Meta, Google, and Microsoft Keep Promising “AI Labels”—So Why Do Most Deepfakes Still Arrive ‘Unmarked’ in 2026?
The standards are real, the cryptography is elegant—and yet labels vanish in upload pipelines, reposts, and UI design choices. In 2026, “AI labeling” fails less at creation than in the supply chain.
By TheMurrow Editorial
March 1, 2026
Key Points
- 1Recognize that most “AI labels” break in transit: upload pipelines, recompression, screenshots, and reposts routinely erase provenance signals.
- 2Distinguish three systems—C2PA provenance, SynthID watermarking, and platform UI notices—because each fails differently and at different points.
- 3Demand consistency: labels must be preserved on upload, surfaced clearly in feeds, and credible enough to avoid false-positive backlash.
Deepfakes have a familiar tell in 2026: not the uncanny eyes or the too-smooth skin, but the empty space where a label was supposed to be.
For years, the largest platforms and AI vendors have promised provenance systems that would travel with synthetic media—digital “nutrition labels” for images, audio, and video. The acronyms are real. The standards exist. Some of the underlying cryptography is elegant. And yet, when manipulated clips land in group chats or ricochet through reposts, most arrive unmarked.
The obvious explanation is “bad actors strip the labels.” That happens. But it is not the full story. Reporting and platform documentation point to a more structural problem: many “AI labels” are optional, inconsistently displayed, hidden in menus, dependent on metadata that often gets erased, and not reliably preserved when content is uploaded and re-shared. The Washington Post’s testing is blunt: even when provenance information exists, common platform pipelines can remove it, blocking inspection of what the clip actually is.
“The defining feature of AI labeling in 2026 is not its sophistication—it’s its fragility.”
— — TheMurrow
What readers are noticing, often without the language to name it, is a paradox: we have more labeling standards than ever, and less labeling in the wild than the public was led to expect.
The Paradox: Why “Provenance Everywhere” Became “Labels Nowhere”
Public messaging around deepfakes has leaned on a comforting idea: the internet will learn to identify synthetic media the way food packaging lists ingredients. In practice, the internet works more like a deli counter. Items are wrapped, rewrapped, sliced, recompressed, screenshot, stitched into compilations, and reposted with new captions and new contexts. The label—if it exists—rarely survives.
The Washington Post reported in late 2025 that major platforms stripped the digital marker from an uploaded AI video, preventing provenance inspection. The test matters because it challenges the assumption that “standards adoption” automatically translates to user-facing clarity. A provenance signal can be technically present at creation and still vanish the moment the file touches a typical upload-and-share workflow.
Even when a label survives, readers often never see it. The Verge observed in February 2026 that Meta’s “AI info” can be tiny, displaced by other interface elements, and inconsistent across surfaces such as mobile versus desktop. Engadget reported in 2024 that Meta began moving certain AI-related notes into a three-dot menu—an area many users never open.
None of this requires conspiracy. Platforms are balancing competing pressures: user experience, legal risk, creator backlash, and the simple reality that no detection method is perfect. Still, the net effect is the same: the public sees fewer labels than promised, in the places that matter most, at the moment decisions are made.
The Washington Post reported in late 2025 that major platforms stripped the digital marker from an uploaded AI video, preventing provenance inspection. The test matters because it challenges the assumption that “standards adoption” automatically translates to user-facing clarity. A provenance signal can be technically present at creation and still vanish the moment the file touches a typical upload-and-share workflow.
Even when a label survives, readers often never see it. The Verge observed in February 2026 that Meta’s “AI info” can be tiny, displaced by other interface elements, and inconsistent across surfaces such as mobile versus desktop. Engadget reported in 2024 that Meta began moving certain AI-related notes into a three-dot menu—an area many users never open.
None of this requires conspiracy. Platforms are balancing competing pressures: user experience, legal risk, creator backlash, and the simple reality that no detection method is perfect. Still, the net effect is the same: the public sees fewer labels than promised, in the places that matter most, at the moment decisions are made.
