Real Image or AI-Generated?

Perfect hands. Readable text. Correct reflections. And the entire image was generated by AI.

Now consider another photograph. The anatomy looks strange, one leg appears to disappear and the angle seems physically impossible. It looks artificial, but it was captured by a real camera at exactly the wrong moment.

That is the problem with trying to detect AI images by sight alone. Generated images are becoming better at avoiding familiar mistakes, while real photographs can look unnatural because of perspective, movement, editing or coincidence.

Fake can look real. Real can look fake. Sometimes the image is authentic while the story attached to it is completely false.

The most useful question is no longer:

“Can I find the AI mistake?”

It is:

“What evidence would justify believing this image?”

What does “real” actually mean?

The phrase “real image” sounds simple, but online images can have several different origins.

• Camera-captured: A camera recorded a real scene.
• Traditionally edited: The photograph was adjusted, retouched or manipulated without generative AI.
• AI-generated: Most or all of the image was synthesised by a generative model.
• AI-modified: A real photograph contains generated objects, people, backgrounds or extensions.
• Real but misrepresented: The photograph is authentic, but the date, location, identity or caption is false.

The last category is especially important. A genuine photograph from a protest in one country can be reposted years later as evidence of an unrelated event somewhere else. Every pixel may be real, yet the conclusion is still false.

This is why “real or AI?” is sometimes the wrong question. You may need to determine whether the image came from a camera, whether part of it was generated, whether the event happened and whether the caption describes it accurately.

Those are different claims, and each requires different evidence.

Why common AI detection tricks are becoming weaker

Most advice about spotting AI images repeats the same checklist: count the fingers, inspect the eyes, read the signs, check the background and look for strange reflections.

These clues can still be useful, but they are not proof.

A generated image can contain correct hands, coherent text and believable lighting. A real photograph can show strange anatomy because two objects overlap, the subject is moving or the camera freezes an unusual moment.

Professional portraits, architecture and food photography may also look “too perfect” without being artificial. Beauty is not evidence, and strangeness is not evidence either.

More useful clues usually involve relationships between objects:

• a finger passing through a glass
• glasses disappearing before reaching an ear
• jewellery merging with skin or clothing
• a reflection showing the wrong objects
• shadows that conflict with the light source
• repeated people, windows or textures
• background structures that stop making physical sense

These details can justify suspicion, but they still cannot reveal the complete history of an image. Compression, motion blur, lens distortion and traditional editing can create similar effects.

A clue tells you to investigate. It does not deliver a final verdict.

Sometimes there is nothing visible to detect

Imagine a convincing image of a street market. The people look natural, the signs are readable, the reflections match the wet pavement and the architecture remains coherent. Nothing appears broken.

The image was still generated by AI.

Its origin can be confirmed because the generation process was preserved. Without that record, however, an ordinary viewer might not be able to verify it by studying the pixels alone.

In cases like this, the most accurate answer may be:

“I do not know yet.”

That answer can feel unsatisfying, but it is better than pretending to have evidence that does not exist.

A large Microsoft Research experiment analysed approximately 287,000 image evaluations from more than 12,500 participants. Overall accuracy was about 62%. That is better than the 50% expected from random guessing, but it is not reliable enough to treat human vision as a universal AI detector.

The result does not mean that every viewer or image will produce exactly the same score. Difficulty changes with the generator, subject, image quality, compression and testing method. The broader lesson is that visual judgement alone is inconsistent.

Remember C.O.R.E.

The C.O.R.E. protocol is a four-step method for examining suspicious images without depending on one temporary visual trick.

C — Context

Before zooming into individual pixels, examine the post surrounding the image.

Ask:

• Who published it?
• Is this the original account?
• What exactly does the caption claim?
• Is there a date or location?
• Is the photographer or creator identified?
• Does the account have a reason to manipulate the audience?
• Is the post designed to create urgency, fear or anger?
• Can reliable sources confirm the same event?

A photograph can be genuine while its description is false. Always separate the file from the story attached to it.

It also helps to define the precise claim being investigated. Instead of asking whether an image “looks suspicious”, ask whether it truly shows a particular event, in a particular place, on a particular date.

A precise question makes it easier to identify the evidence you need.

O — Oddities

Now inspect the image itself. Do not search only for additional fingers or distorted faces. Look at how objects interact with one another.

Check whether:

• hands hold objects naturally
• feet make contact with the ground
• glasses and jewellery connect correctly
• reflections match the scene
• shadows are compatible with the light
• repeated patterns appear in crowds or backgrounds
• thin objects continue behind obstructions
• signs and symbols remain coherent
• perspective and scale make physical sense

The strongest visual clues are usually specific. “The image feels artificial” is weak. “The reflection contains an object that is not present in the scene” is much more useful.

