5 Ways to Tell If a Photo Was Made With AI
A practical checklist, not a party trick.
Most AI photos don't fail one big, obvious test. They fail several small, easy-to-miss ones.
Key takeaways
- No single visual clue reliably proves a photo is AI-generated anymore — modern models have fixed most of the obvious tells. The reliable approach is checking several signals together.
- Fine, high-frequency detail — individual fingers, teeth, small text, jewelry, and patterned fabric — is still where generative models most often slip, because these details require precise, non-repeating structure that's hard to render consistently.
- Blending seams around an edited or swapped region (especially faces) are one of the fastest tells, since color, sharpness, and lighting need to match perfectly at the boundary and rarely do under close inspection.
- Lighting and shadow direction inconsistencies between a subject and its background are a strong signal, because generative models don't always enforce one consistent physical light source across an entire scene.
- Content Credentials (the C2PA standard) is an emerging, harder-to-forge metadata system that records how an image was created and edited — a more durable long-term solution than any visual-inspection trick, though adoption is still uneven across platforms and tools.
- As generation quality keeps improving, the industry's own emphasis is shifting from "spot the flaw in the image" toward "verify where the image came from" — which is worth knowing before you rely too heavily on any single visual test.
A year or two ago, spotting an AI-generated photo was often as simple as counting fingers. That specific tell has mostly disappeared as models improved, and it's tempting to conclude there's no reliable way left to tell real from generated. That's not quite true — it's just that the useful tests have shifted from "one obvious giveaway" to "several smaller signals, checked together." Here are five that still hold up, roughly in order of how easy they are to check without any special tools.
1. Look at hands, teeth, and small repeating objects
Still the single best starting pointGenerative models learn to render faces and bodies extremely well because those are heavily represented in training data with fairly predictable structure. Fine, irregular, non-repeating detail is a different problem: individual fingers, the specific arrangement of teeth, small text on packaging, the links of a chain necklace, or the weave of a patterned fabric. These require the model to get a lot of small, precise decisions right simultaneously, and it's still the area most likely to show smearing, an extra or fused finger, teeth that don't quite align, or jewelry that dissolves into an abstract shape up close. Zoom in before you scroll past.
2. Check the edges where something might have been added or swapped
Especially useful for face swaps and compositesAs we covered in our breakdown of how AI face swap actually works, a generated or swapped face still has to be blended into the surrounding photo — color-matched, resolution-matched, and seamed in around the hairline, ears, and neck. This blending step is where a lot of otherwise convincing results fall apart under close inspection: look for a subtle halo or soft edge around the face, a mismatch in skin texture or grain between the face and the neck, or hair that looks slightly too smooth or slightly misaligned with the scalp. This is usually a faster and more reliable tell than trying to judge whether the face itself "looks real," since the face generation has often improved faster than the blending around it.
3. Check whether the lighting actually makes physical sense
A strong signal, and hard for AI to fake consistentlyIn a real photograph, every object in frame is lit by the same physical light sources, which means shadows fall in consistent directions and highlights land on consistent surfaces. Generative models don't always enforce that constraint across an entire scene the way a camera and physics do automatically. Look for a subject whose shadow falls in a different direction than the background's shadows, catchlights in the eyes that don't match the apparent direction of the light source, or a face lit more evenly and softly than a background that suggests harsher, more directional light. This is a genuinely difficult thing for current models to get consistently right, which makes it one of the more durable tells on this list.
4. Look for backgrounds that almost make sense
Check the parts of the image nobody was looking atBecause most attention — both a viewer's and, often, a model's own generation priority — goes to the main subject, backgrounds are frequently where generation artifacts survive uncorrected. Watch for text on signs or packaging that looks like letters but isn't actually legible words, architectural elements that don't logically connect (a railing that leads nowhere, a window frame that doesn't align with the wall behind it), or a repeating pattern — tile, brick, foliage — that subtly loses coherence or repeats in an unnatural way the farther it gets from the main subject.
5. Check the metadata — specifically, Content Credentials
The most durable long-term signal, when it's availableTraditional camera metadata (EXIF data — camera model, lens, shutter speed) can be stripped or faked fairly easily, so its absence isn't strong evidence either way. A newer, more robust standard called Content Credentials, built on the C2PA specification, is designed specifically to solve this: it cryptographically records how an image was created and what edits were made to it, in a way that's significantly harder to forge than traditional metadata. Where it's available — a growing number of cameras, editing tools, and AI generation platforms now support it — it's a much stronger signal than anything visual. The catch is adoption: not every platform preserves this metadata through upload and compression, and not every AI tool attaches it in the first place, so its absence today means "not verifiable" rather than "definitely fake."
A bonus signal: the "have I seen this face before" feeling
We wrote previously about why AI-generated characters in short dramas often feel oddly familiar — a side effect of generative models defaulting toward statistically common facial features when they aren't anchored to one specific reference photo. That same instinct is worth trusting here: if a face in a photo feels like a slightly-off composite of two or three people you can almost place, rather than one specific, memorable individual, that vague familiarity is itself a mild signal worth factoring in alongside the more concrete checks above.
| Check | Best for | How reliable today |
|---|---|---|
| Hands, teeth, small objects | Fully AI-generated images | Moderate — improving fast, but still the most common slip |
| Blending seams / edges | Face swaps and composites | High — blending still lags behind face generation quality |
| Lighting and shadow logic | Any AI image | High — hard for models to fully solve |
| Background coherence | Fully AI-generated scenes | Moderate — depends heavily on the model and prompt |
| Content Credentials metadata | Any image, when present | Very high when available — but not always present |
What none of these tests guarantee
It's worth being honest about the limits here: none of these checks are foolproof, and as generation quality keeps improving, purely visual tells will keep getting harder to spot. That's exactly why the industry's broader focus is shifting from "can a person spot the flaw" toward "can the image prove where it came from" — provenance and disclosure standards like Content Credentials, rather than an ever-escalating game of finding the next visual giveaway. Treat the five checks above as a useful, practical first pass, not a certainty machine.
A quick glossary
- EXIF data
- Standard metadata embedded by cameras and phones (camera model, settings, sometimes location), which can be stripped or edited and isn't strong evidence on its own.
- C2PA / Content Credentials
- An industry standard for cryptographically recording an image's creation and edit history in a way that's significantly harder to forge than traditional metadata.
- Provenance
- Verifiable information about where a piece of content actually came from and what's happened to it since, as opposed to judgments based on how the content itself looks.
Frequently asked questions
What's the single most reliable way to tell if a photo is AI-generated?
No single tell is fully reliable on its own, since AI image quality keeps improving. The most dependable approach is combining several signals — fine detail in hands and small objects, edge and lighting consistency, background logic, and metadata like Content Credentials — rather than relying on any one clue.
Can I trust a photo's metadata to prove it's real?
Metadata is a useful signal but not proof by itself. Standard camera metadata (EXIF) can be stripped or edited, and while newer Content Credentials (C2PA) metadata is harder to forge, not all platforms preserve it and not all AI tools include it yet.
Will AI detection tools always be able to catch AI-generated photos?
Not reliably forever. As generation quality improves, purely visual detection gets harder, which is part of why the industry is shifting toward provenance standards like Content Credentials — proving where an image came from, rather than trying to spot flaws in the image itself.
Do AI face swap apps make images harder or easier to detect as AI-generated?
Easier, in most cases, because a face swap only needs to make one region of an existing real photo convincing rather than generating an entire scene from scratch. Blending seams around the swapped face are usually the fastest tell.