Somebody shares an image and asks, “is this AI?” The instinct is to look for the classic tells - too many fingers, garbled text, waxy skin - or to run it through one of the “AI detector” tools that have sprung up. Both instincts are reasonable and both have a shelf life. The visible tells are shrinking with every model release, and the detector tools carry a reliability problem that does not go away just because the interface looks confident.
This article covers what is actually true: the specific patterns where AI image tools still struggle, and why no detector - human or automated - can give you a guaranteed verdict on a single image.
The failure patterns that still hold up
As of mid-2026, a few categories remain worth a second look - not because they always fail, and not because avoiding them proves authenticity, but because they still show up often enough in generated or heavily edited images to justify a pause:
Hands and fine anatomy. Complex hand poses, overlapping fingers, and unusual anatomy still produce visible errors more often than faces or simple poses do, though frontier tools have improved a lot here.
Embedded text. Signage, labels, and text within an image remain uneven across tools and versions - some generations look clean; others still show near-words, inconsistent lettering, or text that is crisp in one area and smeared in another. Treat “looks readable” as weak evidence either way.
Precise counting and exact repetition. Ask a model to depict a specific number of identical objects and the count is often wrong. This applies to both generation and to a model’s own description of an image.
Consistency across multiple images or frames. A single generated image can look flawless; the same character or object generated twice, or across video frames, frequently drifts in small but noticeable ways.
Reflections, shadows, and physically consistent lighting. Generators produce plausible-looking light and shadow more often than physically correct light and shadow.
None of these are reliable enough to build a firm verdict on their own - a skilled or lucky generation can avoid all of them, and a genuine photo can contain an odd shadow or an awkward hand. Treat them as things worth a second look, not proof. They also age quickly: verify against current tools rather than treating this list as permanent.
Why “AI detector” tools cannot give you certainty
The clearest evidence for this pattern comes from OpenAI’s own admission about the equivalent problem in text. OpenAI released an AI Classifier for written text in January 2023 and later took it offline “due to its low rate of accuracy” - in OpenAI’s own published evaluation, the tool correctly flagged only 26% of AI-written text while incorrectly flagging 9% of genuinely human-written text as AI-generated (OpenAI, “New AI classifier for indicating AI-written text,” 31 January 2023; discontinued 20 July 2023). That specific number describes text, not images - but it is a clear, publicly admitted example of the underlying problem: a detector trained on one generation of models degrades against newer ones, and it fails in both directions, missing real AI content and falsely flagging real human content.
Image and video detectors face an analogous, ongoing arms-race problem for a structural reason: every improvement in detection can, in principle, be trained against by the next generation of image models, and there is no publicly available, independently verified detector that holds up reliably across all current generators, all edit types, and all compression or re-upload paths an image might travel through. Treat any specific accuracy percentage a detector vendor advertises as a claim about their test set on that date, not a guarantee about the image in front of you.
Do not use a detector tool’s score as the deciding factor in a consequential decision - a workplace dispute, a relationship conflict, a public accusation. A false positive can wrongly brand a genuine photo as fake; a false negative can wave through a convincing fabrication. Use detector output, if at all, as one weak input alongside everything else in this article - never as the verdict.
What actually helps, in order of usefulness
Provenance information, where it exists. A Content Credential under the C2PA standard can show who signed a file and what creation or editing history they asserted - genuinely stronger evidence than a detector’s guess, though it depends on the tool and platform supporting it and surviving the file’s journey to you. See Content Credentials and watermark basics for what that badge does and does not prove.
Finding the original source. The single most reliable check for most images is not analyzing the pixels at all - it is finding where the image first appeared, from whom, and when. A reverse image search or a direct search for the claimed event often surfaces the original context (or the absence of one) faster and more reliably than any visual analysis.
Consistency with what else is known. Does the claimed event show up anywhere else, from an independent source? Does the setting, clothing, or timestamp line up with what is otherwise verifiable? This is slower than a detector button but far more trustworthy.
The failure patterns above, as a supporting signal only. Worth a look, never worth a verdict on their own.
Where this applies beyond a single suspicious photo
This is the same reason synthetic media provenance and consent treats a Content Credential as useful-but-not-sufficient, and it is the direct answer to a common question in the adult deepfake first-response plan: do not delay reporting while trying to “prove” something is synthetic first. If you are generating images yourself and want a practical guide to what current tools do well and where their limits sit day to day, AI image generation 101 covers the tool-selection and prompting side of the same limits described here.
Common pitfalls
- Treating “no visible tells” as proof of authenticity. Absence of the classic errors listed above means a model avoided them, not that the image is genuine.
- Treating a detector’s percentage as a fact about the image. It is a fact about how that detector scored that image against its own training, on that date.
- Spending time “proving” synthetic origin before acting on something urgent. If a decision needs to happen now - reporting harmful content, for instance - do not let detector-chasing delay it. Why AI sometimes gives confident wrong answers covers the same overconfidence problem in a different context; the lesson generalizes.
- Assuming next year’s models will make this easier. The failure patterns shrink over time; the fundamental detection problem does not, because detection and generation are locked in the same arms race.
Try it this week
Next time an image makes you pause, resist the urge to run it through a detector app and trust the number. Instead: look for the original source first, check the specific failure patterns above as a secondary signal, and check for a Content Credential if the platform supports it. Use the AI image limits checklist to build the habit of checking the right things, in the right order.
