Why Instagram’s AI Labels Keep Misidentifying Real Photos

Introduction

Instagram’s “AI Content” label is meant to help people recognize images created or substantially altered with generative AI. In recent weeks, however, users have reported the opposite experience: photos made without generative AI are being labeled, while some genuinely AI-generated images are not.

This is not the platform’s first labeling problem. In 2024, Instagram reportedly applied “Made by AI” labels to photographs that had only received limited generative retouching. Meta said it would refine the system to better reflect how much AI was used, but it has provided little public detail about the signals and scanning process behind the labels.

Key points

• Assistive AI is being confused with generative AI. Users say Instagram labels images edited with Canva’s background remover or minor blemish-fixing tools. These features may use machine learning, but they are not equivalent to systems that generate an image from a prompt or add substantial new content.
• Metadata can create inconsistent results. Canva reportedly told one content strategist that some of its assistive tools had been incorrectly treated as generative. The company said the issue had been fixed, although some users continued to report labels after using background removal.
• Generated images can still slip through. Tests described by The Verge found that images created with Google Gemini, including images containing C2PA and SynthID signals, were not labeled by Instagram. The signal that most reliably triggered a label in those tests came from Meta’s own generative AI tools.
• Meta’s detection criteria remain unclear. The company has previously referred to IPTC and C2PA metadata and to industry indicators embedded by other tools. It has not clearly explained which signals it currently checks or how it separates a small AI-assisted edit from an entirely generated image.

Why it matters

AI labels affect more than casual browsing. They can influence how audiences judge photographers, brands, publishers, and creators. A false positive can make authentic work appear synthetic, while a false negative weakens a user’s ability to identify fabricated imagery. When both errors occur at once, the label becomes difficult to treat as reliable evidence.

The technical challenge is growing as AI becomes embedded in everyday editing. Background removal, object selection, cleanup, and other routine functions may rely on machine learning without producing the kind of synthetic image people usually associate with generative AI. Labeling every intelligent edit would make the system too broad; relying only on metadata makes it vulnerable to missing, altered, or incompatible signals.

Meta should explain its labeling categories and distinguish between fully generated images, generative modifications, and ordinary or assistive editing. Until the system becomes more consistent, users should treat the label as a platform-generated indication rather than conclusive proof of how an image was made.

Source: The Verge AI