AI by going and opening a chatbot. It turns up inside apps we already use, and it's often the first AI a child talks to. These come with their own settings and their own rules, and they're easy to miss because nobody signed up for them.

Here's what I'd know about the ones most families will run into.

Meta AI in WhatsApp, Instagram and Messenger. Meta says it's optional and only "on" when you choose to message it, and that your ordinary end-to-end encrypted messages aren't seen by it (WhatsApp). What you send to Meta AI itself is a different matter, and is used by Meta. In a WhatsApp group you can switch on Advanced Chat Privacy, which stops people in that chat using the messages with AI features (WhatsApp). Separately, Meta uses public posts from UK adults to train its models, not under-18s' accounts, and offers an objection form (Meta).

Snapchat's My AI. This sits at the top of a teenager's chat list, and anything sent to it is kept until deleted and can be used to improve Snap's products and personalise ads (Snapchat). It can be unpinned or cleared from the chat feed (Snapchat), and if you're set up in Snapchat's Family Center you can switch it off for your teen entirely (Snapchat).

Gemini in Gmail and Google Docs. Google says it doesn't use your Gmail or Docs content to train the models behind Gemini (Google). If you'd rather not have the features at all, there are three "smart features" switches in Gmail's settings, under General, that turn them off across Gmail, Workspace and other Google products (Google).

Copilot in Word and Excel. On a Microsoft 365 Personal or Family subscription you can turn Copilot off in Word on a computer (File, then Options, then Copilot on Windows), though not in the phone or web versions (Microsoft).

Apple Intelligence. On recent iPhones, iPads and Macs it's under Settings, then Apple Intelligence & Siri, where it can be turned off (Apple). Its link to ChatGPT is separate and opt-in, for over-13s, and you're asked before any photo or file is sent (Apple).

Voice assistants. Alexa and Siri are now being rebuilt on the same kind of models as the chatbots. Alexa+ arrived in the UK in early access this year (Amazon) and Apple has announced an AI version of Siri (Apple). The thing to bear in mind is that the speaker in the kitchen is becoming a chatbot with a microphone, and one the children can talk to without a screen. The same questions apply: what does it keep, who can hear it, and does it know who's asking?

None of this needs sorting today. It's worth a look through the apps your family actually uses, one at a time, in the same way you'd look at the chatbot settings in section 5.

4. Do I need to pay?

Probably not to begin with, and for a lot of parents, not at all. Here's how I'd think about it.

What a free account gives you

Drafting, summarising, explaining, planning, ideas, questions, the occasional document or photo upload, and a go on voice mode. Free tiers have usage limits, and they may move you onto a smaller model when things are busy. If you use it a few times a day, you may or may not notice that.

What paying adds

Higher limits, the most capable models, the longer research tasks that run for a few minutes, more file uploads, and the newer features where the tool does things for you rather than answering. At the time of writing, the main individual plans sit around £19 to £20 a month. Google AI Pro is £18.99 (Google, UK). Microsoft 365 Premium is £18.99 (Microsoft, UK). ChatGPT Plus, Claude Pro and Perplexity Pro are each listed at $20 in US dollars and converted at checkout (OpenAI, Anthropic, Perplexity). ChatGPT also has a cheaper Go tier. Prices move, so the official page is worth a look before you subscribe.

Working out whether it's worth it

I'd start with one free tool and see what you actually use it for. After a few weeks, one of these will probably be true.

If you keep running into a limit and being told to wait, that's a useful point to consider whether paying for that one tool would help.

If you want it to read long documents or do a proper research job for you, a paid plan on one tool is likely to earn its keep. Where there's a free research mode, it's worth trying that first.

If what you're after is better answers, paying tends to help less than people expect. Asking more clearly, which section 7 covers, usually makes a bigger difference.

If you already pay for Google One or Microsoft 365, it's worth checking what's included before buying anything else. Microsoft 365 Premium replaced the old Copilot Pro (Microsoft).

