Loading today’s stories
Product1 distinct publisher3 min readPublished
Users say Meta's label lands on ordinary retouching while genuinely generated pictures go untagged, which leaves brands publishing edited photography to explain a badge nobody at Meta will explain to them.
The Product Desk · Product desk
Compiled by The Product DeskSomething wrong?How this is made
Follow any of these and your For You feed starts watching them — no settings page required.
product
Meta's Mac app is a data connector wearing a chatbot's clothes6 distinct publishers
build
Meta's Hatch asks for account connectors in the first of three onboarding steps1 distinct publisher
invest
Meta's internal agent gets its own computer to run errands across the open web1 distinct publisher
product
Meta's off-platform ad pipeline is default-on, and the off switch comes in four parts1 distinct publisher
One user, The Verge reports, ran the same picture two ways: the copy that had been poisoned, meaning subtly altered so it degrades models trained on it, picked up the AI Content label, and the ordinary version did not [11]. Whatever the scanner picked up on there, it tracked something other than how the image was actually made.
Meta's stated method dates to February 2024, when it said it would scan for IPTC and C2PA metadata, which can indicate with reasonable certainty whether generative tools created or manipulated a picture [6]. Beyond that it has described its inputs as industry standard indicators that other companies include in content from their tools, with no recent detail on which ones, or when they get read [7]. The feature broke the same way in 2024 under the name Made by AI, catching photographs whose Adobe metadata showed generative retouching so slight the result was substantively the same photograph [5]. Meta's answer then was a promise to better reflect the amount of AI used in an image [9].
The grid worth drawing before anyone leans on this has two axes: whether a generative tool touched the file, and whether the badge appeared. User reports put cases in both of the wrong corners [18]. A signal that errs in both directions is tracking metadata survival rather than provenance: whether a piece of metadata made it from an editing app to Meta's scanner, not whether a generative tool touched the file, and that trip is the part no publisher controls.
Background removal is assistive machine learning of the sort object selection tools in Photoshop have used for more than a decade, not text-to-picture generation [10]. The label does not appear to separate the two, and by Canva's own account to content strategist Jess Bruno, some of its assistive tools were being tagged as generative before it corrected them [12].
The label reads as though it warns the audience when a picture is synthetic, which would let a studio shooting its own product photography treat it as nothing to manage. In practice it annotates whichever tool wrote metadata last, and the person who answers for it is a social manager in a comment thread [17].
The reporting does not quantify any of this. There is no count of affected accounts and no reported consequence to reach or ad eligibility, so the costs so far are reputational rather than tied to media spend [19].
So the practical unit of provenance here is the editing tool that touched the file; the image itself carries no such record. A two-column list gets a team most of the way: for every app in the retouching pipeline, whether its exports carry AI or C2PA metadata, and who owns the reply when the badge shows up anyway. The raw capture and a short edit log are worth keeping for the same reason a receipt is. And an unlabeled competitor image should be read as unlabeled and nothing more, because on the current evidence the absence of the badge carries no information at all [3].
Ranked by verification strength, evidence, and original report placement.
Over recent weeks, users have reported that Meta has been automatically applying an "AI Content" label to images they did not create or edit using generative AI tools.
Many users said the label appeared on images edited with tools like Canva's Background Remover or with negligible use of blemish-fixing tools; one Threads user said it appeared "every time there is bg remover involved," and another said it appeared after using Canva to remove "a speckle."
Instagram's visible AI labels are intended to help people quickly spot synthetically generated content at a glance.
A similar problem hit Instagram photography in 2024 with a "Made by AI" label a few months after release; the detection system reportedly swept up pictures whose Adobe metadata indicated generative retouching, even where the changes were extremely minor and substantively produced the same photograph.
In February 2024 Meta announced it would scan images for IPTC and C2PA metadata, which can prove with reasonable certainty whether generative AI was used to create or manipulate them.
Meta has said it uses "industry standard indicators that other companies include in content from their tools," but there is no recent information on what they are, or how and when it scans for them.
Distinct publishers with included, body-backed reporting in this cluster.
1 article · September 4, 2026
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One newsroom's hands-on against a crowd of anonymous posts
The firmest thing in this story is the part the reporter did personally: The Verge ran the editing tools and could only make the badge appear via Meta AI. Everything that points the other way — the background remover, the removed speckle — comes in as unnamed Threads posts, and Canva's explanation reaches the page second-hand through a content strategist. Meta was asked and said nothing; Canva was asked and said nothing; About Face has not answered either. Real, verifiable at the edges, thin exactly where the causal claim sits.
Platform-wide by construction, unmeasured in practice
This is not a pilot anyone opted into: the tag attaches automatically to ordinary uploads, and it attached to a real brand's photos and to the reporter's test posts. What is missing is a denominator. No count of affected accounts or posts appears anywhere, and no consequence to reach, recommendation, or ad eligibility is reported — so the reach is broad while the impact stays anecdotal.
The causal story outruns what testing showed
"A mess (again)" is fair as a description of user confusion; it is weaker as a finding about background removers. The single controlled test in this reporting reproduced the badge only through Meta AI, Canva says the mislabeling was its own metadata bug and is fixed, and the users who supposedly prove the pattern contradict each other about whether Canva edits were ever tagged. The badge landing where owners say it shouldn't is documented; the mechanism named in the framing is not.
Nobody holding the answer wants to give it
Meta will not describe what it scans for, a silence it can justify as keeping bad actors from gaming detection, and which conveniently also shields a system misfiring in public. Canva's version — our tools were wrongly emitting generative markers, and we have fixed it — cleared Canva and arrived through a third party, after which Canva stopped talking. Facing them are brands and creators with an obvious stake in insisting no AI was involved; About Face's denial is the company's own word about its own photo, and nobody has checked it.
The complaint holds up better than the explanation
Two things are solid: the badge exists and is appearing on images whose owners say no generative tool touched them, and one reporter tested it rather than aggregating outrage. Beyond that, we are working from a mechanism no company will describe, anecdotes from anonymous accounts, and a vendor statement filtered through someone else — inside a single publisher with no independent replication. One clarifying sentence from Meta could rewrite the cause while leaving the grievance untouched.