The uncomfortable truth for users
Labels that require extra taps, specialized viewers, or intact metadata are not “labels” in the everyday sense. They are audits—useful to investigators, unreliable for everyone else.
“A label you have to hunt for is not a warning. It’s a footnote.”
— — TheMurrow
Three Different “AI Labels” People Keep Confusing
Part of the confusion is semantic. “AI labeling” is treated as one thing. It is at least three, each with different strengths and failure modes.
1) Provenance metadata: C2PA / Content Credentials
C2PA’s Content Credentials are a cryptographically signed provenance standard designed to record how a piece of content was made and edited—tool used, edits applied, issuer, and related context. In February 2026, C2PA announced Content Credentials 2.3 and said the coalition now has 6,000+ members/affiliates with “live applications.” Those are meaningful adoption signals.
Provenance has an important virtue: it is not guessing. When the chain is intact, the record is anchored in cryptographic signing. The weakness is equally simple: provenance metadata is only useful when it is present, preserved, and surfaced in a way ordinary users can understand.
If a platform strips metadata on upload—or if a clip is screen-recorded, re-encoded, or passed through editing tools that don’t preserve the chain—the signature can break. In the real social internet, these transformations are not edge cases. They are the default.
Provenance has an important virtue: it is not guessing. When the chain is intact, the record is anchored in cryptographic signing. The weakness is equally simple: provenance metadata is only useful when it is present, preserved, and surfaced in a way ordinary users can understand.
If a platform strips metadata on upload—or if a clip is screen-recorded, re-encoded, or passed through editing tools that don’t preserve the chain—the signature can break. In the real social internet, these transformations are not edge cases. They are the default.
6,000+
C2PA said in Feb. 2026 its coalition has 6,000+ members/affiliates with “live applications,” signaling broad—if fragile—adoption.
2) Invisible watermarking: Google’s SynthID
A different approach avoids visible labels and embedded metadata. SynthID is an imperceptible watermark embedded into content generated by Google’s tools across text, audio, images, and video within its ecosystem. In May 2025, Google announced SynthID Detector, an upload-and-check portal for journalists and researchers, initially rolling out to early testers.
Google also said that over 10 billion pieces of content had been watermarked with SynthID as of May 2025. That is a huge number—evidence that watermarking can operate at platform scale.
The limitation is structural: SynthID is Google-specific. A deepfake made with non-Google tools may carry no SynthID watermark. Even when the watermark exists, the public still needs access to detection infrastructure. A detector available to “early testers” is not the same as a label visible at the point of consumption.
Google also said that over 10 billion pieces of content had been watermarked with SynthID as of May 2025. That is a huge number—evidence that watermarking can operate at platform scale.
The limitation is structural: SynthID is Google-specific. A deepfake made with non-Google tools may carry no SynthID watermark. Even when the watermark exists, the public still needs access to detection infrastructure. A detector available to “early testers” is not the same as a label visible at the point of consumption.
10B+
Google said in May 2025 that over 10 billion pieces of content had been watermarked with SynthID—massive scale, but limited ecosystem coverage.
3) Platform UI labels: what users actually see
The third system is the most socially consequential: the notice a user encounters in a feed. “Made by AI.” “AI info.” “Altered or synthetic.” These UI labels may be driven by several inputs:
- Creator self-disclosure workflows (for example, a disclosure in YouTube Studio)
- Metadata/provenance reads (C2PA-style signals)
- Platform classifiers (still limited at scale, especially for audio/video)
- Separate enforcement rules for ads or political content
UI labels can be powerful. They can also be quietly de-emphasized, inconsistently applied, or designed for plausible deniability rather than clarity.
- Creator self-disclosure workflows (for example, a disclosure in YouTube Studio)
- Metadata/provenance reads (C2PA-style signals)
- Platform classifiers (still limited at scale, especially for audio/video)
- Separate enforcement rules for ads or political content
UI labels can be powerful. They can also be quietly de-emphasized, inconsistently applied, or designed for plausible deniability rather than clarity.
