AI Content label errors are spreading across Instagram for the second time in roughly two and a half years, and the pattern is the same as it was in 2024: ordinary photographs are being tagged as synthetic while genuinely synthetic images sail through untouched. The Verge’s Jess Weatherbed reported on 4 September 2026 that users have spent weeks watching Meta’s automatic system attach an “AI Content” badge to pictures they neither generated nor edited with generative tools.
The badge is supposed to be a shortcut to trust. Instead it has become a coin flip. A photographer who removed a speckle from a portrait gets flagged. A brand that shot on an iPhone and touched the image up in Apple Photos gets flagged. Meanwhile a brand-new account uploading fully generated pictures in rapid succession, carrying verifiable machine-readable AI markers, gets nothing at all.
This article walks through what the badge was designed to do, what is actually happening, where the fault lines run between Meta and third-party editing tools, and what a creator or brand should do when the badge lands on work that is entirely their own. If you follow the platform-governance stories that change how ordinary content gets published, they land in our AI models and tools hub as they break. The short version: the signal has stopped carrying information in either direction, and Meta has not explained why.
Table of contents
- What the AI Content Label Was Supposed to Do
- What Is Going Wrong With the AI Content Label Now
- The Canva Trail Behind Many AI Content Label Errors
- When a Real Brand Gets an AI Content Label It Did Not Earn
- The 2024 Rehearsal for the 2026 AI Content Label Mess
- What Signals Could Be Driving the AI Content Label
- The Test That Should Worry Meta Most
- Why the AI Content Label Failure Matters Beyond Instagram
- What Creators and Brands Should Do About the AI Content Label
- Frequently Asked Questions About the AI Content Label
- Conclusion: A Signal That No Longer Carries Information
- References and Further Reading
What the AI Content Label Was Supposed to Do
Meta did not stumble into image labelling. It announced the programme deliberately, with a named executive and a published technical rationale, and the gap between that rationale and today’s behaviour is the whole story.
The promise Meta made in February 2024
On 6 February 2024, Nick Clegg, then Meta’s President of Global Affairs, published a post committing the company to labelling AI-generated images across Facebook, Instagram and Threads. The mechanism was specific. Meta said it would read the “AI generated” information carried in the C2PA and IPTC technical standards, and it named the tool makers whose output it expected to catch: Google, OpenAI, Microsoft, Adobe, Midjourney and Shutterstock.
That is a metadata-driven design, not a pixel-inspection one. The AI Content label was never meant to be a model guessing whether a photograph looks synthetic. It was meant to be a reader of cryptographic provenance records that generative tools embed at the moment of creation.
The distinction that matters: generated versus modified
Meta then made a second commitment that turns out to be central. In April 2024 it published a revised approach to labelling after the first wave of complaints, conceding that “our labels based on these indicators weren’t always aligned with people’s expectations.”
The fix was a split. Content the system detected as fully generated by an AI tool would keep a visible badge. Content it detected as only modified or edited with AI tools would have the “AI info” label moved into the post’s menu, out of the main view. From 1 July 2024, the wording changed from “Made with AI” to “AI info” to reflect that softer claim. The current AI Content label failure breaks precisely that promise, because the badge appearing on retouched photographs is the loud version, not the quiet menu entry.
Why a visible badge is a strong claim
An AI Content label is not a neutral annotation. The AI Content label tells every viewer that what they are looking at was manufactured rather than observed, which for a photographer, a journalist or a cosmetics brand is a statement about honesty. Getting an AI Content label wrong is not a cosmetic bug; it is an accusation attached to someone else’s work, applied automatically and without appeal.
What Is Going Wrong With the AI Content Label Now
The reports gathering on Threads over recent weeks describe two opposite failures happening simultaneously. Either one alone would be a bug. Together they empty the AI Content label of meaning.
