The fix was never a better eye. It’s a better habit.
GeneralGE
If you've been training your eye (counting fingers, hunting for warped text, looking for that plasticky AI sheen), you've been preparing for the last war. Those tells worked for a while. Most of them don't anymore. The generators read the same critiques you did and closed the gaps.
So here's the uncomfortable reframe: the people who are hardest to fool online were never the ones with the sharpest eyes. They were the ones with the most boring habits. That's the good news. Boring habits are learnable, and they don't expire every time someone ships a new model.
The thing the best fact-checkers do that you don't#
In 2017, the Stanford History Education Group ran a study that should be more famous than it is. Researchers gave the same online-evaluation tasks to three groups: ten PhD historians, ten professional fact-checkers, and twenty-five Stanford undergraduates. Smart people, all of them.
The fact-checkers crushed the other two groups, and they did it faster.
The difference wasn't intelligence or credentials. It was where they pointed their attention. Historians and students read vertically: they stayed on the page, studied the site's design, scrolled its "About" section, weighed its prose, and tried to reason their way to a verdict from inside the thing they were evaluating. The fact-checkers read laterally. They glanced at the page, then immediately left it, opening new tabs to see what the rest of the internet said about this source before forming any opinion of their own.
That's the whole trick. Sam Wineburg and Sarah McGrew, the researchers who named it, summed up the lesson in a phrase: you learn more by reading less of the page in front of you and more around it.
It matters more now than it did in 2017, because the page in front of you is exactly what AI is best at faking. A convincing site, a confident byline, a clean layout, a photo that survives any amount of squinting: those are cheap to generate. What's still expensive to fake is a track record that independent, unrelated sources will vouch for. Lateral reading puts your trust where the faking is hardest.
Lateral reading is the instinct. SIFT is the checklist that makes it a habit. It comes from Mike Caulfield, a digital-literacy educator who packaged the practice into four moves in a 2019 post (he'd been refining the idea for years before that). The four moves are:
Before you read, react, or share, notice the pull. The content engineered to spread is engineered to spike your pulse first and your judgment second. If a headline or image makes you feel something strong and fast, that's the moment to slow down, not speed up. Stop also means: do you even need to evaluate this, or can you just close the tab?
Spend thirty seconds finding out who's actually talking before you weigh what they're saying. Crucially, you don't do this on their site; you do it everywhere else. Open a tab. See how other outlets describe them. A confident "About" page is theater; what strangers with no incentive say about a source is data.
You usually care about the claim, not the page. So go find the best available reporting on that claim, even if it means abandoning the source you started with. If a surprising assertion appears in exactly one place and nowhere else, that absence is itself information. Real events leave footprints across independent outlets.
A quote, a statistic, a clip, a photo: follow it back upstream to where it started. Things get mangled, stripped of context, or invented entirely in the retelling. The original almost always says something more boring, more qualified, or more damning than the version that reached you.
None of these moves requires you to be a detector. They route around the problem. You stop asking "does this look real?" (a question modern generators are designed to win) and start asking "what does the rest of the world say about this?", which they can't.
Detection-by-inspection is in structural decline, and it's worth being honest about why. The visual tells keep degrading as models improve. We cover the moving target in why the old tells are getting harder. Automated detectors that promise to flag synthetic content are unreliable enough that we'd warn you off leaning on them: the reasons are in why AI detectors fail. And the deeper question of whether "real" is even a clean category anymore (given that a phone photo is already a stack of computational guesses) is its own rabbit hole in what counts as real.
Provenance is the one technical thread genuinely worth pulling, and it's improving. Content Credentials (the open C2PA standard backed by Adobe, Google, and others) let a camera or generator cryptographically sign a file with its origin. By 2026 it's shipping in real hardware: Google's Pixel 10 signs photos by default, and Canon and others are rolling out credential-enabled workflows for newsrooms.
But provenance is a clue, not a verdict, and here's the catch that keeps it humble: most platforms strip that metadata the moment you upload, so signed content routinely arrives stripped of its signature. A missing credential tells you almost nothing. Credentials can also be revoked: Nikon had to invalidate a batch after a signing flaw. Even regulation, like the EU AI Act's labeling rules that took effect on 2 August 2026, is enforced unevenly and won't reach everything you see. So provenance belongs in your toolkit as one corroborating signal, never as the single thing you trust.
Which lands us back where the fact-checkers were all along: don't trust the artifact, trust the corroboration. Source reputation, independent coverage, and the original context outlast every generator, because they live outside the file entirely.
Leave the page. Find out who's talking and what unrelated sources say about them. Follow the claim to its origin. Then decide.
It's slower than glancing and guessing. That slowness is the entire point. The reflex this defeats is the one every fake is built to exploit: the snap judgment. Build the habit on low-stakes stuff and it's there when it counts.
That, honestly, is what a daily round of the Spotdle puzzle is quietly training: the pause before the verdict. The eye is optional. The habit is everything.
Doesn't lateral reading just mean Googling it?
More or less, but with discipline. The skill is leaving the page before you've decided what to think, opening other tabs, and judging the source by what independent outlets say about it, not by how polished its own site looks. The polish is the part the AI can fake. The reputation isn't.
Can I still trust a photo if it has a Content Credential?
A credential is useful evidence, not a verdict. It tells you a signed claim about origin exists. But most platforms strip that metadata on upload, the absence of a credential proves nothing, and credentials can be faked or revoked. Treat it as one corroborating signal among several, not as proof.