Most of the advice you have read about spotting AI images is out of date. Count the fingers. Look at the teeth. Check for garbled text in the background. That guidance was genuinely useful a few years ago, and relying on it now is worse than having no method at all — because it produces confidence without accuracy.
Image generation improved. Hands got fixed. Text got legible. The visual tells that people were trained to look for are precisely the flaws that developers prioritized eliminating.
So the method has to change. The reliable question is no longer does this look fake? It is where did this come from? That shift, from inspecting pixels to tracing provenance, is what this guide is about.
Why Squinting at Images Stopped Working
There are two problems with visual inspection, and they compound.
The first is false negatives. A well-made synthetic image simply does not contain the artifacts you are hunting for, so a careful inspection returns a clean bill of health that means nothing.
The second is more damaging: false positives. Real photographs get accused constantly. Compression softens details. Phone cameras apply aggressive computational processing to skin and skies. A genuine photo passed through a few rounds of social media re-encoding can look distinctly plasticky. When “it looks too smooth” becomes evidence of fakery, authentic images get dismissed — and that is arguably the greater harm, because it lets people wave away things that actually happened.
Visual inspection is not worthless. It is just a weak signal that should be applied last, not first.
The Order of Operations
Work through these in sequence. Most questionable content resolves at step one or two, and you never need to zoom in at all.
| Step | Question | Strength of Signal |
|---|---|---|
| 1. Source | Who posted this, and do they have any track record? | Strong |
| 2. Corroboration | If this happened, who else would have captured it? | Strongest |
| 3. Origin trace | Can I find an earlier or larger version of this file? | Strong |
| 4. Provenance data | Does the file carry credentials, metadata, or a label? | Moderate, improving |
| 5. Visual inspection | Does anything in the frame break physics or continuity? | Weak on its own |
Step 1: Interrogate the Account, Not the Image
Synthetic media rarely arrives from a well-established source with a reputation to lose. It arrives from accounts that are new, anonymous, high-volume, or engagement-farming.
Check the posting history. An account created recently that suddenly produces dramatic, emotionally charged content is the single most common vector. So is an aged account whose earlier posts are about something entirely unrelated — a sign it was bought or repurposed.
Look at how the caption is written, too. Content designed to spread avoids specifics. Real reporting says where and when. Manufactured content says “this is happening right now” and leaves you to fill in the rest.
Step 2: Ask Who Else Would Have Filmed It
This is the most powerful test available to a non-expert, and it requires no tools whatsoever.
Any genuinely dramatic public event in 2026 is captured from multiple angles by multiple people with cameras in their pockets. A remarkable event that exists in exactly one image, from one angle, posted by one account, is a claim that has failed its own plausibility check.
Search for the event itself rather than the image. If a major incident occurred in a populated place and no established outlet or second bystander has any record of it, the absence of corroboration is your answer. Silence from every direction is not a cover-up; it is evidence.
Step 3: Trace the File Backwards
Reverse image search remains the workhorse of verification, and it answers a more useful question than “is this AI.” It answers “is this old.”
An enormous share of viral deception is not synthetic at all. It is a real photograph from a different year, a different country, or a different event, recaptioned. Reverse search catches this instantly.
Practical notes that improve results:
- Crop before searching. Isolate a distinctive object — a sign, a building, a vehicle — and search that region alone. Full-frame searches often return visually similar noise.
- Try more than one engine. Different reverse-search tools index different corners of the web, and results diverge more than people expect.
- Hunt for the largest version. The highest-resolution copy is usually closest to the original. Screenshots of screenshots travel downhill.
- For video, search keyframes. Pause on a distinctive frame, screenshot it, and reverse-search that.
Step 4: Check for Provenance Credentials
The industry’s structural answer to this problem is not detection. It is provenance — attaching tamper-evident information about origin and editing history to the file itself.
