Check it yourself
AI can make fake photos, voices and videos that look real. Fake accounts can make a lie look popular.
You don't need a degree to deal with this. You need a few habits. They take a minute or two, and anyone can learn them.
This page won't tell you what to believe. It shows you how to check for yourself. That includes checking us.
A person has checked every link on this page.
Four steps
The SIFT method for checking what you read online was made by Mike Caulfield.
Where this comes from (3 sources)
This comes from university research
SIFT (The Four Moves) library guide
The library guide says its SIFT steps are adapted from the page by Mike Caulfield.
by Mike Caulfield, licensed under CC BY 4.0
Archived copy:Wayback Machine
A person has checked this link.
This comes from university research
The library guide says Mike Caulfield made the SIFT method.
The SIFT method was created by Mike Caulfield
Archived copy:Wayback Machine
A person has checked this link.
This is an original document from the time
Caulfield's own page lists the four moves that SIFT is built on.
The four moves: Stop, Investigate the source, find better coverage, trace the original context.
Archived copy:Wayback Machine
A person has checked this link.
We put the steps in our own words.
1. Stop
Don't share it yet. Ask yourself: do I know who made this?
Be extra careful if a post makes you really angry, or really happy. Those are the posts we most want to be true. Stopping costs you nothing.
2. Who made this?
Don't just read their "About" page. Anyone can write nice things about themselves.
Leave the page. Open a new tab and search their name. See what other people say about them. Who runs it? Who pays for it? What do they want you to do?
3. Who else is reporting it?
Search for the claim itself. See who else has it, and whether they checked it.
Watch for this trap: ten websites can all repeat one report. That is one source, not ten. Look for sources that found it out on their own and have different owners.
Different owners is a good start, but it does not prove two sources checked things on their own. Two news stories about the same report are still one report.
4. Find the original
A post about a study is not the study. A clip of a speech is not the whole speech.
Follow the trail back to where it started. Then ask: does the original really say that? Quotes and clips can change meaning when the part around them gets cut.
See the four steps used on a real story
We used these steps on a news story about bots going quiet. Real example: did the bots really go quiet?
Does this really work?
There is some good evidence, but it is small, so here is what it shows.
Researchers watched 45 people judge websites: 10 historians, 10 fact-checkers and 25 college students.
Where this comes from (2 sources)
This comes from a research journal
Lateral Reading and the Nature of Expertise (full paper)
The study watched 45 people: 10 historians, 10 fact checkers and 25 Stanford students.
we sampled 45 individuals: 10 Ph.D. historians, 10 professional fact checkers, and 25 Stanford University undergraduates
Archived copy:Wayback Machine
A person has checked this link.
This comes from a news report
Study on how fact-checkers read websites
Poynter reports the Stanford study watched 45 people judge websites.
monitored a group of 45 Ph.D. historians, professional fact-checkers and undergraduates
Archived copy:Wayback Machine
A person has checked this link.
Fact-checkers left the page early and looked up what others said about it. They got better answers, faster, than historians and college students who stayed on the page.
Where this comes from (2 sources)
This comes from a research journal
Lateral Reading and the Nature of Expertise (full paper)
Fact checkers read laterally, leaving the site to check it, and reached better conclusions in far less time.
Compared to the other groups, fact checkers arrived at more warranted conclusions in a fraction of the time.
Archived copy:Wayback Machine
A person has checked this link.
This comes from a news report
Study on how fact-checkers read websites
Fact-checkers were fastest because they left the page to look it up.
Fact-checkers found accurate information the fastest because they scanned webpages and Googled them before reading on.
Archived copy:Wayback Machine
A person has checked this link.
In one college class, 3 of 87 students checked a site by leaving it before a lesson. After the lesson, 67 of 87 did.
Where this comes from (2 sources)
This comes from a research journal
Lateral reading: College students learn to critically evaluate internet sources in an online course
Before the lessons 3 of 87 students left the site to check it. After the lessons 67 of 87 did.
At pretest, only 3 of 87 students engaged in lateral reading
Archived copy:Wayback Machine
A person has checked this link.
This comes from a news report
Free curriculum helping students evaluate online sources
Before the course 3 of 87 students read laterally. After it, 67 did.
only three out of 87 students engaged in lateral reading
Archived copy:Wayback Machine
A person has checked this link.
A study of Swedish high school students found that short lessons on spotting false stories had no lasting effect three months later.
Where this comes from (2 sources)
This comes from a research journal
Long-term effects of interventions against misinformation in upper-secondary school
The study followed Swedish students three months later and found no lasting effect.
the study found no long-term effects of the interventions
Archived copy:Wayback Machine
A person has checked this link.
