How To Set Up Claude AI Safely. And The Toggle That Does Not Mean Private.

The Machine

Somebody asked me how to get one of these AI systems running without stepping on a landmine. Reasonable question, and the answer is shorter than people expect. The setup is four minutes. There is nothing to install and nothing you can break.

But there are three or four decisions inside those four minutes that nobody slows down for, and one of them is the most misunderstood setting in consumer AI right now. So I recorded the whole thing, screen and all, and made a new free account live so nobody has to take my word for how it looks.

I teach real estate agents. Before that I spent 20 Yrs LAPD teaching officers. Same lesson both rooms: the technology is almost never the hard part. Somebody just has to walk it through once at a normal speed, without making anybody feel slow for asking.

The Search Bar Is Where People Get Taken

The address is claude.ai. Type it in the address bar at the very top of the browser, not the search box in the middle of the page.

That distinction is the whole security lesson. When you search for a tool instead of navigating to it, you get a page that mixes the real destination with paid ads, review blogs, and products that borrowed the name because the name carries traffic. claude.ai is correct. claude.com is Anthropic as well, the company that builds Claude, so that one is legitimate too. Keep scrolling past those and it starts drifting into things that are not.

Nobody gets taken by anything sophisticated here. They get taken by a page that looks close enough and asks for a card.

The free account never asks for a credit card. That single fact is your tripwire. If a page wants payment details before it will make you a free account, close it and type the address yourself.

Then bookmark it, and you never have to find it again.

Signing Up, And The Agreement Nobody Reads

Two doors. If you already have Gmail, Continue with Google is the fast one and there is no real downside to it. If you do not, type any email address and Claude sends you a verification link. Open your email, click the link, and it finishes the job.

Then come the terms of service and the acceptable use policy.

Almost nobody reads them. Read the shape of them at least. The acceptable use policy is genuinely readable and it tells you what the system will and will not participate in. If a person sets out to use one of these tools to hurt somebody, that gets noticed and it gets handled.

And a word to my own generation, because this came up when I taught this live. We still carry an assumption that privacy is the default and that somebody is guarding it for us. That stopped matching reality a long time ago. Not at this signup screen. Somewhere back at the phone, or the smart TV, or the twenty page agreement nobody scrolled. Knowing what you agreed to is not paranoia. It is being awake.

The Pricing Screen Has Already Picked An Answer

Three tiers show up: Free, Pro, and Max.

Watch the Pro number. When that page loads, the annual plan is already selected, and the price it displays is the yearly rate divided by twelve. Pay for the year up front and it works out to roughly 17 dollars a month. Pay month to month and it is 20.

I do not think anybody built that to trick you. It is how nearly every subscription on the internet presents itself now. But there is a difference between seeing it happen and having it happen to you. Find the monthly and annual toggle, click it, and look at the real number for the plan you would actually pick. On the Max tier, worth knowing, there is no annual discount at all.

Start free. The free plan is not a demo. It is the real assistant with a daily ceiling, and for the first several weeks that ceiling is generous.

The way to think about usage: it works like quarters in an arcade game. You get a certain amount of play, and heavy work spends it faster. Deep research, long documents, detailed back and forth, those burn through it quicker than simple questions. When you start hitting the ceiling regularly, that is the tool telling you it earned the money. Not before.

Skip The Download. For Now.

Next screen offers the desktop app. One correction I want on the record, because I got this slightly wrong on camera: the desktop app is free on every plan, including the free one. Cost is not the reason to skip it.

Skip it anyway on day one. Use the browser. Fewer moving parts, nothing installed, nothing to uninstall if you decide this is not for you. Once it has earned a permanent spot in your week, then install the desktop app or put it on your phone.

When you do go to the phone: search Claude in the App Store or Play Store and check that the maker says Anthropic. Lookalike apps sit right next to the real one wearing similar names and similar icons. Read the name on the badge before you believe the uniform.

The Toggle Almost Everybody Reads Backwards

Here is the one I actually made this video for.

You will see a switch asking whether your conversations can be used to help improve the model. Turn it off. That is the right instinct and I turned mine off too.

Now understand what you did and what you did not do.

People flip that switch and conclude their information is sealed. It is not. That toggle governs one specific thing: whether your conversations get used to train future versions of the model. It does not govern where your words went.

Anything you type into any AI system that is not running on your own machine leaves your building. It travels to a server somebody else owns. It sits there. It is reachable by subpoena. That is true with the switch on and true with the switch off.

