When a suspicious image starts moving through group chats or social feeds, most people do the same thing: zoom in, hunt for obvious mistakes, and hope the answer will reveal itself. Sometimes it does. Often it does not.
That is the uncomfortable part of this topic. The newest synthetic images can look very convincing on first pass, and a lot of older “spot fake images by looking at hands” advice has aged badly. But that does not mean the task is hopeless. It means you need a better checklist.
The thesis here is simple: the most reliable first pass is not one clue but a sequence of questions. You inspect anatomy, style, function, physics, and provenance, then decide whether the evidence is strong, weak, or inconclusive.
Start by asking what kind of image this claims to be
Before you zoom into pixels, ask a basic question: what is the image presenting itself as?
Is it supposed to be:
- a candid phone photo
- a studio product image
- a news image
- a surreal artwork
- a meme or composite
The standard is different in each case. A glossy conceptual artwork can be visually coherent while still being synthetic. A “breaking news photo” faces a much stricter burden. Context changes what counts as suspicious.
Category 1: anatomy
The 2024 guide by Negar Kamali and coauthors is still a good organizing tool because it groups frequent cues into categories instead of pretending there is one universal tell. One of the most durable categories is anatomical inconsistency.
Kamali et al., 2024
Look for:
- fingers that merge, vanish, or multiply
- jewelry that passes through skin or hair unnaturally
- teeth or eyes that repeat in strange patterns
- ears, shoulders, or elbows with impossible attachment points
- faces that seem plausible until you compare both sides carefully
This is less about “bad hands” as a meme and more about object continuity. Bodies are structured systems. Generators often fail at the points where many parts must agree at once.
Category 2: style and texture
Some AI images feel wrong before you can explain why. That reaction is often tied to texture and style.
Common clues include:
- skin that is too smooth in one zone and oddly noisy in another
- fabric folds that look decorative rather than physical
- surfaces that seem detailed but lack material logic
- localized blur that does not match depth of field
- sharpened edges floating inside otherwise soft regions
The important thing is not to overreact to one patch of weird texture. Compression, denoising, and editing can all do odd things. What matters is whether several texture inconsistencies support the same suspicion.
Category 3: function
This is one of the most underrated tests. A generated image may look fine as a composition while quietly failing as an actual object in the world.
Check whether things would work if they were real:
- glasses that do not sit on a face correctly
- straps, zippers, or seams that have no usable path
- keyboards with impossible layouts
- doors, rails, or stairs that do not operate like objects
- product interfaces with fake icons or unreadable logic
Function is useful because it forces the image to behave like reality, not just resemble it.
Category 4: physics
Lighting, reflections, shadows, and perspective are still important, but they need patience. The 2024 paper explicitly includes violations of physics as a category, and for good reason.
Ask:
- do shadows agree on direction?
- do reflections match the object and camera position?
- does the depth of field make sense across the scene?
- do distant and near objects share the right scale relationships?
This is where humans can still do useful work, especially in scenes that claim realism. A physically impossible highlight or reflection is often stronger evidence than a slightly odd face.
Category 5: sociocultural or contextual weirdness
Some images fail because they combine plausible parts into an implausible whole. Kamali and colleagues include sociocultural implausibilities for that reason.
Kamali et al., 2024
Examples:
- a scene with signage that imitates language rather than using it correctly
- uniforms or symbols mixed in ways a real context would not permit
- crowd behavior that feels statistically off
- objects placed for visual effect with no situational logic
This category is especially useful for “viral documentary” images. The scene may be less broken at the pixel level than at the world-knowledge level.
When humans are more and less reliable
The 2025 study by Kamali and colleagues is helpful because it does not just ask whether humans succeed. It asks when. Based on 749,828 observations from 50,444 participants, the authors found that scene complexity, artifact type, display time, and curation affect detection accuracy.
Kamali et al., 2025
That means a very polished AI image can beat quick inspection, especially if it was curated before posting. It also means a low-confidence human judgment should stay low-confidence. “I stared at it and it felt fake” is not nothing, but it is not enough.
Look beyond the pixels when possible
If you can access the original file, inspect metadata. IPTC explains that image files can carry administrative, descriptive, and rights-related information.
IPTC photo metadata overview
If a provenance record is present, that can be even better. The C2PA explainer says Content Credentials can record origin, modifications, and use of AI in a tamper-evident way.
C2PA explainer
This is one of the biggest mistakes in casual verification: people try to solve everything from a screenshot of an image that may have stripped the very metadata or provenance clues that mattered most.
A practical first-pass checklist
If you only have thirty seconds, do this:
- Ask what the image claims to be.
- Check hands, teeth, eyes, and symmetry.
- Check object function: straps, rails, keyboards, clothing construction.
- Check shadows, reflections, and perspective.
- Look for language, signage, and contextual logic failures.
- If the file is available, inspect metadata or provenance.
- If stakes are high, run a detector and compare its explanation with your own observations.
That sequence is not perfect, but it is much better than random zooming.
What this can and cannot tell you
What it can tell you
- which visual clues are still useful in 2026
- how to organize your inspection instead of guessing
- when a file deserves deeper verification
- why a detector can help even after human review
What it cannot tell you
- that every AI image has obvious visible artifacts
- that one artifact proves the entire image is synthetic
- that a polished image without visible errors must be authentic
- that context and provenance can be ignored
The disciplined conclusion
The point of this process is not to make you overconfident. It is the opposite. A good review helps you sort images into three buckets:
- likely synthetic
- likely authentic
- not enough evidence from the image alone
That third bucket matters. It is where most bad certainty comes from. People often force an answer from weak evidence because “unclear” feels unsatisfying. In authenticity work, “unclear” is often the most honest answer available.
If you want a second layer after visual review, try a scan on the Detectiks home page or use the extension when images are already circulating in the browser.
Last reviewed
May 11, 2026.
