Is This Art AI Generated?

Why Your Eyes Can't Tell Anymore: The AI Art Detection Problem

Five fingers used to be the tell. Today, a seasoned fact-checker confidently declared an AI image authentic precisely because it showed five fingers — a mistake that captures just how badly our visual instincts have fallen behind the technology.

The core problem is simple to state and hard to solve: visual inspection alone is no longer reliable. AI image models now produce works nearly indistinguishable from human-made ones, and a peer-reviewed perception study found that humans were only above-chance — not certain — at identifying which artwork in a pair was AI-generated. Strangely, the same participants showed a significant preference for the AI-generated works, overturning the assumption that people instinctively dislike machine art.

The 2023-era tricks are gone. Counting fingers, spotting garbled text, checking misaligned ears — none of it works against current models. Midjourney and DALL-E 3 now render anatomically correct hands and accurate typography, and Midjourney V6 even marketed "accurate text" as a headline feature, according to the GIJN verification guide.

What still works is forensic and provenance-based, not intuitive:

• EXIF metadata analysis as first-pass triage — AI images from major generators often carry a `trainedAlgorithmicMedia` source type, while stripped or transplanted metadata proves nothing either way.
• C2PA content credentials — cryptographically signed manifests chaining origin and editing history, though most consumer cameras don't yet embed them.
• Structured visual analysis — zoom to 100%, trace perspective lines, check shadow direction and mirror reflections against the object reflected.
• Invisible watermarks like Google's SynthID, which survives compression and cropping — but detection requires access to Google's proprietary classifier.

The urgency is commercial, not academic. The generative AI in art market grew from $0.43 billion in 2024 to $0.62 billion in 2025 — a 42.4% CAGR — and is forecast to reach $2.51 billion by 2029. Publishers, buyers, and businesses adopting AI at work will encounter this output constantly, and no single tool or tell settles the question. Detection has shifted "from definitive identification to a probability assessment and informed editorial judgment."

There's also a darker twist: the Liars' Dividend. Because deepfakes exist, bad actors can dismiss authentic evidence as AI-generated — so proving authenticity now matters as much as spotting fakes. Forensic examiners can't rely on gut instinct; they need repeatable, evidence-based methods, and even the most experienced examiners are challenged by generative AI fakes.

For teams building verification into their workflows, the layered checks above can be chained and automated with agent skills — SkillzBench is building a free directory and benchmarking platform so you can see which skills actually lift performance before you trust one. In a market where misplaced certainty is more dangerous than honest uncertainty, treating every verdict as probabilistic is the only defensible posture.

The Layered Verification Method: From Gut Check to Forensics

If a seasoned fact-checker can confidently declare an AI image authentic because it shows five fingers, none of us can trust our instincts alone. That's why journalism trainer Henk van Ess, author of the GIJN verification guide, warns that "misplaced certainty is more dangerous than honest uncertainty" — and why the answer is a layered workflow, not a single test.

Tier 1: The 30-second red-flag check. Start with the gut question: is this image too perfect? Flawless skin, impossibly even lighting, and a scene that feels "too good to be true" all warrant suspicion. Forensic examiners at Envista Forensics recommend maintaining "a posture of incredulity" — but treat this pass as a filter, not a verdict.

Tier 2: The five-minute technical pass. Zoom to 100% and inspect details the eye skips at full size. Van Ess's workflow covers three checks worth memorizing:

• Zoom to 100% and examine textures, edges, and fine patterns for smearing or incoherence.
• Trace perspective lines — AI images often bend architecture and geometry subtly.
• Check shadow direction and mirror reflections; Envista flags reflections that differ from the object reflected as a classic AI tell.

Tier 3: Provenance and metadata. This is where forensic security researcher Kenneth Hartman's evidence-based approach earns its keep. Run EXIF triage with a tool like ExifTool: a genuine Google Pixel 8 Pro photo carried camera settings, GPS, and "Directly photographed" scene data, while a Google Imagen output was labeled with the C2PA source type trainedAlgorithmicMedia. C2PA content credentials cryptographically chain origin and editing history, and Google's SynthID invisible watermark survives compression, cropping, and metadata removal.

One caution: metadata alone proves nothing. Hartman demonstrated the "EXIF Transplant Problem" by copying Canon EOS metadata onto a Pixel photo with unchanged pixels — so never treat any single signal as conclusive. For high-stakes cases, request the original device or file and escalate to forensic experts.

The final mindset shift: verification is a probability assessment, not a definitive identification. Teams running these checks at scale can automate parts of the workflow with agent skills — SkillzBench is building a free directory and MCP connector for exactly that kind of verification automation, with security scanning and benchmarking built in. Layer the methods, stay skeptical in both directions, and let the evidence accumulate.

Know the Limits: Why Every Signal Can Be Faked

Here's the uncomfortable truth about every verification method we've covered: each one can fail, and some can be actively weaponized. Treating any single signal as conclusive is where well-intentioned verification goes wrong.

