How to Tell If Art Is AI Generated?

Why AI Art Detection Is Getting Harder (and Why Your Instincts Aren't Enough)

If you still trust the "count the fingers" trick, you're already behind. A Nexcess study found that even AI-savvy adults correctly identify AI-generated images only about half the time — and the gap between what people think they can spot and what they actually can is widening every month.

The classic tells are dying fast. Extra fingers, fused digits, and garbled text were reliable red flags in 2023, but generators like Midjourney and DALL-E now render anatomically correct hands and flawless typography, according to a GIJN detection guide. A University of Chicago professor echoed this in CBC News testing: the hand-struggle is simply no longer the case. Some widely shared guidance — like listing hands and vivid colors as warning signs — now lags the actual state of the technology.

The stakes are asymmetric. Traditional fact-checking takes hours or days, while AI misinformation generation takes minutes. In one demonstration, a GIJN author fabricated a complete fake political scandal in 28 minutes for $8. By the time a verification team responds, the image has already spread.

Worse, misplaced certainty is more dangerous than honest uncertainty. Henk van Ess of GIJN recounts a fact-checker who verified a fake as authentic because it showed five fingers — a "tell" that once signaled real photos now proves nothing. Meanwhile, detection tools themselves are unreliable in both directions. A NewsGuard audit found three of five leading tools often flagged authentic images as fake, and in 35 of 45 tested images, at least one tool reached a different verdict than the others. Running multiple detectors may increase confusion rather than clarity.

So what still works? Experts converge on a multi-method approach rather than any single test:

• Structured visual inspection: check lighting and shadow consistency, reflections, textures, and — critically — logical details like whether a button placement actually makes sense in context
• Metadata analysis: image files can contain GPS data, device info, timestamps, and AI-generation tags, though metadata can be stripped or altered
• Reverse image search to trace where an image first appeared
• Source verification and contextual reporting before drawing any conclusion

Treat every detection result as a probability assessment, not proof. Compression, resizing, and unusual lighting can all trigger false positives, and bad actors now run fakes through free detectors iteratively until they pass — or cite detector scores to dispute authentic content.

The same principle applies to the AI tools you use at work: trust should be earned through verification, not vendor claims. That's the thinking behind SkillzBench, which is building independent benchmarking and security scanning for AI agent skills so you can see measured performance before you rely on one. If you're assembling verification workflows, agent skills can automate parts of the process — and you'll soon be able to browse and test free options on SkillzBench.

Why AI Detection Tools Can't Be Trusted on Their Own

If a tool tells you an image is "98.9% likely AI-generated," your first instinct should be skepticism — not about the image, but about the tool. Independent testing shows AI detectors fail in both directions: they miss real fakes and, more damagingly, flag authentic images as fabricated.

A NewsGuard audit put five leading detection tools to the test and found that three of them often declared authentic images fake. ScamAI misidentified 40% of real images as AI-generated, and ZeroGPT got it wrong 20% of the time. Even lightly edited photographs — images that are fundamentally real — were flagged as AI by ScamAI 93% of the time and AI or Not 87% of the time.

The tools can't even agree with each other. In NewsGuard's testing, at least one tool reached a different verdict than the others in 35 of 45 images. That means running an image through multiple detectors — a strategy many people assume adds rigor — may simply multiply your confusion rather than sharpen your judgment.

Accuracy also collapses under real-world conditions. Compression, resizing, unusual lighting, high contrast, and blur all trigger false positives, according to ZeroGPT's own CEO, who recommends pairing any detection result with source verification, reverse image search, metadata analysis, and contextual reporting. Detection models learn the statistical fingerprints of known generators, and those fingerprints lag behind new releases — so images from the latest models may sail through undetected.

Why vendor claims deserve scrutiny

Many detector companies market near-perfect performance: AI or Not claims 98.9% accuracy, ScamAI claims 95.3%, and Hive cites a study it says shows it outperforms human experts. Yet independent audits consistently show far worse real-world results. As one detection vendor candidly put it, anyone claiming 100% accuracy is "marketing to you, not informing you." The gap between marketing and performance is exactly why detectors should never be the final word.