Ordinary edits are being tagged
The first failure is the false positive. Photographers using routine tools are finding an AI Content label attached to images that contain no generated content whatsoever. One Threads user reported the badge appearing “every time there is bg remover involved.” Another said it landed after they used Canva to remove a single speckle from a photograph, and the entire image was then flagged on Instagram as AI.
Speckle removal is not image synthesis. Nothing in that operation invents a subject, a background or a scene. The AI Content label in these cases is describing an edit that a darkroom technician could have performed with a spotting brush sixty years ago.
Real AI images are slipping through
The second failure runs the other way, and it is the more consequential one. The Verge ran a controlled test: photographs edited or fully generated with Canva’s Background Remover, Photoshop’s background erasing tool, Adobe’s Firefly generative features, Google’s Nano Banana model inside Gemini, and Apple Intelligence features in iOS. Several of those uploads were confirmed to carry C2PA and SynthID signals. Instagram labelled none of them.
That is the design failing at its own stated task. The metadata Meta said in February 2024 it would read was present, verifiable and ignored. An AI Content label that cannot fire on a signed provenance record is not performing the function it was announced to perform.
The label’s own wording contradicts the evidence
There is a third oddity buried in the tagged posts. The badge text tells users it has detected signals that typically identify content that was entirely AI-generated, rather than content that was only modified with AI. That is the loud AI Content label category from the April 2024 split. Applying it to a photograph that was shot on a phone and lightly retouched is not a borderline judgement call that reasonable systems might differ on. It is the wrong bucket entirely.
| Reported case | What was actually done | AI Content label applied? |
|---|---|---|
| Canva Background Remover edit | Assistive subject selection, no generated pixels | Yes, repeatedly |
| Canva blemish removal | One speckle removed from a photograph | Yes, whole image flagged |
| About Face brand photos | Shot on iPhone, edited in the Photos app | Yes |
| Poisoned image | Anti-training perturbation applied | Yes |
| Unpoisoned version of the same image | Nothing | No |
| Adobe Firefly generative output | Genuine generative AI, C2PA present | No |
| Google Nano Banana output in Gemini | Genuine generative AI, SynthID present | No |
| Meta AI app output | Genuine generative AI, Meta’s own tooling | Yes — the one reliable trigger |
The Canva Trail Behind Many AI Content Label Errors
Not every misfire traces back to Meta. At least one thread of the current AI Content label problem appears to originate upstream, in how a third-party editor described its own tools to the metadata standard.
What Canva told a content strategist
After raising the Background Remover tagging pattern publicly, content strategist Jess Bruno reported that Canva came back with an explanation: some of the platform’s assistive AI tools “were being tagged as generative.” Canva told Bruno that its tools now tag correctly. A note on Canva’s background removal help page states that using the tool does not add Canva’s AI-generated content metadata to a design, although it is unclear whether that note is new, and Canva did not respond to The Verge’s request for comment.
If that account is right, the AI Content label was doing exactly what it was told to do. It read a provenance record that said “generative”, and it acted on it. The error was in the declaration, not the reader.
Why background removal is not generative AI
This distinction is worth holding on to, because it will recur across every editing suite. Background removal, object selection and blemish healing use machine learning, but they use the assistive variety that Photoshop has shipped in selection and removal tools for well over a decade. Nothing is invented. The model decides which existing pixels belong to a subject; it does not hallucinate new ones.
Lumping that class of tool in with text-to-image generation is the category error underneath much of the current confusion, and it is the same error we examined when Pangram’s chief executive argued that detection is harder than “real or fake” in written work. Assisted and generated are different states, and a single badge cannot express both.
The reports that continued after the fix
Canva’s account does not close the case. Some Threads users report that images edited with the Background Remover are still being tagged after the supposed fix. Others report the opposite anomaly: images they edited with the same tool before the fix were never tagged at all, and the tool did not write the C2PA metadata that Meta supposedly reads.
