Instagram Flagged Our Human-Made Post as AI Info

About two weeks ago, we posted a graphic to Instagram about Kate's feature in The Spruce. Alyssa wrote the caption  and created the graphic in Canva.

Instagram flagged the whole post as AI-generated content.

*Cue WTH moment*

We didn't have a clear answer for why at the time, so Alyssa did the simplest thing Google recommended: delete and repost, and the label disappeared. That fixed the immediate problem but didn't answer the question. That fixed the post, but the question remained: Why would a simple background removal tool that we've been using successfully for years be triggering an AI detection on Instagram now?

What's actually happening

With some digging, we've learned that the mechanism behind this is called C2PA Content Credentials, an industry standard built by a coalition that includes both Adobe and Microsoft.

As a file moves through editing tools, it collects a record of what was done to it, which tool did it, and whether a step was performed by an AI model. The idea is that the record helps people trace how a piece of content was made. Some tools write that data in automatically, and Meta reads it and labels the file based on what it finds.

The record is more careful than you would guess from the label. It has separate values for something built entirely by AI and something a person made with AI touching part of it. What it can't say is how much. One synthesized pixel and a wholesale regeneration land in the same bucket, and Meta collapses that bucket into a single label anyway. 

Tools like cropping, manual color correction, and brightness adjustments don't add this metadata. Canva's own help documentation says its background remover shouldn't either, since it's categorized as an "editing tool" rather than a generative one.

Assistive AI and generative AI are not the same thing

We've been using assistive AI tools for years, with spell-check, upscaling, and cropping suggestions. This is the stuff that helps you with a task but doesn't create anything new. Generative AI creates something outright. This could be an image from a prompt, extending (or fully creating) a background. 

Canva's background remover sits somewhere in the middle. For us, it feels assistive in spirit, because it's basically just helping you select the pixels (shout out to Photoshop's magnetic lasso tool of yesteryear), but because the underlying process uses a genAI model to guess where the lines should be, it gets the AI info meta tag. My guess is that's why the tool carries the AI Info meta tag, even though it doesn't feel like that in practice. 

Creator Jess Bruno hit the same wall, and her post about it went viral enough that Canva's VP of Global Comms responded to it directly. She talks about the conversation in this video. The result of their conversation was that they would update their tools so Instagram stops confusing assistive AI and generative AI.

Courtesy of @thatjessbruno on Instagram

A Correction Worth Making

When I wrote about this on LinkedIn, I called the EU's AI disclosure rules blunt, but later learned that isn't accurate. Meta's been tagging posts as "AI Info" (or, initially, "Made with AI") since February 2024, but the regulatory environment has tightened and platforms have gotten twitchier.

I received a comment on this post from user Daniel Green, noting that the EU system is actually tiered, saying "This includes Fully AI Generated, and Partially AI Modified. The problem with Autodetection is evaluating to what degree AI was involved."

Link: EU Icons for labelling AI-generated content

So the bluntness is in the platform implementation. The tool records a category, the platform translates that into a binary yes/no (apparently inconsistently, according to Alyssa's tests), and by the time it reaches a reader, the graphic she built by hand gets flagged the same as an image created by a prompt. 

Where I Land On This

I'm an early adopter by nature and necessity. My role at Queen of To Do covers marketing, IT, strategy, finance, legal, and technical ops. In a lot of ways, I consider my ADHD a superpower, but context-switching across that many disciplines in an afternoon costs me valuable time. 

My Obsidian + Claude setup, a personal knowledge base "Second Brain" coupled with an AI assistant, gave a lot of those hours back. I'm not using Claude to create content outright; I'm using it as a force multiplier and the means to allow our small business to compete at a level we couldn't otherwise play on. 

That nuance isn't communicated with a simple "AI Info" label.

I believe in labelling, but if we want these labels to have actual meaning, we need them to separate 2% AI assisted with 100% AI generated.