In February 2023, an image of Pope Francis wearing a large white puffer jacket went viral on Twitter. Within hours it had been reshared millions of times. News outlets in Italy, Spain, and the United States embedded it without comment. People assumed it was a candid photo taken outside the Vatican.
It was not a photo. It had been generated by Midjourney version 5, which had just been released days earlier. The person who created it, a Chicago construction worker named Pablo Xavier, posted it in a Facebook group almost as a test — and watched it escape into the internet as fact.
When researchers at BuzzFeed News and Bellingcat went back and studied the image closely, they found it: the Pope's right hand had too many fingers. The lace detail on his white garment dissolved into a soft blur at the edges. His glasses had no visible temples — the arms that hook behind the ears had simply been forgotten by the model. And his skin had the smooth, slightly plastic quality that AI images had not yet learned to lose.
The image fooled millions. But once you knew what to look for — once you knew the specific places AI models consistently get wrong — you couldn't unsee it. That's what this lesson teaches you to do.
AI image generators like Midjourney, DALL-E, and Stable Diffusion work by learning statistical patterns. They train on billions of images and learn what things tend to look like near other things. A face usually has eyes above a nose. A hand usually appears near an arm. A street usually has buildings on both sides.
The problem is that hands are incredibly complex. A human hand has 27 bones, dozens of tendons, and can form hundreds of distinct shapes. When you look at a hand in a photo, you're seeing a precise mechanical structure. When an AI generates a hand, it's predicting what a hand-shaped cluster of pixels usually looks like — and it frequently predicts wrong. Extra fingers. Fingers that merge into each other. Thumbs on the wrong side. Nails that face different directions.
This isn't a bug that's going to disappear tomorrow. As of 2024, even the best AI image generators still produce hands that require close inspection. The more fingers visible in the image, the higher the chance something is wrong.
After the Pope jacket image went viral, several digital forensics researchers published guides to spotting AI anatomy errors. They identified a consistent set of places where AI models fail. You can run this checklist on any suspicious image in under thirty seconds.
Fingers and hands. Count visible fingers. Check whether they look independently jointed. Look for fingers that are too long, too short, that merge at the base, or that have the wrong number of knuckles. Look for thumbs that appear on both sides of a hand.
Eyes. AI-generated eyes are often slightly asymmetrical in ways real eyes aren't. The reflections in the eyes — called catchlights — sometimes don't match. One eye might reflect a window; the other reflects nothing. Look at the pupils: in real photos they're round or slightly oval; in AI images they sometimes have subtle irregular shapes.
Ears. Ears are another structure AI consistently fails at. They're often blurred, simplified, or on close inspection structurally impossible — the inner cartilage folds in the wrong direction or doesn't exist at all.
Teeth. When mouths are open, AI often generates teeth that blend into a single smooth surface instead of showing distinct individual teeth. Or conversely, it generates too many teeth for the space.
Hairlines and hair at edges. Where hair meets the background of an image, AI often produces a soft, unnatural blending — or individual hairs that seem to multiply and fork in impossible ways.
You've now got the first tool. Before you read further: think of one image you've seen recently that seemed surprising or unusual. Could any of these features have been off? You don't need to answer definitively. The habit of asking is what matters.
Here's the honest truth: most people who saw the Pope jacket image did not look at his hands. They looked at his face, registered "Pope," and moved on. That is the normal way of seeing. We're pattern-matchers too — and we mostly look for confirmation of what we expect to see.
What forensic analysts and fact-checkers do differently is train themselves to look at the parts of an image that carry the most information. Not the thing that catches your eye — but the thing that's hardest to fake.
You can now do that. You know that hands are the hardest thing for AI to generate correctly. You know what to look for in eyes, ears, and teeth. Most adults who share AI images online have never been told this. You have been.
Knowing where AI images break down puts you ahead of the vast majority of people who encounter them. Not because you're smarter — because you've learned to look in the right places. That is a specific, real skill. The next time someone shares an image claiming to be a real photo, you'll check the hands. That changes everything.
Pablo Xavier didn't claim the Pope jacket image was real. He posted it in a hobby group. But millions of people shared it as real anyway. Who is responsible for the harm caused by a convincing AI image — the person who made it for fun, the platforms that allowed it to spread, or the people who shared it without checking? Does intent matter if the damage is the same?
You're reviewing images flagged for possible AI generation before a news outlet publishes a story. Your AI partner has studied the same forensics literature you have. They will push back on weak reasoning and ask you to justify your calls.
Start by describing a scenario: pick any type of image — a portrait, a crowd photo, a political image, a product ad — and describe what you'd examine first and why. Your partner will respond as a peer investigator, challenge your approach, and add their own analysis.