Meta’s “AI Info” Case Study: When Labels Get Smaller, Not Stronger
Meta’s label evolution is a revealing case study because it shows how quickly “labeling” becomes a user-interface negotiation rather than a technical one.
Meta originally used “Made with AI,” then changed wording to “AI info” after complaints that real photos were being mislabeled. Photographers argued that minimal edits—especially using AI-assisted features in tools like Photoshop—were triggering labels that suggested something more dramatic than what occurred. Engadget reported Meta’s shift as a response to that backlash, acknowledging mismatches between labeling and user expectations.
A second move was less about language and more about placement. Engadget reported in September 2024 that Meta would hide the AI note for images deemed merely “modified” by AI, placing it in the post’s three-dot menu. That is not a neutral design choice. Three-dot menus exist to reduce on-screen clutter. They are also where platforms put information they consider secondary.
The Verge’s February 2026 reporting sharpened the critique: the “AI info” label can be tiny, sometimes pushed aside by other interface elements, requiring extra taps, and not reliably present across surfaces.
Meta originally used “Made with AI,” then changed wording to “AI info” after complaints that real photos were being mislabeled. Photographers argued that minimal edits—especially using AI-assisted features in tools like Photoshop—were triggering labels that suggested something more dramatic than what occurred. Engadget reported Meta’s shift as a response to that backlash, acknowledging mismatches between labeling and user expectations.
A second move was less about language and more about placement. Engadget reported in September 2024 that Meta would hide the AI note for images deemed merely “modified” by AI, placing it in the post’s three-dot menu. That is not a neutral design choice. Three-dot menus exist to reduce on-screen clutter. They are also where platforms put information they consider secondary.
The Verge’s February 2026 reporting sharpened the critique: the “AI info” label can be tiny, sometimes pushed aside by other interface elements, requiring extra taps, and not reliably present across surfaces.
Why Meta would want to be cautious
False positives are not an abstract concern. Over-labeling creates a “cry wolf” dynamic: once users see labels on content they believe is authentic, labels stop meaning “be careful” and start meaning “the platform is covering itself.”
Platforms also face political pressure. Overzealous labeling can look like editorializing, especially when applied unevenly. Under-labeling, meanwhile, can look like negligence. In that bind, many companies gravitate toward less obtrusive UI, narrower triggers, and heavy reliance on “industry standard indicators” they can point to if challenged.
Platforms also face political pressure. Overzealous labeling can look like editorializing, especially when applied unevenly. Under-labeling, meanwhile, can look like negligence. In that bind, many companies gravitate toward less obtrusive UI, narrower triggers, and heavy reliance on “industry standard indicators” they can point to if challenged.
“The hardest part of AI labeling isn’t detecting fakes. It’s choosing who gets blamed when the label is wrong.”
— — TheMurrow
The Metadata Problem: The Internet Eats Context for Breakfast
Provenance systems assume something the social web rarely grants: stable files.
C2PA-style credentials depend on data traveling with the media. But the modern distribution chain is hostile to attached context. Screenshots are a cultural norm. Short-form video is routinely re-encoded. Messaging apps compress. Platforms optimize formats. Even legitimate creators export and re-export across tools.
The Washington Post’s finding—that major platforms stripped a digital marker from an uploaded AI video—shows how this fragility can appear even without malicious intent. A platform may remove metadata for performance reasons, privacy considerations, or to standardize processing pipelines. The impact on provenance is collateral damage, but it is still damage.
C2PA-style credentials depend on data traveling with the media. But the modern distribution chain is hostile to attached context. Screenshots are a cultural norm. Short-form video is routinely re-encoded. Messaging apps compress. Platforms optimize formats. Even legitimate creators export and re-export across tools.
The Washington Post’s finding—that major platforms stripped a digital marker from an uploaded AI video—shows how this fragility can appear even without malicious intent. A platform may remove metadata for performance reasons, privacy considerations, or to standardize processing pipelines. The impact on provenance is collateral damage, but it is still damage.
