How to Spot AI Generated Images?

Why Your Old Detection Tricks No Longer Work

If you learned to spot AI images by counting fingers, 2025 has quietly made that skill worthless. Midjourney and DALL-E have largely fixed hand rendering, and OpenAI and Midjourney now train their models specifically on accurate text — the two tells everyone relied on in 2023 are simply gone, according to a reporter's guide from the Global Investigative Journalism Network.

That guide's warning should sting anyone who feels confident about their detection skills: false confidence from outdated training is "more dangerous than honest uncertainty." The evidence backs this up. Only 0.1% of people correctly identify every deepfake shown to them, and humans detect high-quality deepfake video at an average accuracy of just 24.5% per iProov research. Even participants explicitly told to watch for synthetic content performed poorly.

The tools many people reach for next don't solve the problem either. A NewsGuard audit of leading AI image detectors found the tools collectively declared authentic images AI-generated 13.33% of the time — and in 35 of 45 tested images, at least one tool reached a different verdict than the rest. The tools that caught manipulation best were the most likely to flag real photos as fake, and vice versa. In one striking case, a genuine Netanyahu "proof-of-life" video was labeled AI-generated with 96.9% confidence, apparently because of a light filter edit.

For business owners and managers, this isn't an abstract media-literacy problem. The stakes are financial and rising fast:

• Deepfake fraud in North America surged 1,100% year over year between Q1 2024 and Q1 2025, per Sumsub data.
• The average cost of a deepfake fraud incident hit $450,000 in 2024 — nearly double the $230,000 recorded in 2022.
• A Hong Kong finance worker transferred $25 million to impostors posing as a CFO and colleagues on a deepfake video call.
• 62% of organizations experienced a deepfake attack in the past 12 months, yet 61% of executives say their companies have no protocols for deepfake risks.

Meanwhile, the speed asymmetry keeps widening. Verification trainer Henk van Ess notes that traditional fact-checking takes hours or days, while AI misinformation takes minutes — he built an entire fake political scandal in 28 minutes for $8. Detection researchers describe the field as a never-ending arms race; a review of 400+ studies found no method generalizes reliably to generators it hasn't seen before.

The practical shift is this: stop hunting for a single tell and start treating every suspicious image as a probability assessment. Geometry and physics — vanishing points, shadow consistency, reflections — remain the most durable checks, because as van Ess puts it, "AI assembles images like a collage artist, not a photographer." And because that assessment is a multi-step workflow, parts of it can be automated: reverse image searches, metadata checks, and source verification can run as agent skills, and free directories like SkillzBench (currently in development) are being built to help you find and test which skills actually perform.

The Tiered Workflow: From 30-Second Gut Check to Deep Verification

If you can only remember one thing about spotting AI images, make it this: there is no single "tell" anymore. The extra fingers and garbled text that worked in 2023 have largely vanished — Midjourney and DALL-E had fixed hand rendering and text accuracy by 2025, according to a reporter's guide from the Global Investigative Journalism Network. What replaces the checklist is a tiered workflow that scales your effort to the stakes.

The first tier takes 30 seconds and relies on a counterintuitive instinct: suspicion of perfection. Verification trainer Henk van Ess warns that "too perfect" is itself a red flag — magazine-quality aesthetics in the wrong context, like flawless makeup on a disaster victim. Real faces carry subtle asymmetries, pores, and wear patterns that generators struggle to replicate.

If something feels off, the second tier is a five-minute technical pass. Zoom to 100% and inspect the details AI still fumbles:

• Faces and skin: look for the plastic-smooth texture that lacks pores and natural imperfection
• Hair strands and teeth: real hair strays; real teeth have irregularities AI tends to homogenize
• Jewelry and reflective objects: check whether metal and gems hold convincing 3D realism
• Clothing physics: fabric should drape, fold, and wrinkle in ways consistent with the body and scene

The most durable checks are geometric, because — as van Ess puts it — "AI assembles images like a collage artist, not a photographer." Parallel lines must converge to a single vanishing point, all shadows must point away from one light source, and lines connecting objects to their reflections should meet surfaces at right angles. The viral Pope Francis puffer-jacket image, for instance, was partly exposed by a crucifix chain held aloft with half the chain inexplicably missing.

The third tier — comparative photo analysis, pixel-level forensics, or experts like Hany Farid — belongs to high-stakes cases only. And whatever tier you reach, treat the conclusion as a probability assessment, not a verdict. Human intuition alone is weak: only 0.1% of people correctly identify every deepfake they're shown, per aggregated industry research. Detection tools don't rescue you either — an independent NewsGuard audit found the leading tools declared authentic images AI-generated 13.33% of the time and disagreed with each other on 35 of 45 tested images.

That's why professionals pair any tool output with reverse image search, metadata analysis, and source verification. Because this verification is a multi-step, repeatable workflow, parts of it — like running reverse image searches or metadata checks — can be automated with agent skills, and SkillzBench is building a free directory plus a free MCP connector so you can browse and test skills for exactly this kind of pipeline. In a landscape where a fake political scandal can be fabricated in 28 minutes for $8, a structured workflow beats both blind trust and false confidence every time.

Geometry and Physics: The Most Durable Visual Cues

The fingers lie. Extra fingers and garbled text — the tells everyone learned in 2023 — are now largely obsolete, because Midjourney and DALL-E have fixed surface rendering problems, according to the GIJN reporter's guide. What generators still struggle to fake is deeper: the physics of a scene.

