How to Check If a Photo Is AI Generated: A 5-Point Visual Checklist
Make this article actionable
Send the article context into Vife Agent and turn it into a plan, checklist, or draft you can keep working on.
Quick answer
There is no single test that proves a photo is AI generated. Detector scores are guesses, metadata can be stripped or faked, and a clean-looking image can still be synthetic. What you can build is a repeatable habit: run the same five visual checks every time, search for an earlier source, treat metadata as a weak hint, and then decide how to label the image before you repost it.
That habit takes about two minutes. It will not give you a verdict. It will give you a reason to slow down, which is usually the point.
Turn the useful parts into next steps
Vife Agent can convert this guide into a prioritized workflow with tasks, risks, and reusable prompts.
Why one signal is never proof
Most people make one of two mistakes. They trust the image because it looks fine, or they paste it into a single "AI detector" and accept whatever number comes back.
Both fail for the same reason: every signal is circumstantial.
- Visual artifacts are real but inconsistent. Modern generators handle hands, text, and reflections far better than they did a couple of years ago, so "no visible glitches" is weak evidence of authenticity.
- Detector scores are probabilistic outputs from a model you cannot inspect. A high score is a prompt to look closer, not a finding.
- Metadata (EXIF fields like camera make, lens, and capture time) is easy to remove. Most social platforms strip it on upload, and it can be edited. Its absence tells you almost nothing; its presence is only mildly reassuring.
- Reverse image search proves a photo existed earlier somewhere, which is useful — but an AI image posted months ago will also turn up in old results.
The practical takeaway: collect several independent observations, and let them point the same direction. One oddity is noise. Four oddities in the same image is a pattern.
The 5-point visual checklist
Work through these in order on a full-size version of the image, not a thumbnail. Zoom in. Thumbnails hide exactly the details you need.
1. Count the repeated small things
Hands, fingers, handles, chair legs, buttons, spokes, earrings, and teeth are where generators most often slip.
What a failure looks like: a mug with two handles, a hand with six fingers or a finger that merges into the cup, a chair with three legs on one side, a row of buttons where one is half-formed, a necklace that dissolves into the collar.
What normal looks like: hands can be partly hidden or blurred by motion. A missing finger behind a cup is not evidence of anything. You are looking for structural errors — extra parts, merged parts, parts that connect to nothing.
2. Look for melted or asymmetric edges
Trace the outline of the main subject slowly with your eyes.
What a failure looks like: an edge that softens and smears into the background instead of ending, a rim that wobbles where it should be a clean circle, a handle that fuses into the table, a shadow that has no defined boundary, a straight line that bends for no reason.
What normal looks like: shallow depth of field also softens edges. The tell is inconsistency — a crisp edge on one side of the same object and a smeared edge on the other.
3. Read the background text and reflections
This is the highest-yield check for everyday images, because text and mirrors are hard to fake convincingly.
What a failure looks like: letters that are almost-words, signage in an invented alphabet, a label whose text curves away from the surface it sits on, a reflection in a window or spoon that shows a different scene than the one in front of it, a reflection of an object that is not in the frame.
What normal looks like: real photos also contain unreadable text — motion blur, distance, and odd angles all do that. The question is whether the text is shaped like writing or shaped like letter-ish texture.
4. Check that shadows agree
Pick two or three shadows in the image and compare their direction and softness.
What a failure looks like: the coffee cup casts a shadow to the left while the chair casts one to the right; a shadow that is sharp under a soft light source; an object with no contact shadow at all, so it appears to float; a highlight on the cup that implies a light source the shadows contradict.
What normal looks like: multiple light sources, bounce light off a wall, and overhead lamps all create mixed shadows in real rooms. Look for shadows that are impossible, not merely varied.
5. Inspect repeated patterns for suspicious perfection
Tiles, floorboards, leaves, brickwork, fabric weave, and window grids.
What a failure looks like: every tile identical down to the same smudge, a pattern that repeats on a grid with no perspective change, leaves that are clones of each other, a fabric texture that turns into noise in one corner.
What normal looks like: real repetition drifts. Grout lines vary, one tile is chipped, perspective compresses the far end of the pattern. Perfect uniformity across a whole surface is unusual in a photograph.
Illustrative walkthrough: a coffee cup on a table
Take a generic, unbranded photo of a coffee cup on a table. Here is how the checklist reads on a hypothetical example:
| Check | What you inspect | Possible red flag |
|---|---|---|
Handles and hands | Handle count, any hand holding the cup | Two handles, or fingers merging into ceramic |
Edges | Rim circle, handle join, cup-to-table contact | Wobbly rim, handle fused into the tabletop |
Text and reflections | Label text, reflection in the coffee surface or a nearby window | Letter-shaped nonsense, reflection showing a different room |
Shadows | Cup shadow vs. chair and table-leg shadows | Shadows pointing in opposite directions |
Patterns | Table grain, tile grout, fabric of a napkin | Grain that repeats identically with no perspective shift |
Two flags in one image is worth pausing over. Five flags is worth not reposting.