Those two reports cannot both be explained by a single upstream declaration bug. Whatever is producing the AI Content label is either reading something other than C2PA, or reading it inconsistently, or applying additional signals Meta has never described.
When a Real Brand Gets an AI Content Label It Did Not Earn
The clearest illustration of the cost is a commercial one, because a brand cannot quietly ignore a badge that tells its customers the product photography is fake.
The About Face case
Recent photographs posted to Instagram by About Face, the cosmetics company founded by the singer Halsey, were automatically given the AI Content tag. When a commenter asked about it, the brand’s social manager replied that no AI was used: “This photo was taken on my iPhone and then slightly edited by myself in the photo app. Our brand uses real artist[s] and real people to create everything you see across all our platforms.”
For a beauty brand, “we use real people” is not a throwaway line. It is a positioning claim, and an automated badge contradicting it in public does measurable reputational work before anyone at the company sees the notification.
Which iPhone tools actually add generative metadata
Only a short list of Apple Photos features should attach generative AI metadata: Spatial Reframing, Extend, and the updated Clean Up introduced in iOS 27. The first two add genuinely new generated elements; the third removes distracting objects or people. All three run on Apple Intelligence and embed Google’s invisible SynthID watermark.
The SynthID check that came back empty
Here the trail goes cold in a way that should concern Meta. According to Google’s own Gemini-based verification, the tagged About Face images do not carry a SynthID watermark at all. So even if one of those three Apple tools had been used, the marker Meta would need to read is absent, and the badge still says the content was entirely AI-generated. Something other than the published metadata pipeline is driving this AI Content label.
The 2024 Rehearsal for the 2026 AI Content Label Mess
Anyone who followed Instagram photography two years ago has seen this film. The 2024 episode is not a loose analogy; it is the same failure mode with the same root cause and the same corporate response.
“Made with AI” became “AI info”
A few months after the labelling feature launched, photographers found the “Made with AI” badge attached to straightforward professional work. Reporting at the time indicated the detection system was sweeping up pictures whose Adobe metadata recorded generative retouching, even where the change was trivial and the resulting photograph was substantively identical to the original frame.
Meta’s response was to promise a labelling approach that better reflected the amount of AI used in an image, and to rename the AI Content label’s predecessor. The wording softened. The underlying inability to distinguish a generated scene from a corrected blemish did not.
What Meta admitted at the time
Clegg’s original February 2024 post was unusually candid about the limits. He conceded that signals were not yet present at scale in audio and video generators, that “there are ways that people can strip out invisible markers”, and that “it’s not yet possible to identify all AI-generated content.” Those admissions describe a system with known false-negative exposure from day one. What nobody flagged was the false-positive exposure, which is the half now damaging ordinary photographers.
What has and has not changed since
| Dimension | 2024 episode | 2026 episode |
|---|---|---|
| Badge wording | “Made with AI” | “AI Content” |
| Who was hit | Professional photographers | Photographers, creators and brands |
| Suspected trigger | Adobe generative retouch metadata | Assistive tools declared as generative |
| False negatives reported | Acknowledged in advance | Demonstrated in a controlled test |
| Meta’s public explanation | Published, then revised policy | None — no response to enquiries |
| Remedy offered | Renamed badge, menu demotion | No stated remedy |
What Signals Could Be Driving the AI Content Label
The honest answer is that nobody outside Meta knows, and Meta did not respond to a request for clarification. What can be established is which signals exist and how each behaved in testing.
C2PA, IPTC and SynthID explained
Three provenance mechanisms matter here, and they work differently. C2PA, the standard behind Content Credentials, attaches a cryptographically signed manifest describing how a file was made and edited. IPTC photo metadata is the older embedded-field standard that news and stock photography have used for decades. SynthID is an invisible watermark written into the pixels themselves, which survives cropping and recompression in ways a metadata field does not.
| Signal | Where it lives | Survives re-saving? | Acted on in The Verge’s test? |
|---|---|---|---|
| C2PA manifest | Signed metadata block | Often stripped | No |
| IPTC fields | Embedded metadata | Often stripped | No |
| SynthID watermark | In the pixels | Yes, largely | No |
| Meta’s own generation marker | Undisclosed | Undisclosed | Yes |
The poisoned-image anomaly
One reported case does not fit any of these categories. A user found that Meta applied the AI Content tag to an image that had been poisoned — deliberately perturbed by a system designed to make pictures useless as AI training data or to degrade models trained on them — while the unpoisoned version of the same content went untagged.
