Field note · Everyday AI

How to Tell If a Photo Is AI-Generated (2026)

Published July 2, 2026 · Vita Indarra

Short answer: Sometimes you can spot a fake by its flaws — garbled text, wrong hands, impossible lighting. But the best fakes now have no flaws you can see, so the honest rule for 2026 is this: pixel-checking catches the lazy fakes; source-checking catches the good ones. Learn the tells, but build the habit.

The visual tells that still (sometimes) work

Cheap and fast generations still leak the classic giveaways. Worth thirty seconds before you share or believe:

• Text in the image. Signs, labels, jerseys, license plates — AI still garbles lettering more than anything else. Zoom in. Alien alphabets and melted logos are near-certain tells.
• Hands, teeth, ears, jewelry. Better than they were, still the most common failure sites: six fingers, fused knuckles, earrings that merge into skin.
• Physics that's almost right. Reflections that don't match the scene, shadows falling two directions, glasses bending a background wrong.
• The texture of too-perfect. Waxy skin, hair that dissolves where it meets the background, fabric patterns that repeat like wallpaper.
• Background people and objects. The subject gets the model's full attention; the crowd behind them often melts — faces smeared, limbs miscounted.

Here's the part most guides won't tell you: a clean image proves nothing. These tells appear in careless generations. Top-tier 2026 generators clear every item on that list routinely. Treating "I looked closely and it seems fine" as verification is exactly the instinct fakes are built to exploit.

Why detector tools can't save you either

Paste-your-image "AI detectors" exist, and they are weak evidence at best. They're in a permanent arms race with generators; they miss well-made fakes; and they falsely accuse real photos — especially ones that have been edited, compressed, or screenshotted, which is most of what circulates. Use one if you like, but treat its verdict as one soft signal, never a ruling. A tool that's confidently wrong in both directions is not a safety layer — it's a coin with opinions.

The habit that doesn't expire: check the source, not the pixels

Visual tells decay every time the models improve. Provenance doesn't. Three checks, in order of effort:

• 1. Who's showing me this? An image of a real event should exist beyond one anonymous account. If a photo of something newsworthy appears nowhere reputable, that absence is the evidence.
• 2. Reverse image search. Search the image itself (right-click on desktop, or your search app's camera icon). You're looking for the earliest version and its original context — the most common "fake" isn't generated at all; it's a real photo relabeled from a different year or country.
• 3. Wait for corroboration on anything that matters. Real events accumulate photos from multiple angles and witnesses within hours. Fakes are usually a single image, everywhere at once, from nowhere in particular.

One more layer is slowly arriving: Content Credentials — tamper-evident labels some cameras and AI tools now attach, recording how an image was made. When present, genuinely useful. But adoption is patchy and labels can be stripped, so read it one-directionally: presence is helpful evidence; absence tells you nothing.

The mindset shift that actually protects you

The deep change in 2026 isn't that fakes exist — it's that "seeing is believing" has quietly stopped being a rule you can live by. The protective instinct isn't sharper eyes; it's a small pause: where did this come from, and who else says so? That's a thirty-second habit, it works on images your eyes can't crack, and it will still work next year when the pixels are perfect.

Frequently asked

Can I tell just by looking?

Sometimes — the tells above catch lazy fakes. But a flawless image proves nothing; the good ones have no visible seams.

Do AI-image detectors work?

Partially. They miss good fakes and falsely flag real photos. Weak signal, never a verdict.

So what's actually reliable?

Source and corroboration: who posted it, what a reverse image search finds, and whether anyone trustworthy carries it. Provenance outlives pixels.

← More field notes