Consider the EXIF Transplant Problem. Metadata forensics expert Kenneth G. Hartman demonstrated that ExifTool can copy an entire metadata set from one image to another — he transplanted Google Pixel 8 Pro camera data onto a different image, replacing it with Canon EOS Rebel T100 references while the pixels stayed untouched. His conclusion is blunt: EXIF "cannot serve as the sole basis for authentication in forensic or legal contexts." And the inverse problem is just as telling — stripped or missing metadata proves nothing either way. Many legitimate photos, and many GAN-generated images like those from thispersondoesnotexist.com, arrive with virtually no metadata at all.

SynthID, Google's steganographic watermark, sounds more robust — it survives compression, cropping, and format conversion. But there's a catch: detecting it requires access to Google's proprietary classifier. If you don't have that access, an unwatermarked image tells you only that it didn't come from a Google tool (or that the watermark didn't survive). C2PA content credentials face a coverage gap too: most consumer cameras don't yet embed them, so their absence is expected for perfectly authentic photos.

Even detector tools inherit the weaknesses documented in text-detection research. A University of Chicago study found that all detectors lost accuracy on passages under 50 words, and open-source models sometimes performed close to random guessing — a reminder that detection is a "technical arms race" requiring regular published audits. The University of San Diego Law Library puts it more starkly: AI detectors are "neither accurate nor reliable," and there's likely no silver bullet coming.

The failure mode that worries verification experts most isn't a missed fake — it's misplaced confidence. Journalism researcher Henk van Ess cites a seasoned fact-checker who confidently declared an AI image authentic because it showed five fingers, and warns that "misplaced certainty is more dangerous than honest uncertainty." The "Liars' Dividend" compounds this: the mere existence of AI imagery lets bad actors dismiss authentic evidence as fake, so proving authenticity matters as much as catching forgery.

The honest framing: verification is a probability assessment, not a verdict. Layer your signals, weigh them together, and stay skeptical of "authentic" conclusions just as much as "fake" ones.

• Metadata present? Useful signal. Metadata absent or transplanted? Inconclusive.
• C2PA credentials found? Strong evidence. Not found? Expected for most consumer cameras.
• SynthID detected? Definitive. Not detected? Only meaningful if you can run Google's classifier.
• Detector says "AI"? Treat it as one vote, not a ruling — accuracy varies sharply by tool and input.

For teams running these checks at scale, agent skills can chain metadata analysis, provenance lookups, and detector calls into a single repeatable workflow — SkillzBench is building a free directory where you can browse and benchmark exactly those skills.

Your Practical Verification Playbook (and How to Automate It)

A gut feeling isn't evidence. Even trained fact-checkers get burned — one seasoned verifier confidently declared an AI image authentic because it showed five fingers, a check that no longer works against 2025-era models. What you need instead is a repeatable playbook, not instinct.

Start with the 30-second red-flag scan. Ask whether the image is suspiciously perfect: skin too smooth, hair too flawless, mirror reflections that don't match the object reflected, or shadows going the wrong direction. This won't settle anything, but it sets the posture of incredulity forensic examiners recommend.

Then run the five-minute technical pass:

• Zoom to 100% and inspect edges, textures, and compression artifacts.
• Trace perspective lines to confirm they converge correctly.
• Check shadow direction and darkness for physical consistency.
• Compare mirror reflections against the objects they should show.

Next, check provenance with free tools. Run ExifTool for EXIF metadata as first-pass triage: a genuine camera photo typically carries lens, exposure, and "Directly photographed" scene data, while AI images from major generators often carry a trainedAlgorithmicMedia source type. Then run c2patool to look for C2PA content credentials — cryptographically signed manifests that chain assertions about origin and editing history. One critical caveat: metadata can be trivially transplanted between files, so it can never serve as sole proof of authenticity.

For high-stakes cases, escalate further. Request the original device or file — camera identification and compression analysis can link a photo to its originating device through artifacts invisible to the human eye. If the image still "feels off" and the stakes justify it, hand it to digital forensics experts; even the most experienced examiners are now challenged by generative AI fakes.

The honest framing: verification has shifted from definitive identification to probability assessment. There's no silver bullet, and there likely never will be — detection is an arms race, which is why researchers recommend regular published audits of any detector tooling.

This is also where automation earns its keep. Each step above is a discrete check, and agent skills can chain them into a repeatable verification workflow — run the red-flag scan, call ExifTool and c2patool, log the findings, flag for escalation. SkillzBench is building a free directory where you can browse and test skills for exactly this kind of chaining, with every skill security-scanned and benchmarked so you know which ones actually help. Given that published research found roughly a quarter of agent skills carry at least one vulnerability, that verification layer matters as much for your tools as for your images.

Treat every single-signal verdict — human or tool — as inconclusive, and be as skeptical of "authentic" calls as of "fake" ones. Misplaced certainty is more dangerous than honest uncertainty.