Can you check if a picture is AI?

Why Your Eyes and One Detector Aren't Enough

A viral photo lands in your feed, and you have seconds to decide: real or AI? Your gut says one thing, a free detector says another — and the uncomfortable truth is that neither can be trusted on its own.

The evidence on human judgment is sobering. A preregistered study of 1,276 participants published by CACM found people identified AI-generated images with just 49.4% accuracy — worse than a coin flip. Even participants highly familiar with synthetic media scored only 51.9%, and prior knowledge provided no measurable boost. The study authors concluded bluntly that relying on human perception is "no longer a viable bulwark" against synthetic media threats.

Automated tools don't rescue you either. In NewsGuard's audit of five leading detectors, the tools disagreed with each other on 35 of 45 images tested — and three of the five frequently flagged authentic images as AI-generated. ZeroGPT's own CEO advises users to combine detection results with reverse image search, metadata analysis, and source verification before drawing any conclusion.

The classic visual tells you learned in 2023 are also dead. According to the Global Investigative Journalism Network's detection guide, extra fingers and garbled text no longer work: Midjourney and DALL-E now render anatomically correct hands, and DALL-E 3 was specifically trained on text accuracy. University of Chicago professor Ben Y. Zhao confirms the hand problem "is no longer the case," though subtle details still trip models up.

What still gives humans and tools trouble:

• Small logical details — button placement, hair blending into a collar, reflection geometry
• Low-quality, compressed, or heavily filtered images, which trigger false positives in detectors
• Conflict-zone imagery, where unusual lighting and blur mimic AI artifacts

The most dangerous position isn't uncertainty — it's misplaced confidence. As the GIJN guide warns, a journalist trained on outdated methods may declare AI content authentic because it passes obsolete tests. Honest uncertainty beats confident error every time.

The practical takeaway: treat every verdict — yours or a tool's — as one data point, not proof. Layer multiple checks, expect disagreement, and stay current as generators and detectors race ahead. That same layered-verification mindset applies to the AI agents increasingly doing this work; SkillzBench is building a free directory and MCP connector so you can benchmark which agent skills genuinely help with verification workflows rather than taking any single tool's word for it.

Why Detection Tools Disagree: The Arms Race and Its Blind Spots

Run the same photo through three different AI detectors and you'll often get three different answers. According to NewsGuard's audit of five leading tools, at least one tool reached a different verdict than the rest in 35 of 45 images tested.

Why the disagreement? Tool executives point to the mundane realities of how images travel online. ZeroGPT CEO Rawad Baroud attributes false positives to resizing, compression, and images with "unusual lighting, high contrast, and blur" — noting that conflict-zone photography often contains several of these qualities at once. ScamAI co-founder Dennis Ng adds that detectors "generally require high resolution and high pixel levels" to work reliably.

The deeper problem is an arms race. As CBC News reported, bad actors can weaponize detectors themselves — iterating a fake image until detectors label it as real. Every published detector becomes a free testing rig for anyone motivated to defeat it.

False flags cause real harm. An authentic Netanyahu "proof-of-life" cafe video was flagged by Hive as "likely AI-Generated" at 96.9% probability, and pro-Iran accounts seized on that verdict to spread a rumor he had been killed. A light filter that blurred the background apparently triggered the false positive.

The data reveals an uncomfortable trade-off between the two extremes:

• Hive and Sightengine had zero false positives on authentic images — but detected only 73.33% and 33% of heavily AI-edited images, respectively.
• AI or Not caught 100% of significantly edited images — but flagged 87% of lightly edited authentic photos as AI.
• ScamAI, which claims "95.3% detection accuracy," produced false positives on 6 of 15 authentic images.

In other words, tools tuned to never cry wolf are the worst at spotting a wolf in convincing disguise, while the most aggressive detectors risk discrediting genuine content — a weapon in itself when authentic reporting gets labeled fake.

For anyone building verification into a workflow, the lesson is clear: treat detector output as one signal, not a verdict. That layered mindset — combining tools with reverse image search, metadata, and source checks — is the same principle behind how we at SkillzBench approach evaluating AI tools generally, and it's why our free directory benchmarks skills with and without them rather than trusting any single score.

The bottom line: a detector disagreement isn't a glitch. It's the system working exactly as its limitations allow.