And if you're not sure, staying on the free version costs nothing. Nothing in this guide needs a subscription.

5. Setting up your first AI tool

This takes a little while, but it's a one-off, and it's the part most people skip.

1. 1

   Choose one. Use the table in section 2. If you can't decide, the company whose email you already use is a reasonable tiebreak.

2. 2

   Find the genuine one. Type the address in yourself rather than searching for it, because there are look-alike apps that borrow the names. The real addresses are chatgpt.com, claude.ai, gemini.google.com, copilot.microsoft.com and perplexity.ai. In the app stores, check the publisher name: OpenAI, Anthropic, Google LLC, Microsoft Corporation, Perplexity AI, Inc. (OpenAI's note on checking the publisher). If an app asks for payment before you've seen a chat box, that's a reason to close it.

3. 3

   Use the right account. Family things belong in a personal account. If you also have a work login for the same tool, I'd keep the two apart. Work and school accounts often come with stronger protections (on ChatGPT, Claude, Gemini and Copilot they aren't used for training by default), but your employer can usually see and manage what's in there, and their rules about what you can type in may be stricter than you'd guess. It's worth reading your employer's AI policy before anything from work goes into a personal account, or anything personal goes into a work one.

4. 4

   Secure it. A password you don't use anywhere else, and two-step verification if it's offered. If you sign in with Google, Apple or Microsoft, remember that account's security is now the AI tool's security as well.

5. 5

   Look at the privacy settings once. Each product puts them somewhere different; the notes below have the paths. You're looking for three things: whether your chats are used to improve the model, whether it keeps a memory about you, and how long chats are kept. Section 6 explains what each one does, so you can decide rather than switching everything off just in case.

6. 6

   Try something that doesn't matter. No personal details, and something you can judge yourself. "Give me five ideas for a rainy Saturday with a seven-year-old, nothing that needs a car" is the sort of thing. See what comes back, then reply with a correction or a follow-up. That back-and-forth is most of what there is to learn.

Where the settings live

Menus move around. If a setting isn't where I've said, the app's search box or help page is a better bet than assuming it's gone.

Setup checklist

6. Settings worth understanding

If you've ever opened the settings, seen a row of switches and quietly closed them again, this section is for you. None of these is a plain on or off. Each one is a trade, and a thoughtful parent could reasonably go either way. What matters is knowing what you're choosing.

Using your chats to improve the model

When a company says it uses your conversations to "improve" its models, what does that actually involve? It means what you type, and what the tool replies, can become material for training future versions. On most consumer tools it's a single switch, and section 5 has where it lives. Work, school and business accounts are generally excluded by default.

Leaving it on means you're contributing to better tools for everyone, which some people are happy to do. Turning it off keeps your chats out of that process. If you mostly ask about dinner ideas, you may not mind either way. If you paste in school reports or write about a child's difficulties, you'll probably want it off. Two things worth knowing: switching it off doesn't always cover chats you've already had, and it doesn't delete anything.

Chat history and how long it's kept

Deleting a chat feels final. It usually isn't, quite. The tool keeps a list of past conversations so you can go back to them, and when you delete one it disappears from your view, but the company generally holds it for a while afterwards. ChatGPT and Claude both describe up to 30 days before it's permanently gone (OpenAI, Anthropic). Gemini keeps activity for 18 months by default, which you can change to 3 or 36, and chats a human reviewer has looked at can be kept for up to three years even after you delete them (Google). Copilot keeps conversation activity for 18 months by default (Microsoft).

History is useful for picking up where you left off. A shorter auto-delete period, where you can set one, means less builds up in your account without changing how you use the tool.

Memory and personalisation

Have you noticed a tool bringing up something you mentioned weeks ago? That's memory. The tool builds a short profile from your chats (two children, prefers short answers, vegetarian) and uses it in future conversations. You can usually see, edit and delete the individual items.

The upside is that you don't have to repeat yourself, and answers get more relevant. The downside is that something you said once can resurface, and anyone using your login gets answers shaped by it. One middle path is to leave it on and open the memory list every so often to remove anything you'd rather wasn't there. ChatGPT's teen setup and Gemini both restrict memory for under-18s.

