AI-generated images have become so convincing that spotting them by eye is now close to impossible. Hands look right, text mostly reads correctly, and lighting behaves. Meanwhile, the automated AI detectors marketed as a one-click answer are far less reliable than they claim. The good news is that verification still works. It just starts with the source, not the pixels. These six checks, in order, are what fact-checkers actually use in 2026.

Before inspecting a single pixel, ask a simpler question: where did this image come from? An image's origin tells you more than its appearance ever will. Was it posted by a verified outlet, a named photographer, or an official account, or by a brand-new profile with no history and a sensational caption?

Look for the earliest publication you can find. If a dramatic photo exists only on unknown accounts and no reputable newsroom is carrying it, doubt is warranted. If you can trace it to a known creator or a wire service, the mystery is usually solved before you examine anything else.

Many cameras, editors, and AI generators now attach tamper-evident provenance data using the C2PA Content Credentials standard, which records how a file was created and edited. If a valid credential is present, it is a genuine answer about the image's origin, and free verification tools can read these manifests straight from the file.

Google's invisible SynthID watermark is another planted signal. Since late 2025, users can check for it inside the Gemini app, and since May 2026, in Google Search as well, though this only detects content made with Google's own AI tools. A present watermark settles the question. A missing one, however, proves nothing, since most images online carry no watermark at all.

Upload the image or paste its URL into Google Lens or Bing Visual Search. This often reveals that the same picture existed years earlier, appeared in a completely different context, or was altered before being reposted with a new claim.

Reverse search is especially powerful against misattributed real photos. A genuine photograph with a false date, location, or explanation is still misleading, and finding the original context exposes the trick. It cannot prove an image was AI-generated, but it frequently proves the story around it is wrong.

Visual inspection is the least reliable check, but it still catches sloppy fakes. Zoom to full size and scan the details: garbled or warped lettering on signs and packaging, fused or extra fingers, earrings that do not match, shadows pointing in different directions, reflections that make no sense, and backgrounds that dissolve into mush behind a sharp subject.

Newer models make far fewer of these mistakes, so treat visual tells as supporting evidence only. Uncannily smooth skin or repeating textures can raise suspicion, but heavy retouching and phone HDR processing can make real photos look synthetic too. Never declare an image fake on visual clues alone.

Pixel-based classifiers sound authoritative, but independent testing tells a different story. One 2026 benchmark reportedly found leading tools accurate only 18 to 30 percent of the time against current generators, far below the 95 percent-plus figures vendors advertise on their own test data. Treat any score as one signal among many.

False positives are a real problem: heavily retouched photos, AI-upscaled images, and clean digital illustrations are frequently flagged as generated. Never use a detector score as proof that someone faked an image, especially in a school or workplace where the accusation carries real consequences. When several independent checks agree, confidence rises. A lone verdict means little.

Finally, step back and examine the motive. Does the post demand urgency, money, or outrage? Fake AI images are disproportionately deployed in scam ads and outrage bait. One analysis found hundreds of AI-generated news-presenter ads viewed over a million times on major social platforms. Sensational framing is itself a red flag, whatever the pixels show.

If you cannot verify an image after working through these checks, label it unverified rather than declaring it real or fake, and avoid using it as evidence for anything important. Preserve the original file in case qualified review becomes necessary. In 2026, honest uncertainty beats confident error every time.