Counting fingers or looking for garbled text used to be a real way to catch an AI-generated image — it doesn't work reliably anymore. Every visual flaw that became well-known enough to teach is exactly the flaw a model's next version gets built to fix.
For a while, this was genuinely good advice, not a myth. Early AI image tools were bad at specific, obvious things: hands with too many fingers or fingers that bent the wrong way, text that looked like writing but wasn't actually readable, skin that came out a little too smooth or waxy. None of it was subtle. Anyone could learn to spot it in a few minutes, and for a couple of years, that was often enough to catch a fake.
Here's the part that doesn't get said enough: those flaws were never a secret. They were the most obvious, most mocked, most talked-about problem with AI-generated images — which made them exactly the problem the companies building these tools had the strongest reason to fix first. A flaw that's famous is a flaw with a deadline.
That's roughly what happened. Hands got noticeably better. Text got better too, though it's not fully solved yet — it's just become the newest version of the same problem, and it's already improving the same way hands did. Whatever still looks slightly wrong in an image today is next on that list. It's not a permanent weakness anyone can count on staying there.
It's not only that the specific tells stopped working. It's that leaning on them at all teaches the wrong instinct.
If you check five fingers, count five, and treat that as your whole test, you walk away more confident than before you looked — confident in a photo you actually have no real reason to trust. That's worse than not checking anything. A photo you haven't checked leaves you appropriately unsure. A photo that "passed" a test that doesn't really work leaves you certain, and wrong. Wrong and certain is the dangerous combination, not wrong and unsure.
None of this means there's nothing you can do about an image — it means the checks worth doing aren't about how the image looks. Reverse image search tells you where else a photo has already appeared, and that doesn't depend on whichever tool made the image ever having had a visual flaw at all. What a message is actually asking you to do, and how much urgency it's leaning on, matters more than whether a picture looks a little off.
Both of those hold up no matter which version of which tool made the image, because neither one is testing the image. They're testing everything around it — and that's something you can actually get better at with practice, in a way that staring harder at a photo never was.