In March 2023, a photo of Pope Francis wearing a bright white puffer jacket spread across social media. It looked completely real. People shared it as a genuinely funny, harmless moment: the Pope, dressed like he was heading to a ski resort. It was entirely fake, generated with the AI tool Midjourney by someone who was mostly just testing what the software could do.
Almost nobody who shared it stopped to question it, because there was no obvious reason to. That's the actual lesson here. The danger was never that AI images look fake. It's that, increasingly, they don't.
How much of what you see now is AI
The volume is hard to picture. Back in August 2023, less than a year after these tools went mainstream, Everypixel Journal estimated that around 34 million AI-generated images were already being created every single day. That remains one of the few public figures anyone has actually published, and the tools have only got faster and cheaper since. Be a little wary of the bigger numbers floating around: several widely repeated "current" figures turn out to be attributed to sources that never published them.
Text is easier to measure. A study by Graphite analysed over 65,000 randomly sampled web articles and found that the share written mainly by AI climbed from roughly 36% in the year after ChatGPT launched to about half by early 2025, where it has essentially stayed since. A large share of what you read online was never written by a person.
Worth knowing how that was measured, though: figures like this come from AI detectors, which are useful in bulk but imperfect on any single article. The broad pattern held across three different detectors, which is why it's credible.
When it stops being funny
A puffer jacket photo is a harmless curiosity. Two real cases show what happens when the same technology gets aimed at something that actually matters.
In early 2024, a finance employee at the Hong Kong office of the engineering firm Arup received an emailed request for a confidential transfer. He was suspicious, which is exactly what you'd hope. So he joined a video call with the company's CFO and several colleagues he recognised, and his suspicion evaporated. Every other person on that call was an AI-generated deepfake, built from public video and audio of the real executives. He approved a series of transfers totalling roughly $25 million. Hong Kong police made the case public that February.
In August 2025, an Airbnb host accused a guest of causing around $16,000 in property damage and submitted photos as proof. The guest noticed something odd: the "same" cracked coffee table looked subtly different across supposedly identical photos, which she argued isn't possible in genuine, unedited pictures of one object. Airbnb initially sided with the host. It reversed the decision, refunded her stay and apologised only after a journalist at the Guardian took up the case.
Same technology, wildly different stakes. What connects both stories is that the fake content didn't look fake. It looked exactly convincing enough to be believed by someone with no particular reason for suspicion, and in the Arup case, by someone who was actively suspicious already.
The risk was never obviously fake AI content. It's the content that gives you no reason to double-check at all.
How to actually spot one
No trick works forever as the tools improve, but a few habits still catch a meaningful share of fakes today:
- Check for inconsistency across copies. If the same object or scene appears more than once, compare the details closely. AI often can't keep small details identical, which is exactly what unravelled the Airbnb claim.
- Look at hands, ears, and text. These remain some of the hardest things for image generators to get consistently right.
- Ask where it actually came from. A reverse image search or a quick look for the original source catches a surprising number of fakes in seconds.
- Notice content built to trigger a strong, fast reaction. Outrage, awe, and shock all short-circuit the instinct to pause and verify, which is exactly why manipulated content is often built around them.
- Don't rely on "it just looks wrong." In studies of AI-generated faces, untrained participants have scored below chance, rating synthetic faces as more real than genuine ones. Your gut is a weaker filter than you'd think.
Detection tricks age fast. Every visible flaw in AI-generated content today is a target the next model version is specifically trained to fix. Treat any list of "tells," including this one, as temporarily useful, not permanently reliable.
The real skill isn't detection
Pixel-hunting for glitches is already a losing long-term strategy, because the glitches keep disappearing. The habit that actually holds up is asking a different question entirely: where did this come from, and can I verify it somewhere independent of the post itself? That question works on a fake photo, a fake quote, and a fake statistic equally well, and it doesn't expire the way a list of visual tells does.
It's the same habit we wrote about in why AI makes things up: the problem isn't the technology being obviously wrong, it's the technology being confidently plausible. The fix in both cases is a source, not a gut feeling.
None of this means every image or article you encounter deserves suspicion. It means saving that extra thirty seconds of verification for the moments it actually matters: money, health, safety, or anything you're about to act on or repeat to someone else.
That instinct, pause and verify before you act or share, is most of what we mean by media literacy in the age of AI. Not paranoia. Just a habit worth building before you need it.