Instagram reportedly mislabels non-AI photos as AI-generated while missing actual AI content

Instagram’s AI detection system is producing inconsistent results, flagging images as AI-generated when they were edited with assistive tools like background removers or minor retouching. These tools rely on traditional machine learning for object selection, not generative AI, yet the platform’s labeling treats them as equivalent. The issue mirrors a 2024 incident where Adobe metadata triggered false positives, suggesting Meta’s reliance on metadata standards like IPTC and C2PA may be overly broad or poorly calibrated.

The problem extends beyond false positives: actual AI-generated content is slipping through unlabelled, undermining the system’s purpose. Users report that images created with Meta’s own AI tools are correctly flagged, while third-party AI tools evade detection. This inconsistency raises questions about whether Meta’s detection prioritizes internal tools over external ones, or if the system lacks the granularity to distinguish between generative and non-generative edits.

Canva’s role in the mislabeling highlights the challenges of relying on metadata for AI detection. The platform initially admitted its assistive tools were being misclassified as generative, but users report persistent issues even after Canva claimed to fix the problem. The lack of transparency around Meta’s detection methods, such as which metadata fields it scans or how it weights them, makes it difficult to diagnose or correct these errors. Without clearer standards, similar false positives are likely to recur.

For engineers and content creators, the fallout is twofold. False labels risk algorithmic suppression or reputational damage, while undetected AI content erodes trust in platform moderation. The opacity of Meta’s system also complicates efforts to comply with emerging regulations around AI transparency. Until detection methods improve or platforms adopt more precise metadata standards, users may need to treat AI labels as unreliable indicators rather than definitive markers.