How to Spot AI-Generated Fake News

Learning how to spot AI-generated fake news starts with the claim, not the pixels. Before you examine a single shadow or finger, ask which newsroom gathered the story, on what date, under whose byline — because a generated image attached to a real event is a small problem, and a generated event dressed in a plausible image is the expensive one.

A decade ago a fake usually announced itself: a stolen masthead, a clumsy sentence, a real photograph given a new caption. That advantage is gone. Image models produce streets that look like places, language models produce paragraphs that sound like desks, and audio models can borrow a voice. This page is a reading habit for news claims and the media attached to them. It is not a forensics manual, and it does not cover the legal remedies available to someone deepfaked personally.

How do you spot AI-generated fake news?

Trace the claim to a newsroom before you evaluate the file. Four moves do most of the work, and none of them require looking closely at the image.

Ask who reported it, not who posted it. A newsroom with a name, a dateline, a reporter and a page you can open is doing one kind of work. An account with a dramatic still and no outlet is doing another. If the first place you saw the story cannot name the desk that gathered it, you do not have news yet. You have a file.

Look for a second newsroom. One dramatic clip is a rumor; two independent desks describing the same event, with their own pictures or their own quotes, is a story. Newsrooms covering the same event disagree about emphasis and rarely invent the event itself, which is the whole logic behind following world news across more than one country's press. A claim that exists in one place, in one medium, is the one to slow down on.

Separate the picture from the caption. Generated images are often married to a real headline, and real images are often married to a generated caption. Reverse the pair. Does the same photograph appear earlier, on another day, in another city? Does the caption name a place the background cannot support?

Prefer the publisher's own page. A screenshot of a homepage is trivial to fake. A live article on the outlet's domain, at a URL that matches, is harder. If someone wants you to believe a paper said something, go to the paper.

What still gives a generated image away?

The visual tells are real, they are secondary, and they expire. Poynter's MediaWise lesson on recognizing AI-generated content points at the same families of mistakes that working fact-checkers use: odd hands, strange reflections and shadows, and oddities in backgrounds, with a reverse image search as the actual test. The list below was checked in September 2026 and should be re-checked, because every generation of models retires part of it.

Text inside the picture. Signs, badges, license plates, a newspaper held up in a photograph. Generated type still collapses into almost-letters. A protest sign unreadable in a way that is not motion blur is worth a pause.

Hands, teeth, jewelry, crowds. Extra fingers are rarer than they were. What remains is a softness, a symmetry, or a crowd in which no face resolves into a person. Zoom in: a real crowd is full of ugly, specific detail.

Light that disagrees with itself. Two suns, a shadow falling the wrong way, a reflection showing a different room. Physics is still expensive for a model to keep consistent across a whole frame.

A place that is almost a place. Generated streets look like a memory of a city rather than a block you could walk — shop names that are not names, architecture from two continents. If the caption says a courthouse in Ohio and the building is a hybrid, it is not a wire photograph.

Too much finish. Real news photography is often slightly wrong: a bad crop, a blown highlight, a reporter's sleeve in frame. A glossy, emotionally on-the-nose image of a tragedy is not proof of anything, but it is a reason to ask who the photographer was.

What are the tells in audio, video and text?

Pictures are not the only surface, and the checks change by medium.

Voice. A short clip of a public figure saying something explosive is cheap to attempt. Ask for the original appearance: a full interview, a podium, a broadcast with more than ten seconds of context. Six isolated seconds under dramatic music is a format that favors the fake.

Video. Poynter's guide to detecting doctored or out-of-context videos, published in March 2023, sorts the problem into three kinds: real footage with false context, deceptive editing, and outright synthetic transformation. Its first instruction is not to stare at the face but to find the original source, then to treat unexplained, awkward gaps as evidence of editing. Ears, hairlines, glasses and the moment a head turns were still the weak points in generated footage when this page was checked in September 2026; like every artifact list here, that one expires.