Step 2: Search for an earlier source
Reverse image search is the most concrete step available to you, because it deals in dates and pages rather than impressions.
- Crop or save the image at full resolution.
- Run it through a reverse image search engine.
- Sort or scan results for the oldest match you can find.
- Open that page and check whether the image is presented as a photograph with a plausible context — a photographer's portfolio, a news article, a product listing, a personal post.
- Note the date. An image that appears in 2019 with a photographer credited is a different situation from one that first appears in a stock-photo dump last month.
What this proves: that the image existed at a certain time and place. What it does not prove: that the image is authentic. A synthetic image uploaded years ago will still show up as an old result. And a genuine photo can be reposted with a false caption, which is a separate problem from whether it was generated.
Step 3: Read metadata as a weak hint only
If you can view EXIF data — many operating systems show it in file properties, and some browsers expose it via an extension — look for camera make and model, lens, exposure settings, and a capture timestamp.
Treat it like this:
- Present and internally consistent (a real camera model, plausible settings, a timestamp that matches the file's other dates): mildly reassuring, nothing more.
- Absent: tells you almost nothing. Social platforms strip metadata routinely, and screenshots never had it.
- Present but strange (editing software listed as the only origin, a timestamp that predates the camera model, fields that contradict each other): a reason to look harder, not a conclusion.
Do not build a decision on metadata alone. It is the easiest signal to remove and one of the easier ones to alter.
Step 4: What to do next
Once you have your observations, act on the uncertainty rather than resolving it.
- Ask the sender. "Where did this come from?" is a fair, fast question. Someone who made it with a generator will often just say so.
- Label uncertainty in your own post. If you share it, use plain language: "source unverified" or "I could not confirm where this image originated." Do not write "AI generated" as a fact unless you have a source that says so.
- Avoid presenting it as verified. Do not add a caption that asserts a location, a date, or a person's identity the image does not establish.
- Keep the original file. If you need to revisit the question later, a re-encoded or screenshotted copy loses whatever metadata survived.
- Escalate only when it matters. For a meme, a label is enough. For anything involving a person's reputation, a product claim, or a safety issue, treat it as unverified until you have a source you can name.
Common mistakes
- Trusting a single detector score. A number is not a finding. Use it as a prompt to run the checklist.
- Judging from a thumbnail. Compression and downscaling erase the edges, text, and patterns you need.
- Treating "no visible glitches" as proof of authenticity. Absence of artifacts is weak evidence, especially in newer images.
- Confusing a real photo with a false caption. Those are two different problems; solve the one you actually have.
- Assuming missing metadata means synthetic. It usually means the platform stripped it.
- Over-claiming in your own caption. "This looks AI generated to me" is honest. "This is AI generated" is a claim you probably cannot support.
If you need to produce or edit images yourself
The same attention to edges, text, and shadows applies when you are the one making the image. If you are cleaning up a photo, compositing a product shot, or generating a background to place behind a real subject, the review criteria above double as a self-check before you publish. You can do that kind of work in Vife's AI image editor — and if you are weighing plan options for a small team, the details are on the pricing page.
A copyable review prompt
If you keep a notes file or a shared doc for your team, this is a reasonable template to paste in and fill out. It is a prompt for a human reviewer, not an automated test, and it does not guarantee a correct answer.
Image review — [filename or link]
Reviewer: [name] Date: [date]
1. Repeated small details (hands, handles, buttons, legs):
Observed:
2. Edges (melted, smeared, asymmetric, fused):
Observed:
3. Background text and reflections:
Observed:
4. Shadow direction and softness:
Observed:
5. Repeated patterns (tiles, grain, leaves, weave):
Observed:
Reverse image search — oldest match found:
URL/context: Date:
Metadata — camera fields present? Y/N
Notes:
Flags count: [0-5]
Decision: [repost as-is / repost with "source unverified" label /
do not repost / ask sender first]FAQ
Can any tool tell me for certain whether a photo is AI generated? No. Detector tools produce probabilistic estimates from models you cannot inspect, and they can be wrong in both directions. Use them as one input among several, never as a verdict.
What is the single most useful check? Background text and reflections. Letter-shaped nonsense and reflections that contradict the scene are harder to produce convincingly than clean edges, and they are quick to inspect.
Does missing EXIF data mean the image is synthetic? No. Most social platforms strip metadata on upload, and screenshots never carry it. Missing metadata is close to meaningless on its own.
What if reverse image search finds nothing? It means you could not locate an earlier source. That is common for personal photos, private group chats, and recent uploads. It is not evidence either way.
How many red flags should stop me from reposting? There is no threshold that proves anything. As a working habit: one flag means look closer, several flags in the same image mean do not present it as verified, and any flag on an image about a real person or a real event means ask the sender before you share.
Should I tell people an image is AI generated if I am not sure? Say what you actually know: that you could not verify the source. That is accurate, it is useful to your audience, and it does not require you to make a claim you cannot support.