In March 2023, Eliot Higgins, the founder of the open-source investigative group Bellingcat, generated a series of images purporting to show Donald Trump being arrested by police officers. He created them in Midjourney and posted them on Twitter explicitly labeled as AI-generated, as a demonstration of what the technology could now produce.
Several of the images circulated beyond Higgins's posts without his labels. Fact-checkers at AFP and Reuters were called in to verify them. One of the first things their analysts noted — before checking fingers, before checking metadata — was the lighting.
In one image, Trump's face was lit from the left. The police officers surrounding him were lit from the right. The pavement between them had a shadow pattern that matched neither. Three different scenes had been averaged together by the AI into a single image — and physics didn't care that the statistics said it looked right. The light said something different.
Lighting inconsistency became one of the primary diagnostic tools that major wire services used in 2023 and 2024 to flag AI-generated images before publication. It is not obvious to a casual viewer. It is obvious once you know how to look.
In a real photograph, there is one physical light source — or a coherent set of them. The sun, a lamp, an overhead fluorescent. Every object in the scene is lit by that same source, which means every shadow falls in the same direction, every highlight appears on the same side of a curved surface, and every reflective surface catches the same color of light.
AI image generators don't simulate physics. They predict what a scene usually looks like based on millions of training images. The problem is that those training images come from thousands of different scenes with different lighting. When the AI averages them together, it sometimes produces a scene where different parts of the image were lit by different imaginary light sources — sources that can't coexist in the same physical space.
This produces specific, detectable errors.
Step 1: Find the main light source. Look at the image and identify where the light is coming from. Is the sun visible? Is there a window? A lamp? Once you've identified it, check whether the shadows in the image are consistent with it.
Step 2: Check shadow directions. Pick two or three objects in the scene — a person, a building, a chair — and trace the direction their shadows fall. In a real photo, they all point away from the same source. In AI images, they sometimes point in different directions.
Step 3: Check the catchlights. If faces are visible, look at the eyes. Each eye should have a small bright reflection. In a real photo, both eyes show the same reflection. In AI images, one eye might show a window and the other might show a blank wall — or one might have a reflection that doesn't exist anywhere in the scene.
Step 4: Check highlights on curved surfaces. Noses, cheeks, shoulders, and round objects like glasses should all have highlights on the same side — the side facing the light source. If highlights appear randomly placed, or on the wrong side of a surface, that's an artifact.
Step 5: Check reflected color. In real scenes, surfaces near a colored light source pick up that color in their shadows (this is called color bleeding or color cast). AI images sometimes get this right and sometimes produce surfaces that are completely neutral when they should be tinted.
In 2024, Adobe, Getty Images, and the Associated Press all updated their editorial policies to require photographers to disclose AI-generated content. One of the methods these organizations use to audit submissions is automated lighting analysis — software that checks whether shadow directions and catchlights are consistent. You just learned to do by eye what professional tools do computationally.
Bellingcat's Eliot Higgins made the Trump arrest images to demonstrate capability — to show what was now possible so that the public would be prepared. He labeled them clearly. They still escaped and spread as real.
Researchers who study synthetic media call this the demonstration problem: showing people what a fake looks like sometimes makes better fakes more plausible, not less. The more you demonstrate that AI can produce convincing images, the more people assume any surprising image might be AI — which can also be used to dismiss real photos as fake.
You now have a way to check. But knowing how to check is different from having time to check every image you see. Most people make sharing decisions in seconds. The lighting analysis you just learned takes a full minute of careful attention.
If you know that demonstrating an AI image trick causes some people to use it for harm, but also causes some people to become better at detecting it — what's the right call? Is it better to show the public how sophisticated AI fakes have become, or does that demonstration itself become a tool for bad actors? Who gets to decide?
Wire service analysts at Reuters and AFP use lighting consistency as a primary detection method. You've now learned the same five-step process they apply. This is not general media literacy — this is a specific technical skill used by professional fact-checkers. Most people who encounter AI images never apply it once.
A wire service editor has received three images for a breaking news story. You need to audit them for lighting consistency before they go to print. Your AI partner will analyze your reasoning and challenge any calls you can't back up with specific observations.
Describe a specific image scenario — political, celebrity, natural disaster, sports — and walk through your lighting audit step by step. Your partner will push back, add analysis, and ask you to go deeper on anything vague.
In October 2023, following the outbreak of fighting in Israel and Gaza, social media platforms were flooded with images claiming to document events on the ground. Researchers at the Stanford Internet Observatory and the Israeli newspaper Haaretz analyzed hundreds of viral images in the first two weeks of the conflict.
Many images that had circulated as documentation of current events turned out to be either archive photos from earlier conflicts, or AI-generated. The detection of AI images in this context often came not from faces or hands — many images showed rubble, crowds at a distance, or smoke — but from text and signage in the background.