What this means for ordinary users
When provenance fails, users are left with a handful of imperfect signals:
- The platform’s visible UI label (if present)
- The uploader’s disclosure (if honest)
- Visual and auditory intuition (often unreliable)
- Third-party verification workflows (time-consuming and unevenly available)
The gap between “we have a standard” and “you can trust what you see” sits right here. Standards help the ecosystem coordinate. They do not magically force every intermediary to preserve, display, and prioritize the information.
- The platform’s visible UI label (if present)
- The uploader’s disclosure (if honest)
- Visual and auditory intuition (often unreliable)
- Third-party verification workflows (time-consuming and unevenly available)
The gap between “we have a standard” and “you can trust what you see” sits right here. Standards help the ecosystem coordinate. They do not magically force every intermediary to preserve, display, and prioritize the information.
Watermarks at Scale: Impressive Numbers, Narrow Coverage
Google’s SynthID highlights both the promise and the limits of watermarking.
A watermark is appealing because it can survive some transformations that kill metadata. It is also appealing because it can be applied automatically. Google’s claim of 10 billion+ watermarked items (as of May 2025) suggests the approach can run continuously, without requiring creators to opt in or platforms to read metadata perfectly.
Yet watermarking introduces a different kind of fragmentation. SynthID can only reliably identify content generated within Google’s ecosystem, because it is Google’s watermark. For the broader universe of generative tools, the absence of a SynthID signal does not mean “human-made.” It may simply mean “made elsewhere.”
Google’s SynthID Detector, announced May 20, 2025, is also a telling detail. The company framed it as a portal for journalists and researchers, initially rolling out to early testers. That is valuable, but it reinforces a two-tier reality: professionals may get tools to check, while the average user is still scanning a feed with little more than intuition and whatever UI labels a platform chooses to show.
A watermark is appealing because it can survive some transformations that kill metadata. It is also appealing because it can be applied automatically. Google’s claim of 10 billion+ watermarked items (as of May 2025) suggests the approach can run continuously, without requiring creators to opt in or platforms to read metadata perfectly.
Yet watermarking introduces a different kind of fragmentation. SynthID can only reliably identify content generated within Google’s ecosystem, because it is Google’s watermark. For the broader universe of generative tools, the absence of a SynthID signal does not mean “human-made.” It may simply mean “made elsewhere.”
Google’s SynthID Detector, announced May 20, 2025, is also a telling detail. The company framed it as a portal for journalists and researchers, initially rolling out to early testers. That is valuable, but it reinforces a two-tier reality: professionals may get tools to check, while the average user is still scanning a feed with little more than intuition and whatever UI labels a platform chooses to show.
The risk of misunderstanding
Watermarking can also create a false sense of certainty. A detected watermark might confirm “generated with Google tools,” but it does not tell you whether a clip is deceptive, whether it’s been edited further, or whether it’s being presented out of context. Conversely, no watermark is not proof of authenticity.
Optional, Inconsistent, Hidden: The UI Is Where Labeling Lives or Dies
Even if provenance standards mature and watermarking expands, labeling still faces a basic, under-discussed question: where does a user see the information?
Meta’s experience shows that platforms can comply with a notion of “labeling” while making it easy to miss. The Verge’s reporting about small, inconsistent labels is not just a design critique; it is an argument about incentives. Platforms optimize for engagement, clarity, and minimal friction. Prominent warnings create friction.
At the same time, platforms are understandably wary of making strong claims. Labeling something “AI-generated” is a statement that can be wrong. Even seemingly straightforward categories—“altered,” “synthetic,” “AI-assisted”—collapse under real-world complexity. A photo might be real but retouched. A video might be authentic but have AI noise reduction. A clip might be synthetic but harmless, like a satire filter.
Meta’s experience shows that platforms can comply with a notion of “labeling” while making it easy to miss. The Verge’s reporting about small, inconsistent labels is not just a design critique; it is an argument about incentives. Platforms optimize for engagement, clarity, and minimal friction. Prominent warnings create friction.