Detectors can be weaponized, too

The risks go beyond honest mistakes. Bad actors run their fake images through free detectors repeatedly, iterating until a forgery passes — turning the very tools meant to catch fakes into a quality-control loop for producing them. Detectors are also used to attack authentic content: NewsGuard documented an X user citing a Hive result of 96.9% "likely AI-generated" to falsely claim a real Netanyahu proof-of-life video was fabricated. A single unreliable score became the basis for dismissing genuine evidence.

The practical takeaway for anyone evaluating artwork or images:

• Treat any detector score as one piece of evidence, never as proof.
• Be wary of tools advertising near-perfect or guaranteed accuracy.
• Expect degraded results on compressed, resized, or heavily shared images.
• Remember that a "high AI probability" score can itself be used to launder fakes or discredit real work.

Verification is ultimately a layered process — visual inspection, metadata, provenance, and context — with detectors playing only a supporting role. The same principle applies to evaluating AI tooling generally: independent, transparent testing beats vendor self-reporting every time. It's the philosophy behind SkillzBench, which is building free, sandboxed benchmarks and security scans for AI agent skills so you can see measured performance rather than marketing claims — and if you want to automate parts of a verification workflow, its free skills directory and MCP connector are designed to help you find and test the right skills for the job.

The Multi-Method Verification Approach: Visual Clues That Still Work

The most dangerous mistake in AI-image detection is false confidence. As GIJN's Henk van Ess warns, "misplaced certainty is more dangerous than honest uncertainty" — a fact-checker once verified a fake as authentic simply because it showed five fingers (per the GIJN detection guide).

The expert consensus is that detection is a probability assessment, not proof. No single method works reliably on its own, so professionals combine visual inspection, metadata analysis, reverse image search, and source verification before drawing conclusions (according to NewsGuard's audit).

Forget the old tells. Extra fingers and garbled text are obsolete — Midjourney and DALL-E now render both correctly, a University of Chicago professor confirmed to CBC (CBC News reports). What still works is checking physics and logic.

Ben Y. Zhao of the University of Chicago recommends asking the question models still fumble: "Does it make sense for the button to be placed in this way?" Small logical details — physics violations, scale mismatches, objects that don't make sense in context — remain genuinely hard for generators (per CBC's testing).

Your inspection checklist:

• Lighting and shadows: do all shadows fall consistently with a single light source?
• Vanishing points: do parallel lines converge correctly across the frame?
• Reflections: do mirrors, water, and glass match the objects they reflect?
• Textures: zoom to 100% and check hair strands, fabric weave, and clothing physics.
• Logic: does every object make sense where it sits — buttons, zippers, handles, joints?

Structure your effort in tiers, as the GIJN guide recommends. First, a 30-second red-flag scan for obvious anomalies. If anything feels off, spend five minutes on technical checks — zoom in, trace shadows, examine hair and fabric at full resolution.

Only escalate to deep investigation when stakes justify it. That means metadata analysis, reverse image search, and source verification layered together. One physicist quoted by GIJN notes AI artifacts are "frequency domain fingerprints" — physics-level signatures, not just visual glitches (per GIJN).

The stakes justify the rigor: even AI-savvy adults identify AI images only about half the time, a Nexcess study found. And speed asymmetry is brutal — traditional fact-checking takes hours or days, while AI misinformation takes minutes. The GIJN author fabricated a complete fake political scandal in 28 minutes for $8 (per GIJN).

If you handle verification at scale, agent skills can automate parts of this workflow — reverse-image lookups, metadata checks, report drafting. SkillzBench (skillzbench.com) is building a free, independent directory where you can browse and test those skills, with security scanning and sandboxed benchmarking so you know which ones actually help.

Step-by-Step Workflow: From Metadata to Source Verification

A single detector score tells you far less than you'd hope. In NewsGuard's audit of five leading tools, the tools disagreed with each other in 35 of 45 tested images — which is why a structured, multi-step workflow beats any one tool. Here's a repeatable process for teams vetting images before publishing or purchasing.

Step 1: Inspect the file metadata. Image files can contain GPS coordinates, device information, timestamps, editing history, and AI-generation tags. Right-click the file and check Properties, or use a tool like CleanImages.online to read what's embedded. Remember that metadata can be stripped or altered, so a clean file isn't proof of authenticity — but an AI-generation tag is a strong signal.