In one widely shared image purporting to show a damaged building in Gaza, a researcher at Stanford noticed that a sign on a wall in the background contained letters that did not spell any real Arabic or Hebrew word. They looked like Arabic script — the curves and dots were approximately correct — but the characters had been blended into nonsense. Another sign appeared to be in English but read something like "MEDEICAL ENTRACE." Not "MEDICAL ENTRANCE" — a garbled approximation.
AI image models, as of 2023 and 2024, do not understand text. They generate images of text — shapes that look like letters — without knowing what those letters mean or whether they form real words. This is one of the most reliable tells in any AI image that contains writing.
When you read the word "EXIT" on a sign, you understand it as a sequence of meaning: four letters, one word, one instruction. You know the letters are in an order that produces a specific word in a specific language.
An AI image generator, when it produces a sign that says "EXIT," is not doing any of that. It is generating a cluster of pixels that statistically resembles what a sign with text on it looks like in the training data. If the text happens to be legible, that's coincidental. The AI has no model of language that it applies when generating images — even though the same company's language model understands text perfectly.
Image generation and language understanding are separate systems that don't yet talk to each other well. This is why you will often find, in AI images that contain signs, shop fronts, newspapers, graffiti, or name badges, text that is partially legible but contains errors, invented characters, or near-miss words.
Read every piece of text visible in the image. Signs, menus, name tags, newspapers, bumper stickers, graffiti, labels on equipment. If any text is garbled, partially invented, or uses characters that don't exist in the language it's supposed to be written in, that's a significant artifact.
Check architectural logic. AI images frequently produce buildings where windows don't align, doors appear at impossible heights, or brickwork patterns start and then arbitrarily change. Look at corners where walls meet — AI models often produce walls that don't join cleanly or that continue with different textures on each side.
Look for objects that merge. In real photos, distinct objects are distinct. In AI images, objects near each other sometimes merge — a crowd of people where individual bodies blend into each other at the edges, a bookshelf where the spines of books melt together. Look at the edges between objects, especially in busy scenes.
Check for repeated elements. AI models sometimes tile or repeat visual patterns in backgrounds — the same window appearing three times with slightly different warping, a crowd where you notice the same face appearing twice on different bodies. This is because AI generates from patterns, and sometimes the same pattern gets generated multiple times.
Check for impossible spatial relationships. In real scenes, objects that are behind other objects are partially occluded — you can't see the part that's blocked. In AI images, sometimes an object appears to be behind another object but is fully visible anyway, or a hand holding a cup appears in front of a wall but the wall's shadow somehow falls on the hand from both sides.
In conflict zones, AI-generated images can shape public opinion, justify military decisions, and affect aid organization responses within hours. When the Stanford Internet Observatory published its analysis in October 2023, they noted that some AI images had been shared by verified accounts with hundreds of thousands of followers. The detection skill you're building right now is something governments, newsrooms, and human rights organizations are actively trying to scale.
Most people look at an image's subject. They look at the face, the action, the dramatic element in the center of the frame. This is natural — it's where composition techniques direct the eye.
Developing the background-reading habit means training yourself to zoom out of the emotional core of the image and ask: what's happening at the edges? What does the sign on the wall say? Do the buildings make architectural sense? Are the objects in the background distinct from each other?
This requires deliberate effort. It goes against how images are designed to be read. That's exactly why it works — AI models are trained on images designed to look compelling at the center, and they put less effort into coherence at the periphery.
You've now added background analysis to your toolkit. Most fact-checkers working AI image cases describe the same moment: they were checking a face or hands, then noticed something weird in the background — a sign with nonsense text, a window that didn't exist on the right side of the building. You now know to look there first, not last.
The Stanford researchers who published their analysis of AI images in the Israel-Gaza conflict were helping the public understand what was real. But publishing a detailed guide to how those images were detected also gives future image creators a roadmap to avoid those specific errors. Is publishing detection methods always the right call? Or does it accelerate an arms race between fakers and checkers that the checkers can't win?
A human rights organization has flagged 20 images circulating on social media, claiming to document events in a conflict zone. You're the analyst assigned to triage them by examining backgrounds, text, and object coherence. Your AI partner is on the same team — but will challenge any call you make without specific evidence.
Describe an image scenario involving a location (street, building, crowd, rubble) and walk your partner through your background analysis. Be specific: what text do you see? What architectural details? What object boundaries?
In January 2024, explicit AI-generated images of Taylor Swift were posted on the platform X (formerly Twitter) and spread to millions of viewers before they were taken down. The images had been generated using a Microsoft AI tool and distributed initially through a Telegram channel dedicated to creating non-consensual synthetic images of celebrities.
What made this case different from earlier AI image controversies was how quickly analysts at several organizations, including Wired, 404 Media, and the Stanford Internet Observatory, were able to confirm the images were AI-generated. It wasn't one artifact — it was the convergence of several.
