AI Image Detection Is Failing: Why Even Experts Can't Spot Fakes 2026

AI image detection has become nearly impossible, according to Daniel Púa, head of security at Magnific (formerly Freepik). In a recent interview, Púa admitted that "practically no person at first glance could be 100% sure whether it's AI or not AI." This admission underscores a growing crisis of trust in visual media.

The problem intensified late last year when Google released Nano Banana Pro, a model that produced images indistinguishable from reality for many viewers. As generative models improve at breakneck speed, even daily reviewers lack a foolproof trick to spot fakes.

Why AI Image Detection Is Failing

Púa explained that detectability "depends on the model used and the prompt given." The level of detail in modern AI images has reached a point where the human eye is no longer a reliable tool. Subtle cues like lighting, texture, and anatomy that once betrayed AI are now often flawless.

This shift has led to collective distrust: people doubt all images, even authentic ones. The implications for journalism, social media, and legal evidence are profound.

The Role of Model and Prompt

Not all AI images are equally undetectable. Púa noted that detection difficulty varies by model and prompt. Some models still leave artifacts, especially when prompts are complex or unusual. However, top-tier models like Nano Banana Pro have closed most of these gaps.

Video: The Next Frontier for AI Detection

While still images are largely lost territory, video remains somewhat easier to detect. Púa said, "In video, it's easier to detect because of the sequence of how it advances, certain physics that don't make sense... it's two steps behind."

Yet video generation is advancing rapidly. As models improve temporal consistency and physics simulation, video may soon follow the path of images into undetectability.

What Makes Video Detectable Today

• Physics anomalies: objects moving in ways that defy real-world physics
• Temporal inconsistencies: flickering or morphing between frames
• Unnatural motion blur or lack of it
• Audio-visual sync issues in generated speech
• Repetitive patterns or textures

These cues are shrinking as models like Sora and Runway improve. The window for easy video detection is closing.

Comparing Detection Difficulty: Image vs. Video

As the table shows, video still offers some visual tells, but experts expect those to diminish quickly.

Implications for Trust and Security

The inability to detect AI images erodes trust in all visual content. Púa's comments highlight a security challenge: if even experts can't tell, how can platforms, courts, or the public verify authenticity? The background to this story is set out in AI-Generated Images: Why Photojournalism Ethics Matter Now. A related thread runs through AI Images Threaten Elections as Deepfakes Spread Rapidly. Earlier coverage in AI-Generated Images: Expert Answers on Photography Ethics shows how the argument developed.

Solutions may include watermarking, provenance metadata, and detection tools, but none are foolproof. Magnific and other companies are investing in security measures, yet the arms race continues. A closer look at the same subject appears in UK Children Face Surge in Explicit Deepfake Images, Watchdog. Anyone following the wider picture can start with Meta Ordered to Remove Deepfakes: AI Video Safeguards Inadequate. The same theme surfaces in Getty Images Bankruptcy: How Generative AI Is Disrupting Stock Photography, which takes a different angle.

What to Watch Next

Watch for advances in AI video generation, which will likely make detection even harder. Also monitor regulatory efforts to mandate labeling of AI content. As Púa suggests, the focus may shift from detection to verification and authentication. Context on how this played out before sits in Fake RingGo Text Scam: How to Spot and Avoid It. A companion piece, Spotify Scam Emails: How to Spot Fake Payment Requests, approaches the same subject from another angle.

FAQ