Images: the flaws that still give it away (and the ones that don't)
For a couple of years, the star trick was counting fingers: early AI painted six-fingered hands and impossible wrists. That barely works anymore, because the models have improved exactly where they were most criticised. Even so, cracks remain. Look at the text inside the image — signs, labels, plates — where AI still writes meaningless scribbles; at reflections in mirrors and glass, which often don't match the scene; at jewellery, teeth and backgrounds, which warp or repeat patterns; and at that too-smooth skin, with a plastic sheen, that gives many faces a waxwork air.
The problem is that each of these flaws gets fixed with every new version, so the human eye is an ever-worse detective. Trust visual clues to suspect, never to conclude. It's the exact same rule we apply to the fake GTA 6 gameplay videos that flooded social media: the eye raises the suspicion, but the confirmation comes from elsewhere.
The real proof is in the invisible labels
This is where it gets serious, and where two names are worth knowing. The first is Content Credentials, the C2PA standard: a cryptographically signed label that travels inside the file and states who created it, with what tool and what edits it has undergone. You can check it by dragging the image into a C2PA viewer in your browser, without uploading anything anywhere. The second is SynthID, Google's invisible watermark: unseen, it survives screenshots and crops, and only its own detector recognises it.
Beware the big catch: most images circulating online carry neither, because social networks strip that data on upload. And no label doesn't mean the photo is real, only that nobody left a stamp. This will change fast, because from this very month Europe requires synthetic content to be marked; but the machine-readable mark has a reprieve until December, so for a while you'll still be leaning on your own judgement.
Text: why AI detectors lie to you
Now the uncomfortable part, because there's a whole industry selling snake oil here. AI text detectors — GPTZero and the like — are unreliable, and it's not just us saying it: OpenAI itself shut down its official detector for low accuracy, and these tools mostly fail by flagging human-written text as 'AI', with a cruel bias against people not writing in their native language. A failed grade or a firing based on one of these verdicts is an injustice waiting to happen.
What does help? Reading with an ear. AI tends to write too smoothly: clone-length paragraphs, zero typos, zero risky opinions, and empty filler like 'in today's fast-paced world' or 'it's important to note'. It lacks what a human has to spare: a concrete anecdote, an odd fact, a sentence only someone with something to say would write. If a text is flawless but says nothing you couldn't have guessed, be suspicious. And if it cites facts, check them: AI still invents sources with astonishing confidence.
The one method that doesn't expire
In the end, all the technical clues age: fingers get fixed, marks get stripped, detectors get it wrong. The only thing that holds is as old as journalism: look at the origin before the pixel. Who's publishing this? Does it appear in a source that stands behind it? Does the same image show up elsewhere with a different story attached? A reverse image search and thirty seconds of context catch more fakes than any miracle app.
So the next time something surprises you a little too much, don't just ask 'was this made by AI?', but 'who wants me to believe it, and why?'. Isn't that, in the end, the question we should have been asking long before the machines existed?
The conversation starts here
Sign in with a supporter account to comment. Sign in
Nobody has commented yet. Want to go first?