Anyone can now produce a confident paragraph, a narrated video or a realistic picture in seconds. The tools do not know whether what they made is true. That is the quiet change of the last three years: the cost of making something that looks like knowledge has fallen to almost nothing, while the cost of checking it has not moved. Fact-checked content is becoming the scarce thing, and the habits that find it are worth learning.
AI-generated content is fluent, not fact-checked
A language model predicts likely words. It is very good at the shape of a true sentence: the dates look like dates, the numbers look plausible, the tone is assured. When it has no real source, it fills the gap with the most probable guess, which readers meet as an AI hallucination: a wrong date, an invented quote, a study that does not exist. The sentence reads exactly the same either way. Fluency used to be a weak signal of expertise. It is no longer a signal at all.
The same goes for pictures and video. A photo that was never taken and a map that was never surveyed can look more polished than the real thing. Misinformation no longer needs a motive; it can be an accident of a tool that was only asked to be convincing.
How to verify information online in one minute
You do not need to become a researcher. Three questions catch most of what matters.
- Where does this come from? A claim without a named source is an opinion in a costume. Look for a link to a primary source: the museum, the survey, the agency, the paper, the archive. If the only source is another summary, you have not reached the bottom yet.
- Does the source say that? Click through. The most common failure is not a fake source but a real one that says something slightly different: a smaller number, a hedged claim, a different year. Reliable sources are quoted, not paraphrased into something stronger.
- What does the number mean? A figure is only as good as its unit and its comparison. Taller from where, bigger than what, measured when. A number that cannot be traced to a measurement is a decoration.
Primary sources are the new premium
The institutions that measure and record the world did not stop working because text got cheap. UNESCO still documents heritage sites, USGS and NOAA still publish their surveys, museums still catalogue what they hold, and Wikipedia still cites its references in public where anyone can check them. The difference is that their work now has to be found on purpose, under a growing layer of content that merely sounds like it. Media literacy in this decade is mostly the habit of digging to that layer and staying there.
What we do about it
Culturix is a video encyclopedia of the world with a sense of humour, and the humour only works if the facts underneath are solid. So every video in the World feed is built from a source you can open: the page it came from is linked on the video, the transcript is there to read, and the numbers spoken are the numbers shown. When a figure is contested, we say so and give the highest credibly sourced estimate with its author rather than picking a tidy one.
We use AI tools to make the videos, and we say so. Narration is synthetic, some illustrations are generated, and they are labelled as such on the platforms that ask for it. What we do not let the tools do is decide what is true. New videos go through a review that checks each spoken claim against the source before they are published, and Not So Fast exists specifically to take popular beliefs apart with a named source in hand.
A simple rule for readers and makers
If a claim cannot be traced to a source, treat it as unverified, however good it sounds, and however many times you have seen it. If you make things for other people, link your sources where they can be seen, not in a footer nobody opens. Trustworthy content in the age of AI is not content that avoids the tools; it is content that shows its work. Start with the World feed and open any source link. That is the whole method, and it is yours to use everywhere.