This comes from a news report
Short lessons on disinformation have no long-term effect (Uppsala University news release)
The news release says short lessons did not help Swedish students later on.
short teaching interventions on disinformation have no long-term effect on upper secondary school students
Archived copy:Wayback Machine
A person has checked this link.
So it is a habit, not a trick you learn once. Keep using it.
Go further
- Go to the sourceLaws, court decisions and what was said in government.
- Photos and videoHow to check a picture.
- Has the page changed?Use the Wayback Machine.
- How to find proofA helper for the Prove us wrong form.
Don't trust "AI detector" tools
Some websites say they can tell if a person or an AI made something. Don't count on them.
For writing: OpenAI, the company that makes ChatGPT, shut down its own AI writing detector in 2023. It said the tool was not accurate enough.
Where this comes from (2 sources)
This comes from a news report
OpenAI pulls AI text detector due to low rate of accuracy
OpenAI took down its AI text detector in July 2023 because it was not accurate enough.
the AI classifier is no longer available due to its low rate of accuracy
Archived copy:Wayback Machine
A person has checked this link.
This comes from a news report
OpenAI admits that AI writing detectors do not work
OpenAI stopped offering its AI text detector in 2023.
OpenAI discontinued its AI Classifier, which was an experimental tool designed to detect AI-written text
Archived copy:Wayback Machine
A person has checked this link.
They can blame the wrong people: In one test, seven AI detectors looked at essays by people who learned English as a second language. Over half of those essays were wrongly marked as AI.
Where this comes from (2 sources)
This comes from a research journal
GPT detectors are biased against non-native English writers
Seven detectors wrongly marked over half of essays by non-native English writers as AI.
they incorrectly labeled more than half of the TOEFL essays
Archived copy:Wayback Machine
A person has checked this link.
This comes from university research
AI detectors biased against non-native English writers
The detectors marked 61 percent of essays by non-native English students as AI.
they classified more than half of TOEFL essays (61.22%) written by non-native English students as AI-generated
Archived copy:Wayback Machine
A person has checked this link.
For photos and video: Tools that spot fake photos and video can do well in a lab, then do much worse on real posts.
Where this comes from (3 sources)
This comes from a research or documentation centre
Is Artificial Intelligence Generated Image Detection a Solved Problem?
Image detectors that score well in the lab do much worse on real social media pictures.
despite their high reported accuracy in controlled settings, these detectors suffer significant performance drops on real-world data
Archived copy:Wayback Machine
A person has checked this link.
This comes from a research journal
Tests of 17 detectors on images shared across social media show clear limits in real-world use.
The benchmarking results reveal the limitations of current AI detection methods under real-world conditions
Archived copy:Wayback Machine
A person has checked this link.
This comes from a research or documentation centre
Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circulated in 2024
Deepfake detectors do far worse on fakes found on social media in 2024 than on older test sets.
the performance of open-source state-of-the-art deepfake detection models drops precipitously
Archived copy:Wayback Machine
A person has checked this link.
Some do better than others, but none gets the final word.
Use a detector as one clue at most. "Is this AI?" is the wrong first question. Ask: "Can I trace this back to the original?" A fact is true or false no matter what typed it.
Don't give up on facts
All this can feel like too much. You might start to think nothing can be trusted. That feeling is a trap.
People who lie want you tired.
Two law professors named the "liar's dividend." Once people know fakes exist, someone caught on a real recording can just say it is fake.
Where this comes from (2 sources)
This comes from a research journal
Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security
The law professors who wrote the paper named this the liar's dividend.
deep fakes make it easier for liars to avoid accountability for things that are in fact true
Archived copy:Wayback Machine
A person has checked this link.
This comes from a research or documentation centre
Deepfakes, Elections, and Shrinking the Liar's Dividend
The Brennan Center explains that once fakes are known, people can falsely say real footage is fake.
Law professors Bobby Chesney and Danielle Citron call this dynamic the liar's dividend
Archived copy:Wayback Machine
A person has checked this link.
So don't aim to trust nothing. Aim to check. Some sources earn more trust than others, and you can see which ones by checking them. You won't always get to the bottom of it. "I don't know yet" is an honest answer.
More to read
Every fact above has its sources under "Where this comes from." These pages go further.
- Mike Caulfield's own page on SIFT: hapgood.us
- The Stanford study's record: Stanford
- A second test of AI writing detectors, by European researchers: arXiv
More sources are on the pages for going to the source, photos and video, the Wayback Machine and the bots story.