Stop Trusting What You See on Social Media. Here's How to Check Whether a Photo Is Real or Was Made by AI.

The Photograph Everyone Had an Opinion About

Recently, this photograph spread across social media showing President Donald Trump kissing a woman on a golf course.

Some people immediately said it was real.

Others immediately said it was AI.

It spread precisely because both reactions were emotionally easy.

This article makes no claim about President Trump's conduct or the identity of the woman shown. The image is useful here only as a case study in how difficult visual verification has become.

Fact-checkers found no verified photographer, date, or location for the blurry viral version. Detector tests did not establish that blurry image as AI-generated. A sharper version circulating later had a six-fingered hand, was rated 99.9% likely AI-generated by Hive, and OpenAI Verify found a SynthID watermark associated with OpenAI tools. Crucially, fact-checkers note that the sharper version's existence does not by itself prove the blurry image is fake unless the publication sequence can be established.

That uncertainty is the lesson.

Why This Matters Beyond One Political Photo

AI-generated political media is becoming a serious election-integrity concern, and lawmakers are already fighting over how synthetic political content should be regulated. A federal judge recently blocked part of Montana's deepfake-election law on First Amendment grounds, illustrating how difficult this problem is legally as well as technically.

The danger is not merely that fake photographs can fool voters.

It is that people may eventually stop knowing which photographs deserve to be believed at all.

The Old AI-Detection Advice Is Becoming Obsolete

Everyone has heard the common advice for spotting AI-generated images: count the fingers, inspect the teeth, look at the eyes, search for garbled text, check reflections.

Those are still useful.

But increasingly:

They are clues, not authentication.

A six-fingered hand can prove something is wrong. Five perfect fingers cannot prove something is real.

Why "Ultra-Realistic" AI Often Still Looks Like AI

Prompts filled with terms like "Ultra-realistic, 8K, cinematic, hyper-detailed skin, masterpiece." do not create an identifiable AI fingerprint. But the aesthetic they often request—immaculate skin, perfect lighting, exaggerated depth of field, aggressively finished surfaces—helped create the visual language many people now associate with synthetic imagery. Unfortunately, modern generators no longer need to look like that.

Real photographs contain accidents. AI images often contain intentions everywhere.

However, those stylistic clues are becoming less reliable as models improve.

AI Leaves Traces, But Not Always Visible Ones

Synthetic images may leave: provenance metadata, C2PA credentials, generator-specific watermarks, SynthID signals, editing history, or metadata inconsistencies.

OpenAI now combines C2PA provenance with SynthID watermarking for supported generated media, and it offers verification tooling. OpenAI also explicitly warns that failure to detect a signal does not prove a file is authentic.

A positive provenance signal can establish something about origin. A missing signal cannot establish authenticity.

Compression Is the AI Image's Best Friend

The blurry version of the Trump image hid artifacts that were visible in a sharper version.

Social-media compression can erase exactly the mistakes people rely on to detect synthetic imagery.

Cropping, screenshots, resizing, sharpening, and recompression can destroy both: visible artifacts and useful metadata.

The worse the copy becomes, the less reliable visual inspection may become.

Stop Asking "Does This Look AI?"

The better question is:

Where Did This Image Come From?

That means:

1. Who posted it first?
2. Is there a photographer?
3. Is there an original file?
4. Was it published by a news organization?
5. Are there independent photos from the same event?
6. Does video exist?
7. Can the date and location be verified?
8. Does reverse-image search find an earlier copy?

Authenticity is increasingly a provenance problem, not an eyesight problem.

The Five-Level Reality Check

Here is a method to check a viral image:

Level 1: Source

Who published the original? Was it an anonymous repost, a news agency, a photographer, or an official account?

Level 2: Corroboration

Does independent evidence show the event occurred? Are there other photos, video, witnesses, or press coverage?

Level 3: Provenance

Look for: C2PA Content Credentials, SynthID, or cryptographic origin records. C2PA was created specifically to provide tamper-evident provenance about how digital media was created or modified.

Level 4: Context

Do the location, clothing, weather, schedule, surroundings, and lighting all fit what is known about the event or individuals?

Level 5: Pixels

Only now inspect: anatomy, human contact, shadows, reflections, text, skin, fabric, repeated patterns, and geometry.

Most people currently do this in reverse order.

Why AI Detectors Are Not Lie Detectors

Do not rely on a single AI detector. Different detectors disagree. Compression changes results. Coverage differs by generator. Some systems detect watermarks, others infer statistically from pixels.

An AI detector is evidence. It is not a judge.

The strongest verification combines source, provenance, corroboration, and visual analysis.

Try It Yourself: A Prompt for Examining a Suspicious Image

Do not tell the AI whether I think this image is real or fake. Form your judgment independently first.