Researchers noted: skin texture that was slightly too smooth and uniform across lighting variations. Background elements (crowd figures in some images) that merged at their edges. Proportional inconsistencies in how different body parts related to each other spatially. And in several images, subtle text on clothing or merchandise that was partially garbled.
No single artifact was definitive. But four independent signals pointing in the same direction made the case as solid as forensic analysts could make it without access to the original generation logs. This is how professional image forensics actually works: you don't need a confession; you need a pattern of evidence that has no other plausible explanation.
When you've been through the previous three lessons, you have a substantial toolkit: anatomy artifacts (hands, eyes, ears, teeth), lighting inconsistency (shadow directions, catchlights, highlights), and background analysis (text, architecture, object merging, repeated elements).
The professional approach is to treat each of these as an independent line of evidence. When you find one artifact, you have a flag — something worth investigating. When you find two, you have a serious concern. When you find three or more, especially across different categories, you have a pattern of evidence that is very difficult to explain any other way.
This is important for a specific reason: real photos can have apparent artifacts too. A real photo taken in unusual light might have odd shadows. A real photo taken through glass might distort a sign. A real photo of a crowd taken with motion blur might make bodies seem to merge. The difference is that these real-photo artifacts have specific explanations — you can point to the glass, the blur, the lighting angle. AI artifacts tend to be unexplainable by any single cause.
Here's an uncomfortable truth about where AI image generation is heading: the artifacts you've learned to spot in this module are getting harder to find. Midjourney v6, released in December 2023, produces hands significantly better than v5 — the hands in many v6 images require much more careful scrutiny. The lighting consistency has improved. Some text on signs is now legible.
This means that the visual artifact checklist you've built is a snapshot of AI's current limitations — not a permanent set of tells. As AI improves, some of these tells will become less reliable. This is why professional forensic organizations increasingly combine visual inspection with other methods: metadata analysis (checking EXIF data embedded in image files), reverse image search (checking whether the image or similar images appear elsewhere), and AI detection software (tools trained specifically to recognize patterns in AI-generated images at a statistical level the human eye can't perceive).
The visual skills you've built are the first line of analysis — the thing you apply in the first 30 seconds when you encounter a suspicious image. They are not the last line. When visual analysis is inconclusive, the professional response is not to guess — it's to say "inconclusive" and apply additional methods.
In 2024, the European Union passed the AI Act, which among other provisions requires that AI-generated images be labeled as such at point of creation. Major social media platforms, including Meta and X, announced policies requiring disclosure of synthetic media in political advertising. These policies exist because visual detection alone is insufficient at scale — millions of images per day can't all be hand-checked. The skills you've built matter for the images you personally encounter. At the institutional level, the same principles are being encoded into law and platform policy.
You've now covered four lessons and three major categories of AI image artifacts. Here is the honest summary of what you can and can't do with this knowledge.
You can: Apply a systematic three-category artifact check to any still image in about two to three minutes. Identify specific, nameable artifacts rather than just a vague feeling that something is "off." Build a case from convergent evidence rather than relying on a single indicator. Recognize when your analysis is inconclusive and further methods are needed.
You can't: Definitively prove any image is AI-generated through visual inspection alone — you can only say the evidence is consistent with AI generation and inconsistent with the alternative explanations. Keep pace with improving AI without continuing to update your knowledge. Inspect video using these still-image techniques — video has its own separate artifact categories (motion artifacts, temporal inconsistency, audio sync) that go beyond this module.
The goal of this module was never to make you certain. It was to make you careful. And careful, systematic inspection — even when it ends in "I can't tell" — is infinitely more useful than confident guessing.
You've built a three-category forensic framework that professional fact-checkers and image analysts use in real newsrooms and human rights organizations. You understand not just what to look for, but why AI fails in those specific places — and why those failures are getting harder to spot over time. You also understand the limits of visual detection, which is something many confident "AI detector" articles don't tell you. Knowing the limits of your own tools is what separates a careful analyst from an overconfident one.
The Taylor Swift deepfake images prompted calls in the US Congress for legislation against non-consensual AI-generated images of real people. Some proposed laws would make creating or distributing such images a federal crime. But the same technology that generated those images can also create valuable art, satire, and fiction involving real public figures. Where is the line between protected creative expression and harmful synthetic imagery — and who has the authority to draw it?
You're the lead analyst at a digital verification unit. An image has been submitted to your organization with a claim that it shows a real event. You need to build a full forensic case — documenting what you examined, what you found, and what your conclusion is — across all three artifact categories: anatomy, lighting, and background.
Your AI partner is acting as peer reviewer. They will challenge your reasoning, ask you to justify calls you can't support with specific evidence, and push you to address alternative explanations before you reach a conclusion.