Instagram: AI Labels Deemed Inconsistent by Users

Key Takeaways

1Real photos published by About Face have been labeled "AI Content," with no trace of SynthID according to a verification based on Gemini.

2Canva mentions confusion between assistive and generative tools, announces a correction, but contradictory reports persist.

3Reported tests with C2PA/SynthID content have not resulted in any labeling by Instagram.

Why it matters : The labels intended to clarify the origin of images on Instagram are being challenged for their inconsistency, with reported false positives and false negatives, while Meta remains vague about its methods.

Accounts claim to see their real photos marked "AI Content" while synthetic images slip through. Amid promises of adjustments, uneven metadata, and a lack of recent responses from Meta, Instagram's system appears unpredictable to users.

Reported tests show unlabelled AI content

Publicly described trials involved images modified with Canva's Background Remover, Photoshop's background removal tool, Adobe Firefly's generative functions, Google's Nano Banana model in Gemini, and iOS's intelligence features. Among these visuals, some were entirely generated, and several included confirmed C2PA and SynthID signals. In this case, Meta did not label any of the tested content. The posts were made from a recently created Instagram account with minimal information and rapidly populated, exhibiting behavior akin to that of a bot or an AI farm. Even after Meta announced measures against AI accounts the previous week, this profile was not classified as AI-generated, and the images were not marked at that time. These contents have remained online and public for almost two weeks. In other reported trials, the label only appeared on images modified or entirely generated via Meta's AI application.

About Face contests the label, and no trace of SynthID appears

Recent photos posted on Instagram by About Face, the brand founded by singer Halsey, were automatically marked "AI Content." When questioned in the comments, the person in charge of the brand's social media stated that no AI had been used to produce these images, specifying that one photo was taken with an iPhone and slightly retouched in the Photos app, reaffirming the use of real artists and real people. When asked to detail the exact tools used in the Photos app, the brand has not yet responded. Only a few iPhone tools are expected to add generative AI metadata, including Spatial Reframing, Extend, and the updated cleaning in iOS 27, which rely on Apple Intelligence and incorporate Google's invisible watermark SynthID. However, a verification based on Gemini indicates the absence of SynthID in the labeled images. Even if Apple Intelligence had been employed, it remains difficult to explain why the label mentions signals associated with entirely generated content rather than retouching.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Canva acknowledges inappropriate labeling and announces a correction

Several reports converge on Canva, particularly regarding its Background Remover tool. After complaints, content strategist Jess Bruno reports that Canva explained to her that some assistive AI tools had been labeled as generative. Canva informed her that its tools now label correctly. Its help page also specifies that using the Background Remover does not add metadata for content generated by Canva's AI, though it is unclear if this mention is recent, and Canva has not responded to a request for comment. Despite this, Threads users claim that using the Background Remover continues to trigger the label on Instagram. Others indicate that none of the images modified before the supposed correction were marked and that the tool did not generate C2PA metadata. Cases of AI labeling are also reported without apparent use of the Background Remover. The confusion is exacerbated by the fact that functions like Background Remover fall under assistive AI, similar to selection and object removal tools used for over a decade, rather than text-to-image generators.

What Meta announced: IPTC/C2PA scans and promised adjustments

Meta indicated in February that Instagram would scan images for IPTC and C2PA metadata, which are supposed to attest to some extent the use of generative AI to create or modify a photo. However, the company has remained vague about its methods. In 2024, Instagram's system reportedly classified images whose Adobe metadata mentioned AI retouching as generative, even when the modifications were very minor. Meta had then promised to adjust the labeling to better reflect the amount of AI used and claimed to rely on standard indicators used by other tool publishers. However, no recent information has clarified the nature of these signals or when and how they are scanned. When contacted, the company did not respond. A previous incident, with the label "Made by AI" appearing in 2024, already illustrated these difficulties.

False positives, false negatives, and uncertainty for users

User feedback describes both false positives—real photos marked "AI Content"—and false negatives, where synthetic images go undetected. The causes mentioned vary by case and, in several situations, show no direct link to generative AI. One user reports that a deliberately "poisoned" image was labeled, while the non-poisoned version of the same visual was not. The lack of clear indications on the signals actually sought by Meta perpetuates confusion, in a context where platforms rarely publish the details of their moderation systems to limit manipulation. For now, users say they do not understand the logic behind the application of labels.