Analyze this image as a skeptical digital-image examiner. Do not assume it is authentic or AI-generated. First evaluate only visible evidence:

• Anatomy: Fingers, eyes, teeth, ears, overall body proportions.
• Human Contact: How skin meets skin, objects held, interactions.
• Lighting: Consistency of light source, shadows, reflections.
• Reflections: What is reflected and its consistency with the scene.
• Geometry: Straight lines, repeating patterns, perspective.
• Fabric: Texture, drape, folds, how it interacts with the body.
• Text: Legibility, distortion, consistency.
• Hair: Individual strands, natural flow, fine details.
• Depth: Natural blur, focus, background consistency.
• Compression: Visible artifacts from image processing.

Based on this visible evidence, provide:

• Strong evidence for authenticity: Specific visual cues that suggest it is real.
• Suspicious clues for synthetic origin: Specific visual cues that suggest it is AI-generated.
• Normal photographic explanations: Any anomalies that could be explained by typical camera behavior or editing.
• What cannot be determined from pixels alone: Aspects requiring external verification.

Finally, state your:

• Pixel verdict: Based only on visual inspection, is it more likely authentic, synthetic, or inconclusive?
• Confidence level: (Low, Medium, High)
• Most important reason for your verdict: What single piece of visual evidence or lack thereof was most determinative?

A multimodal AI can hallucinate visual evidence too. Never treat its verdict as authentication. Use it to generate questions for further investigation.

The Even Bigger Problem: Real Photos Can Now Be Dismissed as AI

The frightening consequence of convincing fake images is not merely that we will believe things that never happened. It is that we may stop believing things that did.

If fake images become believable, another thing happens. Real images become deniable. Someone confronted with authentic evidence can simply say: "AI."

Generative AI does not merely create fake evidence. It can weaken real evidence.

This is often described as the liar's dividend: the existence of convincing synthetic media gives people a plausible excuse to deny authentic recordings. And in politics, journalism, law, and warfare, that is potentially as consequential as the fake itself.

Elections Make This More Urgent, Not More Partisan

The United States is approaching the 2026 midterms amid broader concern about AI-generated misinformation and other election-security challenges. Reuters reports that election officials are preparing for misinformation and AI-generated disinformation among several risks, while recent litigation shows that regulating political deepfakes also raises serious free-speech questions.

This is not a Republican problem or a Democratic problem.

A synthetic photograph can target anyone.

And a genuine photograph can be dismissed by anyone.

The Recursum Connection

The deeper lesson is one Recursum keeps encountering in other forms: trust increasingly depends on lineage.

For decisions: Decision Lineage.

For knowledge: Longitudinal Knowledge.

For media: Content Provenance.

When appearance becomes unreliable, history becomes evidence.

Who created it? Where? How was it transformed? What happened to it before it reached you?

The Social Media Rule

The stronger the emotional reaction an image creates, the slower you should be to share it.

Not because it is necessarily fake.

Because emotional content is exactly what bypasses verification.

Verification should happen before amplification, not after embarrassment.

Ending

For most of human history, seeing was powerful evidence.

Then photography gave us mechanical eyes.

For more than a century, photographs became shorthand for:

This happened.

Generative AI has broken that shortcut.

A photograph can now depict something that never occurred.

And something that really occurred can be dismissed as synthetic.

That does not mean we should stop trusting everything we see.

It means we need to stop treating seeing as the end of verification.

Look at the image.

Inspect the details.

Ask AI to challenge it.

But then do something more important:

Find its history.

• Who made it.
• Where it came from.
• What happened to it.

Whether anything else proves the event occurred.

Because in the age of generative AI, the question is no longer:

"Does this look real?"

The question is:

"Can I prove where this came from?"

TL;DR: The core lesson for the AI era is: seeing is no longer verification. A viral image of President Trump kissing a woman (used neutrally as a case study) highlights this. Fact-checkers found no verified source for the blurry version, and a sharper one was clearly synthetic, yet the blurry image's authenticity couldn't be conclusively disproved from pixels alone. This uncertainty is the lesson. AI-generated political media is a growing election concern, risking that people will stop believing any photographs. Old detection advice (counting fingers, checking eyes) is becoming obsolete.

Five perfect fingers don't prove reality. "Ultra-realistic" AI often has a distinct, overly polished aesthetic, but these clues are diminishing. AI leaves forensic traces (C2PA, SynthID), but absence of evidence isn't proof of authenticity. Social-media compression can also hide subtle artifacts. The key question shifts from "Does this look AI?" to "Where did this image come from?" A five-level reality check is proposed: Source, Corroboration, Provenance (C2PA, SynthID, metadata), Context, and finally, Pixels.

AI detectors are evidence, not judges. A custom prompt is provided to make AI act as a skeptical image examiner, but warns that multimodal AI can hallucinate visual evidence. The "liar's dividend" is a major problem: convincing synthetic media allows people to dismiss real evidence as fake. This is urgent for politics, journalism, law, and warfare. Recursum's work on Decision Lineage, Longitudinal Knowledge, and Content Provenance reinforces that trust increasingly depends on history, not appearance. The social media rule: emotional content bypasses verification, so verify before amplifying. In the age of generative AI, the question is "Can I prove where this came from?"

By Ernesto Verdugo, AI Architect, Founder of Verdugo Labs, and Creator of Recursum. He builds systems for what happens when human and artificial intelligence stop working separately.