How to Detect AI Generated Images?

Why Single AI Image Detectors Keep Failing You

You paste an image into a detector, get a confident "AI-generated" verdict, and share the result — only to find out the photo was real. That scenario plays out constantly, and independent testing explains why.

A NewsGuard audit of five leading tools found that in 35 of 45 tested images, at least one detector disagreed with the others. Three of the five tools frequently misidentified authentic images as AI-generated, with ScamAI producing false positives on 40% of genuine photos and ZeroGPT on 20%.

CBC News ran its own test of five free detectors and found the same chaos. Two tools correctly labeled all three test images on both runs, while two others got all three wrong, and a Hugging Face tool mislabeled the real photo as fake, according to CBC's reporting. Same images, same moment, wildly different answers.

The core problem is a built-in trade-off. NewsGuard found that the tools best at confirming authentic images — Hive and Sightengine, with 0% false positive rates — performed worst at catching manipulated ones, while AI or Not, which caught 100% of altered images, also mislabeled real photos as fake. You can tune a detector to be cautious or aggressive, but not both.

Worse, the tools break in predictable ways:

• Resizing and compression can cause real images to be flagged as AI, per ZeroGPT's own CEO.
• Photos with unusual lighting, high contrast, or blur — common in conflict-zone news images — produce inaccurate results.
• Bad actors can run fakes through detectors repeatedly, iterating until the tool labels them real, as University of Chicago professor Ben Y. Zhao warns.

Detector verdicts can even be weaponized: NewsGuard documented how Hive's high-confidence "AI-generated" rating on a Netanyahu "proof-of-life" video was circulated to falsely claim he had been killed.

Meanwhile, human confidence far exceeds human ability. A global survey found 57% of people believe they can spot a deepfake, even though only 29% know what a deepfake actually is. Hive CEO Kevin Guo puts it bluntly: AI imagery "is moving beyond the threshold of human detection."

No single detector — or eyeball check — can carry the full burden of verification. The experts NewsGuard and CBC interviewed converge on the same advice: treat detector output as one signal among several, cross-check with reverse image search, metadata analysis, and provenance, and never trust a single verdict. The same layered logic applies to AI workflows generally — which is why SkillzBench is building independent, evidence-based verification for AI agent skills, so you never have to trust one tool's word alone.

The Layered Detection Method Experts Actually Recommend

No single tool can tell you whether an image is real. When NewsGuard audited five leading AI image detectors, the tools disagreed with each other on 35 of 45 tested images — and three of the five frequently misidentified authentic photos as AI-generated. That's why experts now recommend a layered method: automation as one signal, never as the verdict.

The layers work together like a checklist:

• Automated detectors — run the image through at least two tools, since accuracy varies wildly between them and even between runs.
• Manual visual inspection — zoom in on asymmetries, hands, and small logical details.
• Metadata and EXIF analysis — check when and where the image was created, and whether editing software left traces.
• Reverse image search — find earlier versions or the original source.
• Provenance checks — look for C2PA Content Credentials or SynthID watermarks, which Google made easier to verify with its SynthID Detector portal in May 2025.

Each layer has expert backing. Dr Kristen Moore of CSIRO advises looking for asymmetries typical of diffusion models — earrings that differ on each side, one eye slightly bigger than the other — and checking hands for the right number of fingers. For face-swaps, she points to blending points on the forehead: textural or colour differences, or an off-looking hairline.

But the classic tells are fading. Prof. Ben Zhao of the University of Chicago notes that malformed hands are no longer reliable, and recommends checking whether small details make sense instead — button placement, or hair blending correctly into a turtleneck. "Smaller details are really still tricky for models to get right," he says.

The final layer is source verification. Rawad Baroud, CEO of ZeroGPT, recommends combining AI-detection results with source verification, reverse image search, metadata analysis, publication history, and contextual reporting before reaching any conclusion. CBC News proved how powerful this is when it debunked a claimed AI-generated campaign photo using raw event footage comparison and metadata analysis — no detector needed.

Context matters too. Baroud warns that resizing and compression can flag real images as fake, and that unusual lighting, high contrast, and blur — common in conflict-zone news photos — produce inaccurate results. The arms race cuts both ways: Zhao cautions that bad actors can run their fakes through detectors repeatedly, iterating until the tools label them real.

For teams building verification workflows around image analysis, agent skills can automate parts of this layering — reverse image lookups, metadata extraction, provenance checks. SkillzBench is building a free directory and benchmarking platform for such skills, with a free MCP connector, so you can equip your agents with only the tools you've vetted. The layered method isn't faster than trusting a single detector, but it's the only approach the evidence supports.

Manual Inspection: Small Details Still Trip Up Generators

Your eyes are still one of the best detection tools available — and unlike software, they can't be fooled the same way twice. Automated detectors frequently contradict each other (in NewsGuard's audit, at least one tool disagreed with the others on 35 of 45 tested images), which makes manual inspection a critical layer of any verification workflow.

Start with asymmetries. Dr Kristen Moore of Australia's national science agency CSIRO notes that diffusion-model images can typically be spotted by looking for "asymmetries, like earrings that are different on each side, or one eye being slightly bigger than the other." For face-swaps specifically, examine the forehead and hairline for blending points — subtle textural or colour differences where the swapped face meets the original head.