At the same time, platforms are understandably wary of making strong claims. Labeling something “AI-generated” is a statement that can be wrong. Even seemingly straightforward categories—“altered,” “synthetic,” “AI-assisted”—collapse under real-world complexity. A photo might be real but retouched. A video might be authentic but have AI noise reduction. A clip might be synthetic but harmless, like a satire filter.
Practical implications for readers
Users should treat labels as one signal among several, not as a definitive verdict. A missing label is especially meaningless in 2026, because the absence may reflect:
- Metadata stripped during upload (as the Washington Post testing suggested)
- Content created with tools that don’t embed a given watermark
- Reposts, screen recordings, or compilations that sever provenance
- A platform UI that hides or inconsistently displays the note
The public conversation often treats the label as a stamp. In practice, a label is closer to a brittle handshake between many systems—any one of which can fail.
- Metadata stripped during upload (as the Washington Post testing suggested)
- Content created with tools that don’t embed a given watermark
- Reposts, screen recordings, or compilations that sever provenance
- A platform UI that hides or inconsistently displays the note
The public conversation often treats the label as a stamp. In practice, a label is closer to a brittle handshake between many systems—any one of which can fail.
Key Insight
In 2026, missing labels usually indicate broken handoffs—upload pipelines, recompression, screenshots, reposts, and inconsistent UI—not “no AI involved.”
What “Better” Looks Like: Accountability Without Crying Wolf
The strongest argument for more visible labeling is democratic hygiene. Synthetic media can be used for harassment, fraud, and political manipulation. Readers want a simple cue. Platforms and vendors have a responsibility to reduce confusion, not amplify it.
The strongest argument against aggressive labeling is credibility. Meta’s early mislabeling controversy—where photographers complained that minimal edits triggered warnings—shows how easily trust erodes when labels feel detached from common sense. Users who are told authentic work is “AI” will stop trusting the system, and bad actors will benefit from the cynicism.
A workable middle path has to acknowledge both realities:
- Provenance standards like C2PA Content Credentials can create a durable record—when the chain is preserved and surfaced.
- Watermarking systems like SynthID can operate at enormous scale—within their ecosystems and with appropriate detection access.
- UI labels are where users live—but labels must be legible, consistent, and honest about uncertainty.
The test for 2026 is not whether a standard exists. The test is whether an ordinary user—on mobile, in a hurry, in a reposted clip—can get a clear, reliable signal without needing to become a forensic analyst.
The strongest argument against aggressive labeling is credibility. Meta’s early mislabeling controversy—where photographers complained that minimal edits triggered warnings—shows how easily trust erodes when labels feel detached from common sense. Users who are told authentic work is “AI” will stop trusting the system, and bad actors will benefit from the cynicism.
A workable middle path has to acknowledge both realities:
- Provenance standards like C2PA Content Credentials can create a durable record—when the chain is preserved and surfaced.
- Watermarking systems like SynthID can operate at enormous scale—within their ecosystems and with appropriate detection access.
- UI labels are where users live—but labels must be legible, consistent, and honest about uncertainty.
The test for 2026 is not whether a standard exists. The test is whether an ordinary user—on mobile, in a hurry, in a reposted clip—can get a clear, reliable signal without needing to become a forensic analyst.
Practical takeaways: how to read AI labels right now
- ✓Treat “AI info” as a prompt, not a verdict. It tells you something about how the platform interpreted signals, not the whole history of the file.
- ✓Assume reposts are label-erasing machines. Screenshots, re-encodes, and compilations often break provenance.
- ✓Remember that “no label” is not “no AI.” Missing UI tags can reflect stripped metadata or tool fragmentation, not authenticity.
- ✓Look for consistency across sources. If a clip appears only as a repost with no original upload, skepticism is rational.
Conclusion: The Label Isn’t Failing—The Supply Chain Is
The story of AI labeling in 2026 is not a story of absent technology. It is a story of brittle handoffs.
C2PA’s Content Credentials are advancing—version 2.3 landed in February 2026, backed by a coalition that says it has 6,000+ members/affiliates. Google’s SynthID shows what scale looks like—10 billion+ watermarked items by May 2025, plus a detector aimed at journalists and researchers. Meta’s “AI info” label shows what happens when the messiness of real creative workflows meets public expectations and platform incentives: wording changes, menu hiding, tiny UI, uneven visibility.
