How to Tell If an Image Is AI Generated?

Why Your Eyes Can't Detect AI Images Anymore

Look closely at the next "too perfect" photo you scroll past. Chances are you'll get it wrong — and you're in good company.

According to the iProov Deepfake Blindspot Study 2025, only 0.1% of people can reliably identify a deepfake. A peer-reviewed study of 161 participants — including 71 visual professionals — found that AI-generated images are frequently misclassified as human-made, and that fake faces can actually seem more real than real ones. Training your eye, in other words, no longer guarantees anything.

The stakes are no longer hypothetical. In May 2023, a fake image of a building near the Pentagon "engulfed in black flames" caused genuine U.S. stock market turmoil before anyone verified it, as documented in published research on AI-generated imagery. Earlier that year, a Midjourney image of Pope Francis in a Balenciaga-style puffer coat fooled millions who had no reason to doubt a papal photo.

The uncomfortable truth is that the tells most of us learned in 2023 are now obsolete:

• Malformed hands are gone. Modern generators like Midjourney V6 and DALL-E 3 render anatomically correct hands — one fact-checker even wrongly declared an AI image "authentic" because it had a perfect five-finger hand.
• Garbled text is fixed. DALL-E 3 was specifically trained on text accuracy, and Midjourney V6 markets accurate text as a feature.
• Facial asymmetry no longer helps. Today's faces are so convincing that studies find people rate them as more trustworthy than real ones.

As OSINT trainer Henk van Ess notes in the GIJN's detection guide, a journalist trained on 2023 detection methods "might develop false confidence" — and that misplaced certainty is more dangerous than honest uncertainty. Detection has shifted from definitive identification to probability assessment, based on subtler cues like impossible shadows, inconsistent reflections, and "too perfect" aesthetics.

The volume makes this worse. Deepfakes online reached 8 million in 2025, up 16x from 500,000 in 2023, and deepfake fraud attempts grew from 0.1% of all fraud in 2022 to 6.5% in 2025. No human eyeball can win that race alone — which is why the rest of this guide focuses on structured workflows and tools, not intuition. If you want to go further, agent skills can automate parts of a verification workflow, and you can browse and test free ones on SkillzBench.

The New Detection Signals: Anatomy, Physics, and Pixels

The old tells are dead. A fact-checker once wrongly declared an AI image "authentic" because it had a perfect five-finger hand — a mistake that OSINT trainer Henk van Ess warns exemplifies a dangerous new problem: misplaced certainty. Modern generators like Midjourney V6 and DALL-E 3 now render correct hands and accurate text, so detection has shifted to subtler signals.

The GIJN's detection guide organizes these signals into three categories worth internalizing. The first is aesthetic: AI images often look "too perfect," with flawless makeup or airbrushed skin even in crisis contexts where such polish makes no sense. The second involves geometry and physics — the physical world's rules that generators still struggle to respect consistently.

The third category lives at the pixel level, where technical fingerprints reveal a machine's handiwork. Academic researchers are formalizing this space: a peer-reviewed survey categorizes detection into six frameworks, from spatial-domain analysis to multimodal semantic reasoning. For practical purposes, though, the geometric and pixel checks are what you can apply yourself:

• Conflicting shadows — a single light source should cast shadows in one consistent direction; AI images often break this rule.
• Impossible reflections — mirrors, water, and glass that reflect objects which aren't actually in the scene.
• Vanishing point inconsistencies — parallel lines that should converge at one point but converge at several, or nowhere.

Here is the critical framing: none of these signals proves anything. The goal of image verification, per the GIJN, is now "a probability assessment and informed editorial judgment," not certainty. Detection tools face the same constraint — they hit 96% accuracy in lab conditions but drop 45–50% in real-world conditions, according to deepfake statistics research.

That probabilistic mindset matters because the stakes keep rising. A peer-reviewed study found that AI-generated images are frequently misclassified as human-made, and sometimes fake faces seem more real than real ones. Van Ess puts it bluntly: a journalist trained on 2023 detection methods "might develop false confidence," and that certainty is "more dangerous than honest uncertainty."

For teams building verification into their workflows, agent skills can automate parts of the process — routing suspicious images through verification APIs and triggering review alerts. SkillzBench is building a free directory of vetted, benchmarked agent skills for exactly this kind of task, with every skill security-scanned and tested in a sandbox so you know which ones actually deliver a measurable Lift.

Detection Tools and Provenance Signals: What Works and What Doesn't

When a tool tells you an image "looks AI-generated," how much should you trust it? The uncomfortable answer: far less than the marketing suggests. Detection tools that achieve 96% accuracy in lab conditions drop to 45–50% when deployed against real-world images, according to deepfake detection research — a gap that turns a seemingly reliable tool into a coin flip.