Text. Language models produce confident, even copy, so the tell is rarely grammar. It is specifics that do not exist: a statute with the wrong name, a hospital that is not in that city, a quote no desk can place, a casualty figure that appears nowhere else. Search the distinctive phrase. If the only hits are copies of the same paragraph, you have a chain of repetition, not a confirmation.

Documents and homepages. Whole fake articles in the costume of a known masthead circulate as images, and a leaked memo with perfect letterhead and no originating organization is a prop. Check the live site, then check the URL.

The common move is the same across all four: do not argue with the artifact, leave it and look for the reporting.

What does a five-minute verification checklist look like?

Most of how to spot AI-generated fake news fits inside five minutes. You will not run a full investigation on a train, but you can run this.

1. Name the source. Which newsroom, which reporter, which date. If you cannot fill those in, stop here.
2. Open the original page. Not the screenshot, not the quoted post. The page.
3. Find one more desk. A second independent account of the same event. If none exists an hour into a supposedly enormous story, wait.
4. Uncouple the media from the words. Reverse image search the picture, or search a distinctive spoken line, without the caption attached. See what else it has been.
5. Decide what you are willing to pass on. If you cannot explain the claim without showing the file, you are not ready to share it.

Waiting is allowed, and it is the step people skip. The most expensive shares happen in the first twenty minutes, when the generated version is the only version available.

Why does source-checking beat any AI detector?

Because detection is probabilistic and provenance is not. Poynter's MediaWise material is explicit that detection tools are not fully accurate and that people still have to review the output themselves — which means a score can only ever add to a judgment you were going to have to make anyway. A newsroom, a byline and a live URL are checkable facts. A detector's confidence number is an estimate about a file.

The scale of the underlying problem is documented rather than hypothetical. Pew Research Center's collection of research on misinformation includes its October 2024 study of election news, in which about three-quarters of US adults said they had encountered inaccurate election news at least somewhat often. That is the environment the habit is for.

AI did not invent credulity; it made credulity faster and better looking. The defense is the old one, applied on purpose: newsrooms, names, second sources, original pages. It also explains why a social feed is a poor first draft of the world — a medium that presents reporting, opinion, jokes and synthetic media through the same gesture trains you to give them the same half-second.

The pixels are the appeal. The provenance is the evidence.

What can a news app honestly do about this?

It can refuse to hide the source, and it should not claim more. GetMyNews does not detect generated media, and any reader that says it does is selling a score.

What the app does is structural. Every card names the newsroom. It reads 92 public RSS feeds from 37 US newsrooms and does not rewrite the article, so tapping through opens the publisher's own page in Chrome, with your session — where a correction, if one was issued, sits under the original piece rather than in a copy we invented. When several outlets cover the same event, they collapse into one card that still shows who else is reporting it, which makes the second-desk check a glance rather than a search. That is not proof; it is a faster way to notice a claim with only one parent.

There is no account and no GetMyNews server, personalization runs on the device, and the deck is finite. The app is free on Google Play for Android, funded by a small banner plus one full-screen ad about every twenty stories. The rest of how to spot AI-generated fake news stays your job, which is the correct arrangement — a machine that sorted invention from truth on your behalf would be one more thing to verify. The reason to keep doing it is the one set out in why following the news is still worth the effort.

FAQ

Do AI-detection tools actually work on news images?

Partially, and not well enough to settle a question on their own. Poynter's MediaWise guidance states plainly that detection tools are not 100% accurate and that people still need to review the output themselves. Treat a detector's score as one weak signal beside stronger ones: a named newsroom, a dateline, a reverse image search, and a second desk reporting the same event.

Is a screenshot of a newspaper homepage ever evidence?

Rarely. A screenshot is an image of a page, and an image is the easiest artifact to fabricate or to take out of date. If a paper is said to have published something, open the paper's own domain and find the article at a live URL. Absence there, on a story large enough to be real, is itself informative.