That is a meaningful clue. Poisoning alters pixel statistics without touching metadata. If a poisoned frame triggers the badge and its clean twin does not, then at least one pixel-level classifier is contributing to the AI Content label, and its behaviour has never been documented publicly.
The one signal that reliably works
Across all of that testing, exactly one input reliably produced an AI Content label: images edited or fully generated using the Meta AI app. That result is coherent with a system that trusts its own first-party marker and struggles with everything else, which is a very different product from the industry-standards reader announced in 2024.
The Test That Should Worry Meta Most
The most damaging finding in The Verge’s reporting is not a mislabelled photograph. It is what happened when a reporter tried, deliberately and clumsily, to look like exactly the thing Instagram says it is now policing.
A brand-new account behaving like an AI farm
The test account was created with minimal profile information. Generated and AI-edited pictures were published in quick succession, several of them confirmed to carry C2PA or SynthID signals. As the reporter put it, it would have been hard to try harder to act like a bot or an AI farm.
Fourteen days, no label, no profile flag
The images were not labelled. The profile was not flagged as AI-generated. At the time of writing the posts had been live and fully public for almost two weeks.
Why this undercuts the crackdown announced days earlier
The timing is awkward. Days before that test concluded, Meta had announced new limits on undisclosed AI profiles, including an AI-generated profile label and reduced reach for accounts that decline to disclose. An enforcement regime that misses a deliberately bot-shaped account for fourteen days, while badging a cosmetics brand’s iPhone photographs, is not calibrated in either direction.
Why the AI Content Label Failure Matters Beyond Instagram
It would be easy to file this as a platform bug. It is more useful to read it as a live demonstration of how provenance systems fail in the wild, because everyone building one is about to meet the same problems.
Trust is asymmetric
A labelling system earns credibility slowly and loses it instantly. Users do not maintain separate confidence estimates for false positives and false negatives; they form one judgement about whether an AI Content label means anything. Once a viewer has seen an AI Content label on a friend’s holiday snap, the badge stops functioning as evidence on any post, including the ones where it is correct.
Reach and the disclosure penalty
There is a structural incentive problem sitting underneath the accuracy problem. Meta has said undisclosed AI profiles will see reduced distribution. That makes an AI Content label an economic object, not just an informational one. If the system attaches it inaccurately, it is levying a distribution penalty on creators who did nothing wrong, and doing so through an automated process with no published appeal route.
The regulatory clock
Transparency obligations for synthetic media are no longer voluntary in every market. Article 50 of the EU AI Act sets transparency duties around AI-generated and manipulated content, and platforms will be expected to show that their disclosure mechanisms work. A badge that fires on retouched photographs and stays silent on signed generative output is a weak compliance artefact, whatever its intentions.
What good detection looks like by comparison
The contrast with dedicated detection vendors is instructive. When we looked at whether Pangram deserves its reputation as the gold standard for AI text detection, the argument turned on published error rates, independent testing and a clear account of what the tool measures. Meta has published none of those things for the AI Content label, and declined to answer questions about what signals it currently scans for.
What Creators and Brands Should Do About the AI Content Label
None of this is fixable from the outside, but the exposure is manageable. The practical goal is to reduce the chance of an unearned badge and to have a response ready if one lands.
Check before you post
Run finished images through a Content Credentials inspector before publishing anything commercially sensitive. If a manifest exists and describes an edit as generative when it was not, you have found your trigger and can change the tool or the export path. Exporting through a step that does not carry the manifest forward is a blunt instrument, but it is available.