That doesn't mean tools are useless. It means you need to understand what each type of tool can and cannot tell you.

Provenance signals: the most trustworthy evidence — with a catch

The most practical tool-based approach right now is provenance verification. OpenAI offers a free Verify tool that checks uploaded images for C2PA metadata and SynthID watermarks, reporting one of three outcomes: C2PA metadata detected, SynthID watermark detected, or no supported signal found. An API is also available for teams that want to build verification into their own workflows.

But provenance has a critical asymmetry that many people miss:

• A detected watermark or C2PA signal is positive evidence the image was AI-generated — you can act on it with confidence.
• "No supported signal found" proves nothing. It does not mean the image is human-made; it may simply have come from a tool OpenAI's checker doesn't cover.
• The Verify tool only detects content generated with ChatGPT, the OpenAI API, or Codex — not images from Midjourney, Stable Diffusion, or other providers.

In short: provenance signals can prove AI origin, but they can never prove authenticity. An image with no watermark is not a real photograph.

Why "just build a detector" is a losing bet

Given the stakes — deepfake fraud attempts grew from 0.1% of all fraud in 2022 to 6.5% in 2025 — it's tempting to think a custom in-house detector would solve the problem. OSINT experts strongly advise against it. Building a serious detection system requires massive computing power and Ph.D.-level teams; without that infrastructure, you'll fail, according to guidance from the GIJN detection guide. Academic research backs this up: the field spans six distinct technical paradigms — spatial-domain, frequency-domain, fingerprint-based, patch-based, training-free, and multimodal reasoning — and is advancing too quickly for ad hoc efforts to keep pace.

The numbers tell a similar story. While some studies report commercial detectors exceeding 98% accuracy in controlled contexts, the lab-to-field collapse means those figures rarely survive contact with messy, real-world images that have been compressed, resized, or screenshotted.

The pragmatic takeaway

Treat every tool output as a probability assessment, never a verdict. A sensible workflow layers provenance checks (fast, reliable when positive) with human judgment on geometry, physics, and context — and reserves expert review for anything financially or legally consequential. Teams that want to automate parts of this pipeline, such as routing suspicious images through verification APIs and triggering alerts, can browse and test free agent skills on SkillzBench, which benchmarks skills in a sandbox so you can see measured performance before you rely on one.

A Tiered Verification Workflow for Teams and Businesses

Most companies discover their image-verification gap the moment a fake invoice photo or AI-generated "executive" video call slips through. Fraud data shows deepfake attempts exploded from 0.1% of all fraud in 2022 to 6.5% in 2025 — a 2,137% increase — while US deepfake fraud losses tripled to $1.1B in 2025. Yet the same research reports that 80% of companies have no deepfake response plan at all.

The fix is not asking employees to spot fakes — only 0.1% of people can reliably identify a deepfake. The fix is a tiered protocol that matches verification effort to stakes, adapted from the GIJN's journalist model for business use.

Tier 1: The 30-second scan (low stakes). For social content, marketing assets, and routine communications, run a quick red-flag check: "too perfect" aesthetics, conflicting shadows, impossible reflections, and vanishing point inconsistencies. Remember that classic tells like malformed hands no longer work — modern generators render them correctly, and as GIJN's Henk van Ess warns, misplaced certainty is "more dangerous than honest uncertainty."

Tier 2: The 5-minute technical check (medium stakes). For anything headed for publication or external distribution, add provenance verification. OpenAI's free Verify tool checks for C2PA metadata and SynthID watermarks — but "no supported signal found" does not mean an image is human-made, and the tool only detects content from OpenAI's own tools. Treat provenance hits as evidence of AI origin, never as proof of authenticity.

Tier 3: Deep investigation (high stakes). Anything financially or legally consequential — payment requests, identity documents, executive communications — gets pixel-level analysis and expert review. The $25.6M Arup scam in Hong Kong shows what skipping this tier costs. Keep in mind that detection tools hitting 96% accuracy in lab conditions drop 45–50% in real-world conditions, so frame every conclusion as a probability assessment, not certainty.

A practical business protocol looks like this:

• Route every inbound image tied to money, identity, or legal exposure to Tier 3 automatically
• Log verification decisions so patterns of fraud attempts become visible over time
• Train teams on the new detection categories — anatomy, geometry, physics, pixel fingerprints — not outdated tells
• Assign clear ownership: someone must be accountable when a suspicious image arrives

Agent skills can automate the routing itself — skills exist that push suspicious images through verification APIs and trigger alerting workflows before a human ever sees the file. SkillzBench is building a free directory of these skills plus a free MCP connector so any agent can be equipped with vetted verification capabilities; the waitlist is live now. Given that only 28% of C-suite executives feel prepared, per WEF data, a documented tiered workflow is cheaper than a single incident.