Verification trainer Henk van Ess frames it memorably: "AI assembles images like a collage artist, not a photographer." A collage artist pastes convincing pieces next to each other; a photographer captures one scene governed by one set of physical laws. That distinction is why geometry outlasts texture fixes.

Three structural checks are hardest for generators to fake:

• Vanishing point analysis — parallel lines in a real photo (building edges, floor tiles, road markings) must converge to a single point. AI images often give each object its own, conflicting perspective.
• Shadow consistency — every shadow in a scene must point away from one light source. Mismatched shadow directions are a classic collage giveaway.
• Reflection geometry — lines connecting objects to their reflections should converge and meet the reflecting surface at right angles. Faked mirrors and water rarely obey.

The famous Pope Francis "Balenciaga puffer jacket" image illustrates how structural slips expose even viral fakes. Beyond its dreamlike polish, viewers spotted a crucifix chain inexplicably held aloft with half the chain missing — a physical impossibility that no amount of surface realism could patch.

That's the key insight: surface artifacts get patched in the next model release, but a scene that violates geometry requires the generator to understand the scene, not just render it. Academic reviews of 400+ detection studies similarly note that robust detection research centers on generalizability — catching generators the model has never seen — precisely because surface-level tells don't transfer, per a systematic review in Forensic Science International: Synergy.

For teams building verification workflows, these checks are procedural enough to partially automate — reverse image searches, metadata analysis, and consistency checks can run as agent steps. SkillzBench is building a free directory and MCP connector for exactly that kind of agent skill, so you can browse and test verification skills as the platform takes shape.

Until then, trust the physics. Hands heal; geometry doesn't.

Don't Trust AI Detectors Blindly: What the Evidence Shows

Paste any suspicious image into an AI detector and you'll get a confident percentage. The problem? That number may be wrong more often than you'd expect.

NewsGuard's independent audit of five leading detection tools found they collectively declared authentic images AI-generated 13.33% of the time. Worse, in 35 of 45 tested images, at least one tool reached a different verdict than the rest. Consensus, not confidence, is what you should look for.

The audit exposed a troubling trade-off. Tools that caught manipulation best (AI or Not flagged 100% of significantly manipulated images) also misflagged real ones, while tools that got every authentic image right performed worst at spotting manipulation. There's no single tool that wins both ways.

False positives carry real consequences. Bad actors exploit them to dispute reality — a Netanyahu "proof-of-life" video was falsely labeled AI-generated by one tool at 96.9% probability, apparently because of a simple filter edit. A wrong "AI" verdict can be weaponized just as easily as a fake image.

Treat any detector output as one signal, never a verdict. Even ZeroGPT's own CEO advises combining detection results with source verification, reverse image search, metadata analysis, and contextual reporting. A layered approach looks like this:

• Run two or three detectors and note where they agree or conflict — disagreement is itself a red flag.
• Reverse image search to find earlier appearances and original contexts of the picture.
• Check metadata (EXIF data, creation dates, editing software traces) for consistency.
• Verify the source: who published it, when, and do they have a credible track record?

This layered workflow is exactly the kind of multi-step verification that agent skills can partially automate — running reverse searches or metadata checks on demand. SkillzBench is building a free directory and MCP connector so you can browse, benchmark, and equip exactly those skills for your own verification stack.

Detection remains, as analysts describe it, "a never-ending back-and-forth" with no silver bullet. Your best defense isn't a better tool — it's a process that assumes any single tool can be wrong.

Build Your Verification Habit (and Automate the Busywork)

A single suspicious image can now cost a business $450,000 — the average cost of deepfake fraud, which has nearly doubled since 2022. And with 62% of organizations reporting a deepfake attack in the past year, verification can't depend on one person's intuition. It needs to be a habit, with clear roles and escalation paths.

The good news: the workflow is teachable. Verification trainers recommend a tiered approach that scales effort to stakes:

• 30-second red flag check: Is the image "too perfect"? Magazine-quality aesthetics in inappropriate contexts — flawless makeup on a disaster victim, for instance — deserve skepticism.
• 5-minute physics check: Zoom to 100% and verify vanishing points (parallel lines should converge to one point), shadow consistency (all shadows should point away from a single light source), and reflections. AI "assembles images like a collage artist, not a photographer," so geometry violations are harder to fake than surface details.
• Deep investigation: For high-stakes media, run reverse image searches, check metadata, compare against known photos, and escalate to forensic experts.

Then define who does what. A practical policy for teams: front-line staff run the 30-second check on anything suspicious and escalate anything they can't clear in five minutes. A designated verifier (finance for invoices, IT for video calls) owns the 5-minute check and provenance work. Anything touching money or reputation goes to the escalation path — and for payment requests, verification happens out-of-band, through a known phone number, never through the channel the request arrived on. The stakes justify it: a Hong Kong finance worker transferred $25 million to deepfake impostors posing as a CFO on a video call, and North American deepfake fraud surged 1,100% year over year.

One caution: don't lean on detection tools as the final word. An independent audit found leading detectors declared authentic images AI-generated 13.33% of the time, and the tools disagreed with each other on 35 of 45 tested images. Treat tool output as one input alongside reverse image search and source verification — never as a verdict.

Finally, automate the repetitive parts. Reverse image searches, metadata extraction, and provenance checks are exactly the kind of multi-step busywork that AI agent skills handle well — freeing your people to focus on judgment calls. You can browse and test free agent skills for workflows like these on SkillzBench's directory, with its MCP connector for wiring skills into your agents currently being built ahead of launch. Given that only 0.1% of people correctly identify every deepfake they're shown, a well-designed process beats raw intuition every time.