Temporary or private chats

Sometimes you want to ask something without it sitting in your history. Most tools have a mode for that: a chat that doesn't go in your history, doesn't create memories and isn't used for training. ChatGPT calls it Temporary Chat, Claude calls it Incognito, Gemini calls it Temporary chat, and Perplexity calls it Incognito.

It's good for one-off questions. What it isn't is a guarantee that nothing is kept. ChatGPT and Claude hold temporary chats for up to 30 days for safety reasons, Gemini for 72 hours, and Perplexity says incognito sessions expire within 24 hours (OpenAI, Anthropic, Google, Perplexity).

Four things that sound the same and aren't

These get muddled together, so here they are side by side.

Turning off training stops future chats being used to improve the model. It doesn't delete them.

Deleting a chat removes it from your account after a holding period. It doesn't undo any training that already used it.

Turning off memory stops facts carrying between conversations. The existing memories may need deleting separately, and the chats themselves are still there.

A temporary chat keeps one conversation out of history, memory and training. The company still holds it briefly.

Doing all four gives you the most privacy a consumer tool currently offers. None of them means nothing is kept.

Connected apps and permissions

What happens when you let a tool see your calendar? It can answer "what's on Friday?" without you pasting anything in, which is handy. Gemini calls these Connected Apps, Claude calls them connectors, ChatGPT calls them apps.

Every connection widens what the tool can read, and sometimes what it can do. My suggestion would be to connect one thing at a time, when you've got a use for it, and to look at the list now and again to see whether you still need everything on it.

Sharing a chat with a link

Most tools let you make a link to a conversation, and on personal accounts anyone with the link can read it. In some cases uploaded images and files are visible too (OpenAI, Google).

Before sharing, it's worth rereading the whole chat, not just the end, for names, schools and anything you wouldn't want a stranger to see. Each tool has a page listing your shared links where you can delete them.

Voice, camera, photos and files

Talking to it, pointing your camera at something, uploading a photo or a document: all of this is useful, and the free tiers of most tools include some version.

The thing to bear in mind is that audio, images and files are data too, sometimes with their own rules. ChatGPT keeps voice clips for 30 days (OpenAI). Gemini Live recordings go into your activity if Keep activity is on (Google). Giving the app camera and microphone access only while you're using those features is one way to keep it in check, and it's worth a second look before photographing anything with other people's faces or a child's school details in it.

If you decide it's not for you

What if you try one of these, decide you'd rather not, and want to close the door properly? Every tool in this guide lets you download your data and delete your account, and it's worth doing both rather than just deleting the app. ChatGPT: Settings, then Data controls, to export; Settings, then Account, to delete (OpenAI). Claude: Settings, then Privacy, to export; Settings, then Account, to delete (Anthropic). Gemini's data comes out through Google Takeout (Google), and deleting it means deleting the Google account or clearing Gemini activity. Copilot's history is exported and deleted from the Microsoft privacy dashboard (Microsoft). Perplexity has both on the account page (Perplexity).

Deletion isn't instant. Most of these say the data goes within about 30 days, and Microsoft gives you a window to change your mind before the account closes. Cancel any subscription first, or it may keep billing.

7. How to ask for useful help

If you've typed "help me with X" and got back something bland and generic, the problem usually isn't you or the tool. It's that it didn't have enough to go on. The way I'd think about it is briefing a capable temp on their first morning: what the job is, the background that matters, anything to steer clear of, and what you'd like back.

The four parts of a clear request

You won't always need all four. But when an answer disappoints, one of them is usually what's missing.

Vague to useful, one example

Suppose you start with "Help me with a birthday party." You'll get a generic list, because that's all it can do with that.

Try instead: "Plan a two-hour birthday party at home for eight children aged about six. Budget under £60, no cooking, one adult helper, garden if it's dry. Give me a timed schedule and a shopping list."

