Can You Trust What You See Online? | Lesson 2.10, Free Course

Why the page itself is a poor witness.

The idea

A website can look official and still mislead. Logos and web addresses are easy to fake. In one study, history experts and students who mostly stayed on the page were often fooled by them. An official-looking logo or web address is just the kind of feature that took them in. Professional fact-checkers learned most about a site by leaving it. They opened new tabs to see what others said about who runs it.
Checking a false story by searching for it can backfire. When people searched online to check false stories, they became more likely to rate them as true. Searching a false story's own headline turned up unreliable sites far more often than other searches did. Pasting in the headline of a startling story about a famous person is exactly that kind of search. So the four quick moves below search for the topic in new words, on sources you already know.
AI chatbots sometimes guess when they're unsure. A guess can look just like a real reference, with an author, a title, a journal and a year. In tests of earlier chatbots, many of the references they produced turned out not to exist. Newer ones may differ, so a reference still needs checking. A reference from a chatbot is a lead to check, not a source.
Voices and video can be faked as well. Criminals make short fake audio clips of a loved one's voice, posing as a relative in a crisis. Such fakes are often hard to spot, so official advice leans on checks instead. Calling back on a number you already have is one. A secret word agreed with your family, to check it's really them, is another. Fake live video calls from a 'boss' or other authority figure are another form these fakes take.

Want the research, names and dates? They are in Go deeper, at the end.

A smart uniform proves very little.

Picture it

Picture a stranger at your door in a smart uniform. The uniform tells you very little, because anyone can buy one. A call to the company they say they work for tells you much more. A website's look is its uniform. What others say about it is the phone call.

For a child: A note on your desk says, 'No homework tonight! From your teacher.' It even has a gold star on it. You only know a note is real once you check with the person it claims to come from.

Four quick moves, none of them on the page.

How it applies

It comes down to four quick moves. The first is a pause. Do you know the source? If you do, that may be enough. The second is finding out who's behind it, from other sites, in a new tab. The third is seeing what sources you already know and trust report about it. The last is tracing it to where it first appeared. A quote leads back to a book or speech, and a photo to its earliest copy. A number leads back to the report it came from. Other sites tell you more about a page than the page itself.

Four quick moves · Drawing: Hacks Vitae

Group chats, video apps and AI answers.

Where you’ll see it

In a group chat: a screenshot of a shocking headline, with no link to the story. Nothing says who wrote it.
On a video app: a famous face seems to recommend a product you have never heard of. The voice sounds just like theirs.
In an AI answer: a confident summary of a news event. Nothing shows where any of it came from.

Signs it’s happening

You can't tell who made it, or who runs the account that shared it.
A famous name with no date, book or speech attached is worth a second look.
A call or message from someone you know asks for money, and fast.

Same stirring line, two different next moves.

Before and after

A friend shares a stirring line credited to a famous scientist. It sits on a neat quote card with her photo.

The quick answer

"That sounds just like her." It gets shared again, name and all.

The second look

"Where and when did she say it?" No book, speech or date turns up. Without an original, there's no sign the scientist ever said it.

Real paintings, and words nobody wrote then.

A story from history: The Napoleon line that began as a cartoon

Napoleon spent his last years as a prisoner on St Helena, the island where he died in 1821.
In 2023 a line began spreading online as his: 'There is nothing we can do.'
It started in a TikTok photo slideshow posted on 15 August 2023. Two cartoon figures, drawn as Napoleon and one of his men on St Helena, trade two lines. One asks, 'Napoleon, what will we do?' The other answers, 'There's nothing we can do.'
The slideshow then cuts to real paintings of Napoleon on the island. Most later versions end on such a painting too. The first slideshow drew roughly 1.2 million plays in a month.
We found no letter, speech or memoir in which he says it. The line was written for two cartoon figures, about two hundred years after he died.

The real paintings lent a made-up line the look of history. Tracing it to where it first appeared showed what it was all along, dialogue for a meme, not a quotation. The feeling in the line belongs to 2023, not to the man in the painting.

Quick check3 questions · nothing is saved

1. In the study, what did fact-checkers do with an unfamiliar website?

2. What can searching a false story's exact headline do?

3. Where did 'There is nothing we can do' first appear?

One question worth keeping.

In one line

How a post looks says little. Who's behind it, and where it first appeared, says much more.

Which line have you passed on lately without knowing where it first appeared?

Go deeperThe research, names and dates

Where this comes from

Sam Wineburg and Sarah McGrew of Stanford University published a study in Teachers College Record in 2019. They watched 10 PhD historians, 10 professional fact checkers and 25 Stanford undergraduates evaluate real websites. Historians and students tended to stay within the site, and "often fell victim to easily manipulated features of websites, such as official-looking logos and domain names". Fact checkers left to see what others said. In the authors' words: "Fact checkers, in short, learned most about a site by leaving it." This habit is called lateral reading.
In five experiments published in Nature in 2024, Aslett and colleagues asked people to check news articles using online search. Searching increased the chance that people rated a false or misleading article as true. Searches of a false article's own headline or web address often returned unreliable links. That happened in 77% of those searches, against 21% of others.
A 2025 paper by Kalai and colleagues says large language models (the kind of AI behind chatbots) "sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty". In a 2023 study, 55% of the citations produced by GPT-3.5 were fabricated, and 18% of those from GPT-4. A 2023 study in the medical journal Cureus found 47% of ChatGPT's medical references fabricated. Only 7% were both real and accurate.
The FBI warns that "Criminals generate short audio clips containing a loved one's voice to impersonate a close relative in a crisis situation". It says such content "is often difficult to identify". Its advice: "Create a secret word or phrase with your family to verify their identity". The US Federal Trade Commission adds: "Call the person who supposedly contacted you and verify the story."

Staying on the page vs leaving it

Compared	Staying on the page	Leaving it
Who did it	Mostly the historians and students	The professional fact checkers
What they judged by	The site itself, its logos and web address	What other sites said about who runs it
What happened	Often fooled by official-looking logos and domain names	Learned most about a site by leaving it

From a Stanford study published in 2019, with 10 historians, 10 fact checkers and 25 students. Two of the ten historians behaved more like the fact checkers. The authors note that small samples exaggerate differences.

More examples

Some everyday cases, and the check that tells you more.

If a news site you don't know looks polished, what other sites say about who runs it tells you more.
If a cousin's voice note asks for money fast, calling back on a number you already have tells you more.
If a chatbot gives you a neat list of references, opening each source yourself tells you more.

What if…?

What if I've seen it many times before?: Seeing a claim again makes it feel truer, and that has been found reliably in the lab. One set of experiments found it even when people knew the right answer. How long the effect lasts is not settled.
What if most of the internet is bots now?: We found no measurement showing that most accounts on a major platform are bots. Some bot traffic is real and large, but some figures count page requests, not people. And an estimated 6 billion people used the internet in 2025.
What if I lean on an AI tool to check things for me?: Leaning too hard on automation is called automation bias. In one experiment, radiologists read mammograms with suggestions they were told came from AI. Some suggestions were deliberately wrong, and those pulled accuracy down at every level of experience.

Words you may meet

lateral reading: Checking an unfamiliar website by leaving it after a quick look and opening new tabs to see what other sources say about it. Glossary →

deepfake: A video, photo or audio recording that seems real but has been manipulated with artificial intelligence. Glossary →

misattribution: Wrongly saying that something was said or written by a particular person. Glossary →