The only genuinely private setup is a model running on hardware you own, and that means real hardware. NVIDIA sells a desktop AI machine called the DGX Spark around 5,000 dollars. AMD has a comparable box nearer 5,700. A loaded Mac Studio lands in the same neighborhood. Those run a model locally, where nothing leaves the room.

For almost everybody watching, that is not the setup, and that is fine. You are using a system that belongs to somebody else. Just operate like you know it.

So the working rule is simple. Everyday questions, drafting, research, learning, planning, all of that, go ahead. Social Security numbers, bank and account numbers, passwords, executed contracts you are not free to share, anything under medical or credit or financial privacy law, none of it, ever. Not because this company is out to get you. Because that is street sense on any website.

I went deeper on the data side of this in You Are Training AI For Free. The tool is not the advantage. What you own and what you feed it is the advantage.

Introduce Yourself, Then Overrule Yourself

Claude asks your name and what you do. Writer, educator, engineer, health care, finance, legal, student, founder, consultant, scientist, and a long list past those, including Other.

Pick something. It is not a commitment. That choice shapes how it talks to you and what it offers. Pick educator and it leans toward explaining concepts. Pick operations and it starts offering to draft procedures.

Here is the part worth knowing. You can overrule it at any time by simply saying so. On camera I told it I had made a mistake, I was not an educator, I was a rocket scientist, and it turned on a dime. No settings menu, no hunting. You talk, it adjusts.

That is the thing to absorb about this whole category. It is not software with buttons you have to locate. It is closer to working with a very sharp assistant who listens.

The First Four Things To Type

You are in. There is a box at the bottom. Start here, in this order.

One, ask it how to work with it. Type: Hello, can you tell me the best way to use you? It will tell you. Be specific about context, iterate instead of starting over, bring it a whole workflow instead of one isolated question. Straight from the tool.

Two, teach it who you are. Type: Can you help me teach you about how I like to do things? This one matters more than it looks if you are going to produce anything with your name on it. The gap between generic AI slop and something that sounds like a person is almost entirely this step, and almost nobody does it.

Three, ask what the buttons mean. Projects, Artifacts, Code, Customize. Do not go hunting for a manual. Ask it. It will explain its own interface.

Four, make it come down to your level. This is the most useful sentence I have ever handed anybody: Explain that to me like I am in the 7th or 8th grade.

There is no insult in it. Strip out the jargon and the light comes on dramatically faster. I ask for that on subjects I already know well, because the plain version is usually the version that sticks. Ask as many times as you want. It never sighs, never rushes you, never makes anybody feel slow.

That patience is not a side effect. It is the whole reason this thing works for people who spent thirty years being told they were bad with computers.

Who Builds What, So The Names Stop Being Noise

Quick map. Anthropic builds Claude. OpenAI builds ChatGPT. Microsoft has Copilot. xAI has Grok. Meta, which is Facebook, has its own. There are more.

As of today, August 22, 2026, the two out front for general use are Claude and ChatGPT.

One difference worth knowing while you are choosing. Claude states plainly that it shows no ads and does not let advertisers influence answers. ChatGPT began showing ads to free and lower tier users in the United States back in February 2026. OpenAI's stated position is that those ads are labeled, sit separately from the answer, and do not change the answer itself, and their paid tiers stay ad free. I am watching how that plays out rather than calling it. The point is to know how the thing in front of you is funded.

If you want the deeper wiring on Claude specifically, look up constitutional AI. That is a real rabbit hole and a good one. Not required to use the tool.

One more thing, and it connects to what I have been covering all year. The versions the public gets are not the most capable versions that exist. The sharpest systems go to large organizations first, under tighter terms, while everyone else gets something more constrained. I wrote about that dynamic in what happens when one company gets there first, and about the timeline argument in Mo Gawdat's case for 2027. This post is the other end of that same story. While the frontier argument runs hot, there is a free tool sitting there that would solve a real problem in front of most working people this week.

Bonus: The Link Structure I Actually Use

One more thing for anyone who wants to go a step further than the setup itself. I recorded a short screen walkthrough of the actual link structure I use to keep my own AI tools, bookmarks, and resources organized so none of it gets lost once this becomes part of your week. Companion to everything above, not a replacement for it.

What To Do Tonight

Do not try to master this. Do one loop.

Go to claude.ai. Make the free account. Skip the download. Turn the training toggle off. Pick any role. Ask it one real question about something actually on your plate this week. When the answer comes back too technical, tell it to explain like you are in the 8th grade.

Fifteen minutes, start to finish.

And once it answers you well one time, ask it the follow up you would have been embarrassed to ask a person. That is the moment most people stop treating it like a search engine and start treating it like a tool.

The people the internet left behind were never the older ones. They were the ones who never sat down and tried.