ChatGPT Vision: How to Use Image Analysis in 2026

The screenshot category was the quiet winner. We expected OCR to top the table, and it nearly did, but the most useful moments came from showing ChatGPT a broken settings screen or an error dialog and getting a fix in one reply. It reads layout, not just words, so the position of a button or the wording of an alert gives it clues that a typed description loses. Twelve out of twelve screenshots understood was the cleanest result of the whole test.

The failures were never loud. When ChatGPT misread a handwritten word or guessed at an unlabeled bar, it did so confidently and in passing, buried inside an otherwise correct answer. That is the real risk with image input. It is not that it fails often, it is that the failures look exactly like the successes. We started asking it to flag anything it was unsure about, and that one habit caught most of the quiet misreads.

Photo quality mattered more than we assumed. A flat, evenly lit shot of a page beat a slightly angled or shadowed one by a wide margin, especially for small text. Re-shooting a blurry receipt fixed more problems than any clever prompt. If a transcription comes back wrong, the fix is almost always a better photo, not a better question.

The privacy guardrails were consistent and worth knowing about up front. Asking it to name a private person in a photo was declined every time, politely and without exception, while a request to describe what the person was doing was answered normally. That line is deliberate, and it means image input is the wrong tool for any task that hinges on identifying who someone is.