Tell generated from real, or know when you can't

Somewhere in your feed this week, a generated image scrolled past you as a photo, and nothing about it warned you. The trick you already use, zooming in for warped hands and garbled signs, works on today's fakes and quietly expires as the models improve, and no alert will fire when it does. This lesson gives you the checks that don't expire: who posted it first, whether a reverse search finds an earlier original, who gains if you believe it, and how much checking the stakes deserve. It also legitimizes a verdict school doesn't offer, "can't tell", which is a real result with its own next moves, including the one that beats a cloned voice on the phone: hang up and call back on the number you already have. By the weekend you'll have run the whole procedure on three items from your own feed, and you'll walk away with a method you can defend instead of a feeling.

Somewhere in your feed this week, a generated image scrolled past you as a photo, and nothing about it warned you.

You zoom into an image and find nothing wrong. What did that establish?

Little. Well-made fakes pass the eye test, so a clean result leaves the verdict open and hands the job to provenance.

You zoom in and find garbled text on a background sign. What did that establish?

A lot. An artifact hit is strong evidence of generation today. Hits mean plenty, all-clears mean little.

Why does an artifact-only detection habit fail silently?

A fake that beats your eye gets filed as real, so you feel no failure while the miss rate climbs.

Reverse search finds the same photo from years earlier under a different caption. Verdict?

Real photo, false caption: the most common fake in circulation. The earlier dated original beats the newer story.

Reverse search finds no earlier version of an image. Verdict?

Open. A fresh real photo and a fresh fake both come back empty, so the weight shifts to the poster and the stakes.

The counter to a family member's voice on the phone asking for money?

Hang up and call back on the number you already have. The clone gets the voice right and has no way to answer the real phone.

What does a byline on a news photo buy you?

A named person and outlet who lose something if it's fake. Accountability, which a month-old aggregator account can't offer.

Is "can't tell" a failed analysis?

A finding. It ends the eye test and starts the provenance checks, and the stakes then set what you do.

How much verification does an item deserve?

In proportion to the cost of being wrong: a meme earns a shrug, a forward earns a thirty-second search, money earns a callback.

1. You zoom into a suspicious image and find nothing wrong: the background text is crisp, the shadows agree, the hands have five fingers. What did the inspection establish?: Little: well-made fakes pass the eye test, so the question stays open
2. Your class group chat shares a list: "How to spot AI images: check the hands, check the background text, check the reflections." What's the honest shelf life of that list?: It works on much of what circulates today and rots as models improve, so it needs a backup that doesn't depend on the image looking wrong
3. Your phone rings. It's your mother's voice, stressed: she's in trouble and needs you to transfer money in the next ten minutes. It sounds exactly like her. What settles it?: Hang up and call her back on the number saved in your phone
4. A dramatic image of a burning building is going around. You reverse-search it and find the same image on a news site from 2019, attached to a fire in a different country. What do you now know?: The image is real and this week's caption is false: a true photo recycled onto a new event
5. You reverse-search a different image and find nothing: no earlier versions, no matches anywhere. What's the verdict?: Open: a fresh real photo and a fresh fake both come back empty
6. Two accounts post the same dull photo of flood damage. One is "NewsFlashCentral," created last month, 400,000 followers from viral reposts. The other is a regional paper, photo credited to a named staff photographer. Why does the byline matter?: A named person at a named outlet loses a job, and the outlet loses standing, if the photo proves fabricated, so one of the two carries a cost for lying
7. A friend's rule: "If I can't verify an image, I treat it as fake." Where does the rule fail?: It converts "can't tell" into a verdict, when the honest version lets stakes set what you do