The main effort here is the C2PA standard, developed by the Coalition for Content Provenance and Authenticity, whose members include major technology, camera, and media organizations. The concept is often described publicly as “Content Credentials,” and where it is implemented, you can inspect a file to see what device or software created it and how it was subsequently modified.
Two things are worth understanding about this.
First, its usefulness is asymmetric. Present, valid credentials are meaningful positive evidence. Absent credentials prove very little — plenty of authentic images have no credentials at all, and metadata is routinely stripped when a file passes through messaging apps and social platforms. Treat presence as a green light and absence as a shrug.
Second, adoption is uneven and still expanding. Support across cameras, editing software, generation tools, and platforms varies considerably, so check the current situation for whatever tool you are using rather than assuming coverage.
What about AI detector websites?
Use them, but hold their output loosely. Automated classifiers can be helpful as one input, and they are also known to misfire in both directions — flagging authentic photographs, clearing synthetic ones, and behaving unpredictably on heavily compressed files. A detector score is a hint, not a verdict. Never let one override an actual corroboration failure or an actual reverse-search match.
Step 5: What Still Shows Up Visually
Once the first four steps are done, inspection can add a little confirming weight. The durable tells are less about anatomy and more about logic and continuity — the relationships between things rather than the things themselves.
- Light that disagrees with itself. Shadows falling in inconsistent directions, or a subject lit differently from the environment around them.
- Reflections that do not match. Windows, water, glasses, and mirrors have to agree with the scene. They frequently do not.
- Background structures that fail. Railings that change count, brickwork that shifts alignment, patterns that lose their rhythm behind a subject.
- Physics of contact. Where an object meets a surface, where fingers grip a handle, where a foot meets the ground — contact points are hard, and they remain a soft spot.
- In video: temporal drift. Details that quietly change between the start of a clip and its end. Clothing patterns, jewelry, background objects, the length of a shadow. Scrub back and forth rather than watching once forward.
- Audio in video. Room acoustics that do not match the visible space, unnaturally even breathing or its total absence, and consonants that land slightly out of sync with lip movement.
Any single item on this list can appear in a real photograph. A cluster of them, in content that also failed the corroboration test, is worth acting on.
The Habit That Matters More Than Any Technique
The mechanism that spreads misleading media is emotional, not technical. Content engineered to travel is engineered to make you feel something strongly enough to share before you think.
So the most effective personal defense is procedural: when an image or clip produces a sharp emotional reaction — outrage, vindication, alarm — treat that reaction as the trigger to verify rather than the reason to share. The feeling is the signal. Ten seconds of checking the account, and one search for whether anyone else reported it, filters out the overwhelming majority of what you would otherwise pass along.
You do not need forensic expertise. You need to stop being the fastest link in the chain.
Frequently Asked Questions
Is there a single reliable tool that just tells me if an image is AI?
No, and treat any product claiming to be one with caution. Detection tools are probabilistic, they are trained against yesterday’s generators, and their accuracy degrades on compressed or edited files. Provenance and corroboration are more dependable than classification.
Does checking metadata still help?
Sometimes, and it is quick enough to be worth trying. Camera model, timestamps, and location data can support or undermine a claim. But metadata is trivially editable and routinely stripped by platforms, so it can suggest a conclusion and should not carry one alone.
Are AI videos easier or harder to spot than images?
Video gives you more to work with, because it has to stay consistent across time and audio. A still frame only has to hold up once. That said, quality has risen sharply, so the method is the same — start with the source, not the pixels.
What should I do if I already shared something that turned out to be fake?
Post a correction rather than quietly deleting. Deletion leaves the original impression intact in everyone who saw it, while a visible correction reaches the same audience and does more to slow the spread than the deletion does.
Is it always wrong to share AI-generated images?
Not remotely — synthetic media has legitimate creative, illustrative, and commercial uses. The problem is unlabeled synthetic content presented as documentation of something that happened. Label it, and the ethical issue largely disappears.