Deepfakes flourish in the gaps between systems: between creation and upload, between upload and repost, between metadata and interface, between what a standard can store and what a platform chooses to show. The public has been promised labels as guardrails. Instead, labels behave more like confetti—present at launch events, missing on the street.
The next phase of this debate will not be won by another acronym. It will be won by mundane decisions: whether platforms preserve provenance by default, whether they display it consistently, and whether they can do so without punishing legitimate creators with false alarms. Trust will not come from perfect detection. Trust will come from honest, visible signals that survive the journey.
C2PA’s Content Credentials are advancing—version 2.3 landed in February 2026, backed by a coalition that says it has 6,000+ members/affiliates. Google’s SynthID shows what scale looks like—10 billion+ watermarked items by May 2025, plus a detector aimed at journalists and researchers. Meta’s “AI info” label shows what happens when the messiness of real creative workflows meets public expectations and platform incentives: wording changes, menu hiding, tiny UI, uneven visibility.
Deepfakes flourish in the gaps between systems: between creation and upload, between upload and repost, between metadata and interface, between what a standard can store and what a platform chooses to show. The public has been promised labels as guardrails. Instead, labels behave more like confetti—present at launch events, missing on the street.
The next phase of this debate will not be won by another acronym. It will be won by mundane decisions: whether platforms preserve provenance by default, whether they display it consistently, and whether they can do so without punishing legitimate creators with false alarms. Trust will not come from perfect detection. Trust will come from honest, visible signals that survive the journey.
2.3
C2PA’s Content Credentials version 2.3 landed in Feb. 2026—progress in standards, but still dependent on preservation and display.
May 2025
Google announced SynthID Detector on May 20, 2025—useful for journalists and researchers, but not a universal, user-facing label.
T
About the Author
TheMurrow Editorial is a writer for TheMurrow covering technology.
Frequently Asked Questions
Why do so many deepfakes still have no visible label?
Many labels depend on metadata or provenance information that can be lost when content is re-uploaded, compressed, re-encoded, or screen-recorded. Reporting has also found cases where platform upload pipelines strip digital markers, preventing provenance inspection. Even when a platform has a label, it may be small, inconsistent, or hidden in menus.
What is C2PA / Content Credentials, in plain English?
C2PA Content Credentials are cryptographically signed provenance records that can travel with digital media, describing how something was made or edited (tools used, edits, issuer). C2PA announced Content Credentials 2.3 on Feb. 9, 2026, and says it has 6,000+ members/affiliates. The system works best when the metadata survives sharing and platforms display it clearly.
What is Google SynthID, and what does “10 billion watermarked” mean?
SynthID is an imperceptible watermark Google embeds in content generated by its tools across text, audio, images, and video. Google said in May 2025 that over 10 billion pieces of content had been watermarked. The number signals scale inside Google’s ecosystem, but it does not imply coverage of content made with other AI tools.
If a post doesn’t have a SynthID watermark, does that mean it’s real?
No. SynthID is Google-specific. Content made using non-Google tools may have no SynthID watermark at all. Also, watermarks and provenance signals can be lost or become harder to detect after certain transformations. “No watermark detected” should be read as “no watermark detected,” not “definitely authentic.”
Why did Meta change “Made with AI” to “AI info”?
Meta changed wording after complaints that real photos were being mislabeled, particularly when minimal AI-assisted edits triggered “industry standard indicators.” Engadget reported the shift as a response to photographers’ concerns. The change reflects a broader tension: labels that are too aggressive can create backlash and weaken trust.
Where did Meta’s AI labels go on some posts?
Engadget reported in September 2024 that Meta would hide the AI note for images deemed “modified” by AI by moving it into the post’s three-dot menu. The Verge later described Meta’s “AI info” label as sometimes tiny and inconsistent across surfaces. These choices can make labels easy to miss in normal scrolling behavior.
