AI Art Detection Tools: Do They Actually Work?

When a photographer named Kris Kashtanova submitted their graphic novel to the Copyright Office and disclosed AI involvement, it sparked a national conversation about AI in creative work. Since then, platforms have rushed to build or procure AI detection tools — systems that can automatically identify AI-generated images, text, and audio.

The pitch is compelling: upload your work, get a verdict. Human or AI.

The reality is more complicated.

How Detection Tools Work

Most AI image detectors work by looking for statistical patterns that differ between human-made and AI-generated content. AI image generators tend to produce images with specific frequency signatures, smoothness patterns, and pixel-level regularities that differ subtly from photographs or hand-drawn work.

For text, detectors look for patterns in word choice, sentence structure, and the statistical distribution of language that differs between human writers and language models.

These approaches work reasonably well when the content was generated by well-known models using standard settings. They start to fail in several important ways.

The Failure Modes

False positives on human work. Multiple independent studies have shown that AI detection tools flag genuine human-made content as AI-generated at significant rates. Digital art with clean lines and smooth gradients — common in illustration and graphic design — frequently triggers false positives. Highly polished photography is sometimes flagged. One study found that essays by non-native English speakers were flagged as AI-generated at disproportionate rates because their writing patterns statistically resembled model outputs.

For an artist whose income depends on being recognized as human, a false positive isn't a minor inconvenience. It can cost them contracts, platform access, or their professional reputation.

Evasion through fine-tuning. The models used by open-source communities are being actively fine-tuned to reduce the statistical signatures that detectors look for. When Stable Diffusion or Midjourney updates, detection tools trained on older outputs become less reliable. The cat-and-mouse dynamic means detection is always catching up, never ahead.

Local models leave no trace. Cloud-based AI services often embed metadata that can be detected. Models running locally — RVC, Stable Diffusion, LLaMA-based tools — leave no server-side record and can be configured to produce output with minimal detectable signatures. Detection tools that rely on metadata or API fingerprints are blind to these.

Editing evades detection. Running an AI-generated image through post-processing — color grading, noise addition, compression, composite work — is often sufficient to defeat detection tools. The more post-processing applied, the less confident the detection.

The Platform Consequences

Despite these limitations, platforms are using detection tools to make consequential decisions.

Artists have reported having their portfolios suspended on stock platforms after detection tools flagged legitimate human-made work. The appeals processes are slow and often opaque. In some cases, the only way to appeal is to provide process documentation that the artist doesn't have — because no one told them they'd need it.

This is where the failure of detection as a primary mechanism becomes concrete. It shifts the burden onto the creator in a way that is difficult, sometimes impossible, to satisfy retroactively.

What Detection Can and Can't Do

Detection is useful for identifying obvious, mass-produced AI content from known models with default settings. It provides a reasonable signal in cases where the question is whether content was generated entirely by AI with no human involvement.

It is not reliable for:

• Distinguishing AI-assisted work from AI-generated work
• Identifying outputs from fine-tuned or locally run models
• Correctly classifying highly polished human-made digital art
• Providing the kind of certainty needed for legal or commercial decisions

The underlying technical limitation is fundamental: detection is a classification problem, and the classes are not cleanly separable. Human-made and AI-generated content exist on a spectrum, tools blur the line further, and the models generating content are constantly evolving.

The Better Approach

The reliance on detection reflects a reactive mindset — trying to figure out after the fact whether something is human-made. It will remain a losing position for as long as AI capabilities continue to improve.

Provenance-based approaches work differently. Instead of trying to detect the absence of AI, they establish the presence of human creative process. A documented, timestamped record of creation — generated during the work, not reconstructed afterward — provides evidence that no detection tool can supply and no evasion technique can undermine.

For artists who have been flagged incorrectly, the lesson is painful but clear: process documentation is your defense. A detection tool can say "this looks like AI." A certified process record says "here is exactly how and when a human made this."

Those are not equivalent evidence. The second wins.