Checking sources, links and quotes

You ask an AI about a scholarship in your district. The answer comes back neat and clear, with three links at the bottom. It looks checked, so you copy the whole thing into your family group without opening a single link.

Why the links are the most dangerous part

A wrong sentence with no source makes you suspicious. A wrong sentence with a source makes you relax. That is the trap.

A list of sources is a promise. It says somebody already went and looked. Often nobody did. The most convincing mistakes an AI makes are the ones dressed up as proof.

There are real numbers for this. A large study of free AI assistants was published in October 2025. It put 3,113 questions to the free versions of four popular assistants. Nearly half the answers had at least one serious problem. The biggest problem of all was sourcing, not raw facts.

The same study found something else worth knowing. Out of 3,113 questions, only 17 got a refusal. These tools almost never tell you they do not know. Silence about doubt is the default setting, not a sign of confidence.

You meet this even without opening an AI app. The answer box at the top of a normal Google search is AI too. One analysis published in April 2026 ran over 4,000 searches. It found about 9 of every 100 answer boxes wrong, and Google disputed the method. Either way, scroll past the box to the real links when the answer matters.

Trained lawyers in India have filed court papers quoting decisions that do not exist. Courts made them pay for it. If a lawyer can be fooled by a fake reference, so can you. You already met the basic checks in the two-chat test. This lesson is only about sources.

The three ways a source can be wrong

Only three things go wrong, and each one needs a different move from you.

First, the page is not there. The address was invented. Models build names and addresses that look right, because looking right is the whole job. That is the same habit you saw in why AI makes things up.

Second, the page opens, but it does not say that. This is the common one. The link is real. The line attached to it is not. In the study above, one assistant named a radio programme as its source. The link actually opened a video from a different news company, and the material used was comedy, not news.

Third, the page opens and says it, but it is nobody’s authority. A coaching blog is not the exam board. An old news copy is not the notice. For a scheme, a fee or a last date, only the official site or the office counts, as the schemes lesson explains.

The three-tap check

This takes under a minute for each link.

Tap one. Tap the link and let it open.

Tap two. See if it loads. A blank page, an error, or an unrelated homepage all count as a failure.

Tap three. Search inside the open page for one key word from the claim. Your browser menu has an option for this. Look for the words Find in page.

Pick that key word before you tap anything. It is usually a date, an amount, a name or an age limit.

Claim from the AI: the last date to apply is 30 September.
Key word to search for on the page: September

If the page does not contain your key word, the source does not support the claim. Do not argue with the app about it. Drop that line and move on.

Government notices often open as a PDF. Find in page may not work there. Then read the top of the first page. Check the date and the office name before you believe anything below it.

Read the end of a web address, not the start

Official Indian government sites end in gov.in or nic.in. Many state offices and departments use nic.in. Both endings are real.

The important word here is end. Read the address from the left until the first single slash. The part just before that slash tells you who really owns the page. A scam address puts a trusted word at the front and its own name at the end.

Both lines below are made-up examples. Neither one is a real site.

Safe shape:  scheme.gov.in/apply
Trick shape: scheme-gov-in.someothersite.com/apply

The first address truly ends in gov.in. The second one only starts that way. Its real owner is someothersite.com, and that tells you nothing good.

A familiar name is not proof either. One well known national scheme’s own web address was checked on 9 August 2026. It opened a foreign online gambling site. The address had not expired and it loaded without any error. So a name you have heard for years can still point somewhere else today.

Never type an address the AI gave you and stop there. Open it, look at what actually loaded, and read the end of the address in the bar. An app with search switched off cannot check any address at all, as the tools lesson shows.

Treat every quote as unverified

Suppose it gives you a line from a book, a speech, a court decision or a government notice. Treat that line as unverified until you have seen it in the original.

A quote is cheap to build and slow to check. The model works by guessing the next word, which the guessing lesson explains. A sentence that sounds like the author is exactly what it is good at making.

So ask for the exact place first.

Give me the exact place for that quote.
Book name, chapter and page number, if it is a book.
Notice number and date, if it is a government notice.
If you are not sure, write "not sure" instead of guessing.

Then open the original and look for the words. If you cannot see them there, do not use the quote. Never quote what you have not opened and read.

This matters more than it looks. A wrong quote in your answer sheet costs you marks. A wrong quote in an application or a complaint costs you much more.

Make it mark its own sources

You can ask for the marking inside the question itself. Copy this and put your own question under it.

Answer my question in short points.
After every fact, put the source in brackets.
If a fact has no source, write "no source" after it.
Do not guess a web address.
Use search and give me links I can open.

A good answer then looks something like this.

Admission forms open in June. (no source)
The college puts its fee on its own website. (link given)
The last date is 30 June. (no source)

Now you know exactly which lines to check. The lines marked no source are the ones to carry to the office or to the official site. Nothing here is proved yet. The marking only tells you where to point your three taps.

When the answer really matters, add one more line.

Only answer if you are sure.
If you are not sure, say "I am not sure" instead of guessing.

This is not a cure, and you should not treat it as one. It costs you one line, and it sometimes gets the app to admit doubt.

Search helps too, but it does not fix this. A test published in February 2026 asked hard questions with search switched on. Even the best setup left about 3 answers in 10 carrying a claim the cited page did not support. So search changes your job. The question is no longer only whether the claim is true. It becomes whether that page really says it.

Before you forward it to anyone

One rule covers everything above. Never forward anything you have not opened and read yourself.

Family groups move fast. One forward becomes fifty by night. If you pass on a scheme message with a bad link, your uncle will act on it, not question it. Nobody in that chain will check what you did not check.

Forwarding also puts your name on it. Inside your group, you are now the source. If it turns out wrong, the next person will not blame the app. They will stop believing you, and that is hard to win back.

Links are the easy case, because you can open them. Pictures, voices and videos are harder, and spotting fake images and videos in the safety module takes that up properly. Your next lesson gathers every check you now know onto one short page that fits in your notes app.