And then follow up: "Swap the craft for something less messy, and tell me which bits fall apart if it rains."

Notice that nothing in there identifies the child, the house or the guests. "Eight children aged about six" is all it needs.

Placeholders

When real people are involved, placeholders do the job: [child], [school], [teacher]. Get the draft written with the placeholders in, then fill them in yourself in your own email. If you're pasting in something you already have, like a letter from school, take out the names, address and reference numbers first. It'll do the task just as well without them.

Talking back

The first reply is a first draft, not a verdict. If it's wrong, say so and say why. Ask for three alternatives. Ask "what in this are you least sure about?" Ask it to explain a word it used, or for a plainer version. This back-and-forth is where most of the usefulness comes from.

It tends to agree with you

Have you asked one of these tools whether something is a good idea and noticed it nearly always says yes? That's worth knowing about. These models are trained to be helpful and agreeable, and it shows. Suggest a plan and it'll usually find the merits in it. Push back and it'll often fold, even when it was right.

For a lot of tasks that doesn't matter. When you're using it to think through a decision, it does. Two things help. Ask for the case against as well as the case for, or ask it to argue the other side properly. And treat a warm, encouraging reply as exactly that, a reply, rather than a second opinion from someone who knows your situation.

Three first prompts to copy

8. Key phrases explained

These are the words you'll keep bumping into. Each one has a plain meaning, an everyday comparison where one helps, and a note on where the comparison stops being accurate.

How they fit together

Every reply is produced by a model during inference, working only from its context: your prompt, the chat so far, and anything loaded in from memory, a project, custom instructions or a skill. Given tools and connectors, it can look things up or take actions. When it strings several of those steps together on its own, it's acting as an agent, and an automation decides when that work starts. Training is what made the model in the first place, and hallucination is the price of the way it works.

9. Chat, search, research and doing work

Did you know an ordinary chat doesn't necessarily look anything up? Four quite different things happen in the same box, which is why they get muddled.

Asking a question

A plain chat answers from what the model learned in training plus whatever's in context. It doesn't automatically search the internet. ChatGPT may search on its own when it decides a question needs current information (OpenAI). Gemini grounds some answers in Google Search (Google). Claude searches once web search is on and it judges a search would help (Anthropic). Copilot searches and cites sources (Microsoft). Perplexity searches every time. If it matters, you can say "search the web for this" and look for citations in the reply. No citations usually means no search happened.

Searching the web

When it does search, it reads a handful of pages and summarises them. That's quicker than doing it yourself, and worse at noticing a page is old, biased or plain wrong. The summary makes a good starting point; the sources are where you go for anything that matters.

A longer research job

Most tools now have a research mode (Deep Research in ChatGPT and Gemini, Research in Claude and Perplexity) that runs for a few minutes, reads a lot of pages and produces a report with references. It suits "what are my options for X" questions. It can still get things wrong, especially in how it describes what a source says, so the references it leans on most are worth opening. Gemini offers it free to over-18s (Google); Claude's Research is on paid plans (Anthropic); Perplexity gives free users limited access (Perplexity).

Letting it do things

The newest features let a tool act rather than answer: fill in a form, edit a document in your files, send something through a connected app, book something. This is what people mean by "agentic", and section 8 has the word agent in more detail.

Chat versus projects: a chat is one conversation. A project is a folder of related chats and files with standing instructions. A project earns its place when you'll keep coming back to a topic, like a school application or a house move. A plain chat does for most other things.

Chat versus agentic tools: in a chat, you get a reply and decide what to do with it. In an agentic mode, you describe the outcome and the tool works through the steps, using the files and services you've given it. At the time of writing that includes Claude's Cowork, on paid plans as a desktop app and in beta on web and mobile (Anthropic), ChatGPT's "Work" mode, rolling out on paid plans (OpenAI), and browser-based actions in Perplexity's Comet and Microsoft Edge. Gemini's agent, Gemini Spark, isn't available in the UK at the moment (Google).