Even then, the correct conclusion is not automatically “AI-generated”. The inconsistency may come from another form of editing or from the photograph itself.

R — Reverse search

A reverse image search can help locate earlier appearances, original publications, uncropped versions, different captions and the real date or location.

Search the complete image first. If that fails, try a crop containing a distinctive building, sign, face or object. Screenshots, borders and captions can prevent a search engine from recognising the original.

Reverse search is particularly valuable when an authentic photograph is being used deceptively. A supposedly current image may have appeared online years before the event it is claimed to show.

Finding an older result is not always enough. Look for a consistent source trail: a credited photographer, related photographs, reliable reporting and a publication date that fits the event.

The goal is not simply to find an earlier copy. It is to reconstruct where the image came from.

E — External evidence

Finally, look beyond the image itself. Useful evidence may include:

• independent photographs or videos
• reporting from credible local sources
• the original photographer or organisation
• metadata
• Content Credentials, which can record a file’s origin and editing history
• generation records
• platform labels
• compatible watermark systems

One repeated screenshot is not independent confirmation. Ten accounts may all be copying the same unverified source.

Look for evidence created separately: different camera angles, original uploads, identifiable creators or multiple reports documenting the same event.

Metadata and provenance systems can strengthen an investigation, but they also have limits. Social platforms may remove metadata, screenshots may destroy provenance records and not every generative tool uses the same watermark system.

The absence of a credential does not prove that an image is real. Its presence may explain part of the file’s history, but it does not automatically prove that the caption is true.

A quick workflow for checking an image

Put the C.O.R.E. method into practice

The quick workflow above is designed to help you examine suspicious images without relying on one visual trick or automatic detector.

Bookmark this article so you can return to the C.O.R.E. protocol whenever an image makes an extraordinary claim, asks you to act quickly or seems designed to provoke an immediate reaction.

Join the AtomicCurious newsletter for future investigations, visual challenges and practical tools →

You can also watch the full YouTube episode above to test your instincts against the complete real-or-AI challenge.

Use conclusions that match the evidence

Not every investigation should end with a confident “real” or “fake”. Use a conclusion that reflects what the evidence actually supports.

Verified camera capture

There is a traceable source and supporting material showing that the image originated from a camera. The caption may still need separate verification.

Verified AI-generated or modified

There is direct evidence that the image was generated or substantially altered using AI.

Misleading context

The image may be authentic, but the date, location, identity or explanation attached to it is false.

Likely generated or manipulated

Several clues support this conclusion, but the complete creation history is unavailable.

Unverified

There is not enough evidence for a stronger conclusion.

“Unverified” does not mean real. It does not mean fake. It means that the investigation has not produced enough information yet.

This may be the most important category because online platforms constantly pressure users to choose a side before the evidence is available.

Why confidence can be dangerous

Images create conclusions quickly. You see a face, a crowd, a disaster or an accusation before you begin analysing the evidence.

Once an initial interpretation forms, every detail can start supporting it. If you believe the image is fake, every blur becomes an AI artefact. If the image confirms something you already believe, every realistic detail becomes proof.

The most useful habit is therefore simple:

Pause before deciding what the image means.

You do not need to distrust every photograph online. Verification should be proportional to the consequences.

An ordinary landscape image may require little investigation. An image used to request money, accuse someone, provoke outrage or document an extraordinary event deserves much stronger evidence.

The most dangerous image may be real

AI-generated misinformation matters, but focusing exclusively on artificial pixels creates a major blind spot.

A real photograph can still be used to:

• invent a breaking-news event
• exaggerate the size of a crowd
• falsely identify a person
• support an unrelated political claim
• advertise a scam
• create a false accusation

A misleading caption does not require an advanced model. It only requires a convincing photograph and an audience willing to share it without checking.

The future of image verification therefore cannot depend only on detecting AI. It also requires source literacy, context, provenance, corroboration and the willingness to remain uncertain.

So, how can you tell if an image is real or AI?

Sometimes a visible error will reveal a generated image. Sometimes a reverse search will expose an authentic photograph used with a false story. Sometimes provenance information will show that only part of a photograph was generated.

There will also be cases where the pixels reveal nothing and the available evidence remains incomplete. In those situations, the most responsible answer is:

“I do not know yet.”

Remember C.O.R.E.: context, oddities, reverse search and external evidence. Do not trust an image merely because it looks real, and do not reject one merely because it looks strange.

Investigate why you should believe it.

Bookmark this article so the C.O.R.E. protocol is available whenever you need to examine a suspicious image.

Then take the complete real-or-AI test in the YouTube video and compare your visual instincts with the results of a full investigation.

⚛️ AtomicCurious — Exploring science, technology and intelligent curiosities.