Does ChatGPT search the internet? How AI uses tools

You ask an AI for the last date to send a form. Back comes a date, a link, and a very sure tone. You open the link, and the page says nothing at all about a last date.

Nothing broke. The machine did exactly what it is built to do. This lesson shows you what that is, and where your own job starts.

Most of the time it is not online

By default the app is not on the internet. It is answering from what it read while it was being built.

That reading stopped on a date. The date is called the cut-off. Think of a newspaper that stopped arriving at your house. Everything after that day is missing. Worse, the last months before that day are thin, so fairly recent things are the shakiest of all.

How a model is made explains where all that reading comes from.

It does not even know today’s date by itself. The app writes the date into a hidden note that is placed in front of your message. So type the date and the year you mean into your question.

Searching is a separate action. It happens when you switch the feature on, or when the app decides to use it for you. It is not going on quietly in the background every time.

What really happens when it searches

Here is the part almost nobody is told.

The model cannot open a website. What it can do is write a request, like a slip you hand across a counter. A different program takes that slip, runs the search, and fetches the pages.

The text that comes back is then placed in front of the model, as more text inside your chat. Only then does it write the answer. The lesson on the context window calls that space the table it can see.

So it is reading a page that was fetched for it. It is not browsing the way you browse. Picture an open book exam. Somebody put the book on the desk. The student still writes the answer in their own words.

Two things follow from this. It learned nothing from that page, so tomorrow’s chat starts blank. And the answer is only as good as the page that landed on the table.

When the wrong page lands on the table

Suppose the search brings back an old page. Or a copycat site. Or a page about another state’s rule.

The model does not stop and tell you the page looks wrong. It builds the answer on that page, in the same calm voice it always uses.

That is the real failure. Not a wild invented answer. A careful answer built on the wrong paper.

Turning search on is still the biggest single thing you can do for facts. It cuts made-up answers a lot. It does not end them.

One 2026 test used hard questions and kept search switched on. About three answers in ten still carried a claim that the linked page did not support.

A large study of free assistants in 2025 found something similar about news. Nearly half the answers had at least one serious problem. The biggest single problem was not the facts. It was sourcing. The assistants also almost never said they did not know.

Turn search on for these questions

The rule is short. If the answer contains a date, a price, a deadline, a form number, a scheme rule or anything in the news, turn search on.

Where is the switch? Do not hunt for a fixed corner of the screen. Look near the box where you type for a small globe, or a word like search or browse. Many apps now decide by themselves and print a line saying they searched.

You can also just ask for it in the message. This works in almost any app. Fill in today’s real date before you send it.

Search the web before you answer. Today is [write today's date here].
My question: which papers are needed for a new voter ID card in Uttar Pradesh?
Give me:
1. The answer in three short lines.
2. The date of the information.
3. The link of the page you used.
If you cannot find a reliable page, write "I could not find it".
Do not guess.

A good reply gives you three short lines, a date, and a link you can open. A weak reply gives you a long sure paragraph and no link at all. Treat the second one as unchecked.

A link is not proof. Open it.

This is the harder half, and the half nobody teaches.

A link that opens proves nothing. The page has to actually say the thing the answer claimed. In one big study, an assistant gave a link that led to a completely different news outlet’s video.

Names that look right can also be pure invention. Tests of AI-written computer code keep finding that several names in every hundred belong to software that does not exist. Web addresses fail the same way. The machine is writing text that fits the shape of an address, not copying a real one. Why AI makes things up covers the habit behind that.

In India, courts have thrown out filings built on judgments that never existed, and made the filing side pay for it. If it can happen to trained lawyers, it can happen to you.

So use this line every time the facts matter.

For every fact in your answer, add the source name and the link in brackets.
If a fact has no source, write "no source" next to it.
Then list the facts I should check myself on an official website.

Then do three things with each link. Open it. Look at the address of the site, not only the heading. Read the part that is supposed to carry your fact.

For a scheme, a form, a fee or a last date, the official site or the office is the answer. The AI is only your starting point. Checking sources, links and quotes turns this into a two minute habit.

How it reads a photo

A photo goes through the same kind of machinery. The picture is cut into small squares. Each square becomes a piece the model can read, much like the pieces it makes out of a sentence. Then it predicts, exactly as it does with words.

That explains what it is good and bad at. It can describe a scene well. It miscounts the things in it, because counting is not what it is doing. It misreads small print, handwriting and blurry corners.

So give it a good photo. Flat page, daylight, no shadow of your hand, one page at a time, cropped to what you need. Send a blank form or a book page where you can. If your name, your address or any number is printed on the paper, cover it before you take the photo. Then make it show its reading before it explains anything.

I am sending a photo of one page of a form.
Step 1: type out every word you can see in the photo.
Step 2: tell me any word you could not read clearly.
Step 3: explain in simple Hindi what this page is asking me for.

Step 1 is the whole trick. You can check those typed words against the paper in your hand. If the reading is wrong, the answer will be wrong too. Giving it your own book, notes or photo has more on this.

Photos of Hindi pages are the hardest case of all. How well an app reads Devanagari off a real printed page varies a lot between apps, so check the typed text every single time.

What you now know that most people do not

Most people believe the AI went and looked. You now know the truth. It wrote a slip, a program fetched a page, and that page was put in front of it. Search on means better facts. It does not mean checked facts.

Three habits carry all of this. Turn search on for anything with a date, a price or a rule. Open every link and read the part that matters. Make a photo answer type out what it sees before it explains.

A few wrong ideas about AI survive even after all this. The next lesson clears out four wrong ideas about AI, and the whole module finally clicks into place.