What changes when it can act

When a tool can only talk, a mistake costs you a bad paragraph. When it can reach your files, edit documents or use other services, a mistake can overwrite a document, send a half-finished email or fill in a form wrongly. A few habits help.

Give it the smallest area that does the job: one folder rather than your whole drive.

Keep "ask before acting" switched on where it exists. Most of these tools have a setting somewhere between "always ask" and "just do it".

Look before anything leaves your device or can't be undone: sending, submitting, paying, deleting.

And start with jobs you can see and check, like sorting a folder of photos into dated subfolders, before trusting it with anything that touches money or other people.

10. Family information and data safety

Where does it go, the thing you just typed? To a company's computers, where it sits for a while, might be seen by staff in limited circumstances, and depending on your settings might be used to improve the product. That's different from a note on your own phone. None of the services in this guide can fairly be called completely private, and if one tells you it is, that's a reason to look more closely rather than less.

I don't think this calls for anxiety. It calls for a habit: describe things in general terms, take out anything identifying, and keep a short list of things out altogether.

A bit of nuance. Health is where it's most tempting to overshare, because the tool is helpful with it. Asking about a symptom or a condition in general terms is fine; uploading the letter from the consultant is the bit I'd stop short of. For photos, a picture of a maths worksheet works perfectly well without the child's name at the top, and a thumb over it does the job. And when a message has details about someone else in it, it's worth asking whether they'd be comfortable with it sitting on a server they've never heard of.

11. When to check an answer

Has it ever given you a school term date that turned out to be wrong, or a reference you couldn't find anywhere? That's the thing to understand about checking: the answers that need it look exactly like the ones that don't.

Four ways it goes wrong

Made-up facts: a confident statement with nothing behind it, whether a statistic, a date or a rule that doesn't exist.

Bad citations: a reference that looks real and isn't, or a real page that doesn't say what it claims. This is common in research modes.

Out-of-date information: training has a cut-off, and unless it searched, it may be presenting last year's rules or prices as current.

Confident slips: an arithmetic error, a misread question, or two true facts joined into a false conclusion, all in the same assured voice.

A short checking routine

For everyday things, drafts and ideas, your own judgement is usually the check. You know whether the email reads well.

For anything touching health, money, legal matters, or a decision that matters about a child, I'd do four things.

1. 1

   Ask where it came from. "What's the source for that?" Then open it and see whether the page exists, whether it's official or reputable, and whether it actually says that.

2. 2

   Check the date. Is the source current? Did the tool search, or answer from training?

3. 3

   Cross-check once, somewhere that isn't the chatbot: the official site, GOV.UK, the school, the GP, a professional.

4. 4

   Ask what it's unsure of. "What in this would you want a professional to confirm?" often produces a useful list.

Why "are you sure?" isn't a check

Asking whether it's sure just produces more fluent text. It might double down, or it might apologise and change a correct answer to please you. Neither tells you anything about the truth. Checking means a source that isn't the chatbot.

Keeping it in proportion

None of this means distrusting everything. It means matching the checking to the stakes. A party plan needs no checking at all. A question about which medicines a child can take together needs a pharmacist, and the tool's role there is to help you get your question straight before you ask.

12. What about children using AI?

There's a real difference between you using AI to help with family life and a child using one of these services directly. Most of this guide is about the first. This section is the ground for the second.

Ages

At the time of writing, the official minimum ages for consumer accounts are: ChatGPT 13, with a parent's permission needed under 18 (OpenAI terms); Claude 18 (Anthropic consumer terms); Gemini 13 for personal accounts, and not available on Family Link supervised accounts for under-13s in the UK (Google, Google Families); Copilot 13, with under-18s' chats kept out of training and personalisation (Microsoft); Perplexity 13 (Perplexity terms).

A minimum age is a legal line, not a sign that the product was designed for children. These are general-purpose adult tools with some teen safeguards added.