The Layered Verification Method Experts Actually Endorse

No single tool, test, or trained eye can tell you definitively whether an image is AI-generated. What experts actually endorse is a layered verification method — treating detection as a probability assessment rather than proof.

The journalism verification community now frames the question as "a probability assessment and informed editorial judgment," acknowledging that perfect detection may be impossible. As one AI detection statistics analysis puts it, a detector score is evidence about a tool's classification, not proof of an image's provenance.

The layered workflow combines several checks that compensate for each tool's weaknesses:

• Run the image through multiple AI detectors and note disagreements as a red flag.
• Perform a reverse image search to trace where the image first appeared.
• Analyze metadata for editing software signatures or missing camera data.
• Verify the source and publication history of whoever shared it.
• Apply contextual reporting — does the image fit the claimed event, location, and timing?

CBC News demonstrated this method while debunking a viral claim that a Mark Carney campaign photo was AI-generated. Detectors alone produced mixed results, but comparing the image against raw event footage and analyzing its metadata settled the question. The NewsGuard audit reinforces why layering matters: in 35 of 45 images tested, at least one tool reached a different verdict than the rest.

Alongside tools, an updated manual checklist catches what detectors miss. GIJN's Henk van Ess advises checking the physics: vanishing points, shadow consistency, and reflection geometry, because "AI assembles images like a collage artist, not a photographer." University of Chicago professor Ben Y. Zhao recommends scrutinizing small details instead of the old tells — "Does it make sense for the button to be placed in this way? Does it make sense for the hair to blend in with the turtle neck that way?" Finally, watch for "too good to be true" perfection in contexts where real photos are usually imperfect.

The verification mindset applies to AI workflows generally: treat any single automated output as one signal, then verify with independent sources before acting. That principle — layered checks, no blind trust in one tool — is exactly what SkillzBench is building into its approach to AI agent skills, benchmarking each skill in a sandbox before it earns a trust score. For images, the same discipline means running detectors, searching provenance, and applying the physics checklist before you share, publish, or debunk.

Your Step-by-Step Image Verification Workflow

Verification isn't about finding one definitive answer — it's about stacking independent checks until the evidence points clearly in one direction. Here's a repeatable workflow your team can run on any questionable image, at work or in the field.

Step 1: Run 2–3 detectors and treat disagreement as a red flag. NewsGuard's audit of five leading tools found that in 35 of 45 images tested, at least one tool reached a different verdict than the rest. If your detectors disagree, don't average the scores — escalate to manual verification. ZeroGPT's own CEO recommends combining detection results with source verification, reverse image search, metadata analysis, and contextual reporting before concluding anything.

Step 2: Reverse-search and check provenance. CBC News demonstrated the value of this approach when it debunked a claim that a political campaign photo was AI-generated by comparing it against raw event footage and analyzing its metadata. A reverse image search revealing an image's first appearance — or the absence of one — often tells you more than any detector score.

Step 3: Verify the source and context. Who published it, when, and what's their track record? Then apply the manual checks that still work:

• Scrutinize small details like button placement and hair blending with clothing — still tricky for models, per University of Chicago's Ben Y. Zhao.
• Check physics: vanishing points, shadow consistency, and reflection geometry, as GIJN's reporter's guide advises.
• Discard outdated 2023-era tells like extra fingers and garbled text — they no longer work.

Step 4: Double your skepticism for low-quality and conflict-zone imagery — in both directions. Detector executives themselves acknowledge that compression, resizing, and unusual lighting trigger false positives, and NewsGuard documented an authentic Netanyahu video falsely flagged at 96.9% probability and weaponized to spread a death rumor. A confident detector verdict on a blurry war photo proves very little, either way.

Finally, guard against misplaced certainty. A colleague trained on 2023 detection methods may confidently declare AI content authentic because it passes outdated tests — honest uncertainty beats confident error. And don't over-trust your own eyes: a preregistered study of 1,276 participants found human detection accuracy of 49.4%, below a coin flip.

For teams adopting AI, much of this workflow can be automated — agent skills exist for reverse-searching, metadata inspection, and multi-detector checks. SkillzBench is building a free directory of vetted agent skills plus a free MCP connector (currently in active development — join the waitlist to browse and test them early), so verification becomes a routine pipeline step rather than a scramble every time an image looks suspicious.