Hands remain a classic checkpoint, though a weakening one. Check for the right number of fingers and realistic size and proportion. But Prof. Ben Y. Zhao of the University of Chicago cautions that malformed hands are no longer a reliable tell as generators improve — smaller details are still where models struggle.

Your manual checklist should cover:

• Asymmetries in paired features — earrings, eyes, glasses arms
• Hand and finger anatomy, including size and realism
• Blending points on foreheads and hairlines in suspected face-swaps
• Logical consistency of small details — button placement, hair blending into a collar, background objects

The last point deserves emphasis. Zhao's advice is to ask whether small details make sense: does hair blend unnaturally into a turtleneck? Are buttons evenly spaced and attached to actual buttonholes? Do background objects — light fixtures, reflections, text on signs — hold up under scrutiny?

One caveat: image quality affects your judgment too. ZeroGPT's CEO has acknowledged that resizing and compression can cause real images to be flagged as AI, and images with unusual lighting, high contrast, and blur — common in conflict-zone news photos — produce inaccurate results from tools and humans alike. A heavily compressed social media repost strips away exactly the fine detail you need to inspect.

If you want to build this kind of verification into an automated workflow — combining visual checks, metadata analysis, and reverse image search — agent skills can handle parts of it. SkillzBench is building a free, independent directory of agent skills, each security-scanned and benchmarked, plus a free MCP connector so your AI agent only sees the skills you choose.

Provenance, Watermarks, and Metadata: Verify the Source, Not Just the Pixels

When an image's pixels won't give you a straight answer, its history often will. The most reliable verification layer today isn't a detector at all — it's provenance: where the file came from, what metadata it carries, and whether anyone can vouch for it.

ZeroGPT's CEO recommends combining detection results with source verification, reverse image search, metadata analysis, and publication history before reaching any conclusion, according to NewsGuard's audit of five leading detection tools. That audit found tools disagreed with each other on 35 of 45 tested images, which explains why experts treat provenance as the anchor of any verification workflow. CBC's own debunking of a claimed AI-generated campaign photo relied on raw event footage comparison and metadata analysis — no detector needed (CBC News).

The provenance layer is maturing fast through two industry standards:

• C2PA Content Credentials — an industry standard for cryptographically signed capture and edit history. In September 2025, Google added C2PA support to the Pixel 10 camera and Google Photos (Data Bridge Market Research).
• SynthID watermarks — invisible signals embedded in AI-generated content. Google launched the SynthID Detector verification portal in May 2025.
• Real-world stress tests — AFP combined C2PA certification with invisible encrypted watermarks to verify news photos during the 2024 U.S. elections.
• Standardized evaluation — NIST published an evaluation framework for AI-generated image discriminators in March 2025, alongside its January 2025 deepfake forensics guidance.

Watermarks are powerful but not bulletproof. Google DeepMind has stated that SynthID remains detectable after common modifications like compression and resizing, but is "not foolproof against extreme manipulations." And the approach fails entirely when a generator simply doesn't embed a watermark — a persistent gap in the current ecosystem.

The practical takeaway: treat provenance signals as one layer in a stack, never a verdict on their own. A C2PA credential strongly suggests authenticity; its absence proves nothing, since most images circulating today carry no credentials at all.

For teams building verification into broader AI workflows, agent skills can automate parts of this process — reverse image search lookups, metadata extraction, and cross-referencing detection APIs. SkillzBench is building a free directory of security-scanned, benchmarked skills plus a free MCP connector to equip agents with exactly these capabilities, so verification stops being a manual chore.

Putting It Into Practice: A Repeatable Workflow (and Where Agent Skills Fit)

Knowing what to look for is one thing; doing it consistently is another. The most reliable way to catch AI-generated images is a repeatable routine you run the same way every time, regardless of how urgent the image seems.

Start with the source, not the pixels. Reverse image search, metadata analysis, and publication history often settle the question faster than any detector — CBC debunked a claimed AI-generated campaign photo using raw event footage and metadata alone, no detector needed (according to its own reporting). If the image is compressed, cropped, or unusually lit, treat every automated result with extra suspicion, since NewsGuard's audit found these conditions cause real images to be flagged as fake.

Next, run at least two free detectors and treat their output as one signal, never a verdict. In that same audit, tools disagreed with each other on 35 of 45 tested images, and the tools best at avoiding false positives performed worst at catching manipulations. Then do the human pass experts keep emphasizing:

• Check small details for logic — button placement, hair blending into fabric, locker counts, light fixtures.
• Look for asymmetries: mismatched earrings, uneven eyes, and blending points on foreheads or hairlines.
• Check provenance markers like C2PA Content Credentials and Google's SynthID watermarks, which Google expanded to Pixel 10 and Google Photos in 2025.

Test free detectors on your own samples before paying for anything. Paid plans can cost hundreds of dollars per year, and accuracy varies dramatically between tools and even between runs of the same tool — so run your own benchmark first (CSIRO researchers recommend pairing this with fact-checking against trusted sources).

Finally, consider automating the tedious parts. Agent skills — instruction packages for tools like Claude Code, Cursor, and Gemini CLI — can handle steps like reverse image search, metadata extraction, and cross-checking detector outputs. SkillzBench is building a free directory where every skill is security-scanned, license-checked, and benchmarked with and without the skill, plus a free MCP connector so your agent only sees the skills you choose. Given that published research found roughly 26% of skills carry at least one vulnerability, verified, benchmarked skills are the safer way to automate your verification workflow — and everything on the platform is free, so you can test before you commit.