Know which tools declare what
Build a short internal list of which editors in your workflow write provenance metadata and what they claim. Canva, Adobe and Apple all behave differently, and the declarations change without notice. For a marketing team publishing daily, this belongs alongside the rest of your content and marketing operations documentation rather than in someone’s head.
If you are wrongly labelled
Do not delete and repost reflexively. Screenshot the AI Content label and the post, note the exact tools used in the edit chain, and reply publicly to the first commenter who asks, exactly as About Face did. A visible, specific correction from the account owner is currently more informative to your audience than an AI Content label is.
If you do use generative AI
Disclose it yourself, in the caption, in your own words. Self-disclosure is the one part of this system that is not broken, it costs nothing, and it removes any question of whether an automated AI Content label caught you or missed you.
| Situation | Action | Why |
|---|---|---|
| Routine retouching only | Inspect Content Credentials before export | Catches a false “generative” declaration early |
| Badge already applied | Screenshot, log the edit chain, reply in thread | Creates a record and answers viewers directly |
| Genuine generative use | Disclose in the caption yourself | Removes reliance on automated detection |
| Brand or client work | Document the tool list and its declarations | Makes the trigger reproducible when it recurs |
| Repeated false badges | Change the tool in the offending step | The declaration, not the edit, is usually at fault |
| High-stakes campaign | Publish a provenance note on your own site | You control the record the platform does not |
Frequently Asked Questions About the AI Content Label
Why is Instagram putting an AI Content label on my real photo?
Most likely because an editing tool in your chain wrote provenance metadata declaring the edit as generative, even though it was assistive. Canva has acknowledged that some of its assistive tools were being tagged that way. Some reported cases have no identifiable trigger at all.
Does removing a background count as generative AI?
No. Background removal, object selection and blemish healing are assistive machine learning of the kind Photoshop has shipped for over a decade. They rearrange or delete existing pixels rather than inventing new ones, which is why an AI Content label on that work is a category error.
Can I remove the AI Content label from a post?
There is no reliable published route to appeal an automatic AI Content label on an individual image. The practical options are to change the tool in the offending edit step and repost, or to leave the post up and explain the discrepancy in the comments yourself.
Has Meta explained what is causing this?
No. Meta did not respond to The Verge’s request for clarification about which signals it currently scans for, or how and when it scans for them. Its last detailed public statement on the mechanism remains the February 2024 announcement.
Does an AI Content label reduce my reach?
Meta has said reduced distribution applies to undisclosed AI profiles under its newer account rules. It has not published a distribution effect for the per-image badge, so treat any reach change as unmeasured rather than confirmed either way.
Should I trust the AI Content label as a viewer?
Not on its own. In current testing the badge missed genuinely generated images carrying industry-standard markers and fired on ordinary photographs, so its absence is not evidence a picture is real and its presence is not evidence a picture is synthetic.
Conclusion: A Signal That No Longer Carries Information
A detection system has two ways to fail, and Instagram is currently failing both at once. It is telling users that authentic photographs are manufactured, and it is telling them nothing at all about images that were genuinely generated and that carry the very markers Meta announced it would read.
The uncomfortable part is that this is a repeat. The company met this failure in 2024, described it publicly, renamed the AI Content label’s predecessor and promised a labelling approach that reflected how much AI was actually involved. Roughly thirty-one months later the same class of error is back, the explanation is absent, and the AI Content label has quietly stopped being a reason to believe anything.
References and Further Reading
The Verge: Instagram’s AI detection is a mess (again)
Meta: Labeling AI-Generated Images on Facebook, Instagram and Threads
Meta: Our Approach to Labeling AI-Generated Content and Manipulated Media
Coalition for Content Provenance and Authenticity
Content Authenticity Initiative
Canva: Background Remover help
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.