Accounts and parental controls

If a teenager is going to use one, an account in their own name and age works better than sharing yours, because the safeguards depend on the tool knowing the user is under 18. ChatGPT has parental controls you link from your own account, covering quiet hours, memory, voice, image generation and training, with safety notifications for serious concerns. They don't let you read your teen's chats (OpenAI). Microsoft points parents to Family Safety and the controls on the device itself (Microsoft). Anthropic doesn't offer under-18 consumer accounts at all.

Separately, "AI companion" apps, the ones built for relationship-style chat, are a different category from anything in this guide. Common Sense Media rates them unacceptable for under-18s (Common Sense Media), and the UK government has said it will require an 18+ minimum age for them, expected to take effect in 2027 (GOV.UK).

Doing it together

For younger children, the approach I'd lean towards is together, on your account. You type, you read the answer out, and you talk about whether it seems right. It teaches the one thing that matters most: that the tool is often helpful and sometimes wrong, and you can't tell which from the wording.

School rules

Schools set their own rules on AI for homework and coursework, and they vary a lot, so it's worth asking before assuming, especially with older children where AI-written work can count as malpractice. The Department for Education publishes guidance for schools (DfE). If the school uses Gemini through Google Workspace for Education, those accounts aren't used to train Google's models, and some features are limited to over-18s (Google).

Talking about what it can't do

Children tend to assume a confident answer is a right one, and that there's someone on the other end. Two conversations are worth having early: that it produces likely-sounding text rather than checked facts, and that it isn't a friend and doesn't know them, however friendly it sounds. The NSPCC and Internet Matters both have short, practical guides for these conversations (NSPCC, Internet Matters).

13. AI coming the other way

So far this has been about AI you use. There's also AI that gets used on you, and it's worth a short word because it's where a lot of the worry about families and AI actually sits.

Voice and message scams. The "Hi Mum, I've lost my phone" message asking for money has been around for years (Action Fraud). Cheap voice cloning means the same trick can now arrive as a call that sounds like your child or your parent. The defence hasn't changed: don't send money on the strength of a message or a call alone, ring them back on the number you already have, and agree within the family that an urgent request for money always gets a second check (Take Five).

Fake pictures and video. AI can make a convincing image or clip of anyone, including children. Creating or asking for sexual deepfakes of adults became a criminal offence in England and Wales in February 2026 (GOV.UK), and images of children were already illegal. For the ordinary version of this, spotting an AI picture in the family group chat, Internet Matters has a short guide (Internet Matters). The habit worth building, for you and for children, is that a photo or a clip is no longer proof on its own.

Things written by AI. Reviews, articles, adverts and messages are increasingly written by the same tools you've been reading about. Knowing how fluent and confident they sound, which section 1 covered, is most of the protection: fluent isn't the same as true, whoever pressed the button.

This is a big subject and a short section. If it's the part that worries you most, the NSPCC and Internet Matters pages linked at the end are a good next stop.

14. A first step

You don't need a plan for this. An afternoon and one small job will do.

Pick one tool, whichever is nearest. As section 2 said, the choice matters less than it looks.

Open its settings once and find the training setting and the memory setting. Decide what you'd like them to be. Those settings are a useful place to start; it's also worth a glance at what you've connected and a think about which family details you're comfortable sharing.

Try one thing that doesn't matter much. No personal details, and a result you can judge for yourself. Rewrite a message you've been putting off. Get five ideas for tomorrow's tea from what's in the fridge. Ask it to explain something your child asked that you couldn't answer.

Then ask yourself one question: was that useful? Not clever, not impressive. Useful. If it was, you've found your first use, and you can stop there for now. If it wasn't, try asking once more with a bit more detail, and then leave it for another day.

One useful thing is enough to begin with. The rest of this will still be here when you want it.

Features, prices, settings and age rules change often. If something here doesn't match what's on your screen, the product's own help pages are the place to look.

Related reading: NSPCC, Talking to children about AI · Internet Matters, AI chatbots guide for parents · Ofcom, AI chatbots and online regulation · DfE, Generative AI in education