Step 2: Run a reverse image search. Trace where the image has appeared before and whether its claimed origin matches its actual publication history. If an "original" artwork surfaces on multiple marketplaces under different names, or a news photo has no traceable first publication, treat that as a red flag. Reverse image search is a starting point for provenance, not a verdict.

Step 3: Verify the source and context. Combine your findings with source verification and contextual reporting — who published the image, when, and does that story hold up? ZeroGPT's CEO recommends pairing AI detection with exactly this kind of source verification and metadata analysis, because compression, resizing, and unusual lighting all trigger false positives on real images.

Step 4: Run multiple detectors — and log the disagreement. Treat each result as evidence, not proof. In CBC's testing of five free tools, AI or Not and Hive scored three out of three correct while a Hugging Face tool flagged a real image as fake, and Illuminarty scored zero. Collective false positive rates across tools reached 13.33% on authentic images in NewsGuard's audit. Skip any tool marketing 100% accuracy — as one vendor puts it, anyone claiming that "is marketing to you, not informing you."

Step 5: Make a weighted judgment. Stack the evidence: metadata, provenance, source history, and detector consensus. When signals conflict, escalate to deeper visual inspection of logical details — whether buttons, reflections, and physics make sense in context — which remain tricky for generative models to get right.

• Document what you checked and what each tool returned, so decisions are auditable across your team.
• Set a threshold policy: for purchases, require consistent provenance plus detector agreement; for social media posts, a 30-second red-flag check may suffice.
• Re-test periodically — detector accuracy degrades on compressed, resized, or newer-generator images, so last quarter's results won't hold.

For teams running this workflow at volume, agent skills can automate the repetitive parts — metadata extraction, reverse search triggers, and result logging. SkillzBench is building a free directory and benchmarking platform for these skills, with a free MCP connector, so you can browse, verify, and equip only the skills your agents need.

Automating Your Image Verification Workflow with Agent Skills

If you verify images regularly, you already know the workflow: inspect the file, check metadata, run a reverse image search, cross-check multiple detectors, then document what you found. Done once, it's manageable. Done across dozens of images a week, it becomes a grind — and a Nexcess study found even AI-savvy adults identify AI images only about half the time when working unstructured.

That's where AI agent skills come in. These are instruction packages (SKILL.md files) that plug into agents like Claude Code, Cursor, or Gemini CLI and teach them to perform specific tasks. For image verification, relevant skills can automate the repetitive parts of the process:

• Extracting and flagging metadata — GPS data, device info, timestamps, and AI-generation tags — from image files
• Running reverse image searches and summarizing where an image has appeared online
• Cross-checking results across multiple detection tools and logging disagreements in one report
• Producing structured documentation of each verification step, so your conclusions are auditable

The multi-tool cross-check matters especially. A NewsGuard audit found that in 35 of 45 images tested, at least one detection tool reached a different verdict than the others — so systematically comparing outputs, rather than eyeballing them, is exactly the kind of tedious step an agent handles well.

But there's a catch worth knowing before you install skills blindly. Published research found that roughly 26% of agent skills carry at least one vulnerability — a real concern when a skill might handle your files, your API keys, or your published conclusions. Verification workflows demand trust in your tooling, since a compromised skill could quietly skew the very results you're relying on.

This is why we're building SkillzBench (skillzbench.com), a free, independent directory of AI agent skills. Every skill is pinned to a specific version, license-checked, and security-scanned, and each one gets benchmarked in a sandbox — tested with and without the skill, using the same model — so you can see the actual performance lift in percentage points rather than trusting a README's promises. A 0-100 trust score combines security, quality, license, and maintenance signals, and benchmark results are always shown separately so ratings stay honest.

SkillzBench is currently in pre-launch, but you can join the waitlist today and, when the free MCP connector goes live, connect any agent to the verification skills you choose — with per-agent loadouts so your agent only ever sees the tools you've approved. Skills, proven: browse and test free, security-scanned skills at SkillzBench, and let your agent handle the metadata pulls and cross-checks while you focus on the judgment calls that still require a human eye.