The news has been full of incredible reports recently. Like this one:

BREAKING: Iowa Farmers Discover "Glitter Milk" from Unicorn Cows
DES MOINES, IA - May 25, 2026

A handful of Iowa dairy farmers say they've started milking unicorn cows, and the results have local nutritionists scratching their heads.

The milk sparkles.

"It's real pretty in the morning sun," said Polk County dairy farmer Dale Hutchins. "First time I saw one with a horn, I figured I'd accidentally bought somebody else's livestock. Then it started making glitter milk."

Researchers examining the milk say the shimmering particles appear to be naturally occurring protein crystals rather than actual glitter. Preliminary tests found the milk to be perfectly safe, with unusually high levels of vitamins and minerals. One eight-ounce serving reportedly contains an entire day's recommended vitamins A, C, D, E, B1, B2, B3, and B12.

"The numbers keep coming back looking impossible," said one nutritional biochemist involved in the testing. "Either we've discovered something genuinely remarkable, or one of our graduate students has been replacing the samples with breakfast cereal."

Children participating in a small nutrition study reportedly loved the milk, although several parents complained that the spilled cereal was "way harder to clean because the glitter goes everywhere."

Federal regulators have not commented, and the Iowa Department of Agriculture says it's waiting for additional testing before making any official statements.

If production continues, glitter milk could begin appearing in a few Midwestern co-ops later this year for about $8.99 a half-gallon.

- Staff Reporter, Heartland Agricultural Digest

As proof of this incredible story, we have a photo of a farmer milking a unicorn cow!



According to the metadata:

• The photo is from a Google Pixel 10 Pro.


• The picture has cryptographically signed C2PA metadata. This data says it is "Created by Pixel Camera". The C2PA metadata even includes a 1024x768 preview image of the photo. Everything in the cryptographically signed manifest is consistent with a real photo from a Google Pixel camera.


• In my blogs, I have repeatedly detailed ways to create "authenticated forgeries" using C2PA. However, the one thing I cannot forge is the cryptographic signature itself. This picture has a valid X.509 certificate chain that traces back to the C2PA-managed trust list. The certificate is issued by Google for the Pixel cameras. To my knowledge, nobody can forge this signature; this was really signed by a Google Pixel camera.


• The cryptographic signature includes a signed timestamp. The timestamp is dated "2026-05-25 17:04:19 GMT" and the signer is Google. Again, I cannot forge this signed timestamp; this is real.


• The C2PA organization has a list of conforming products. If we upload this glitter-milk picture to Adobe's Inspect service (a conforming product), it reports that this is a legitimate photo from a Pixel Camera, recorded on May 25, 2026.


• The Adobe-run Content Authenticity Initiative (CAI) provides C2PA implementations. Their CAI Verify validator reports that the contents shows "captured media", came from Google LLC, issued by a Pixel Camera with a notation that it is "Conformant" (a conforming product), and includes a verified timestamp of "May 25, 2026 at 11:04 AM MDT". (They show the time relative to your own time zone, and I'm in MDT.)

Everything says that this is a legitimate photo from a Google Pixel camera.

There's just one problem: It's a forgery. The picture is AI generated and the news article is fiction, but Google's signatures are real.

Industry best practices for responsible disclosure suggest giving vendors 45-90 days to respond. Since we are 90 days past the vendor notification, I'm following industry best practices and making the details public.

Early Reporting History

I've been working closely with a group of researchers at the University of Maryland, Baltimore County (UMBC). They have a Provenance and Authenticity Standards Assessment Working Group (PASAWG) that has been formally evaluating solutions like C2PA. (While I'm a regular attendee, I'm there as a guest and resource, not a member.) One of the things I like about PASAWG is that they have a more formal way to report bugs than my typical "shouting into the blogosphere".

Nearly a year ago (September 2025), Google made a big announcement about the Pixel 10 product line. They explained "How Pixel and Android are bringing a new level of trust to your images with C2PA Content Credentials". Their bullet points (with their bold emphasis):

• The Pixel 10 lineup is the first to have Content Credentials built in across every photo created by Pixel Camera.


• The Pixel Camera app achieved Assurance Level 2, the highest security rating currently defined by the C2PA Conformance Program. Assurance Level 2 for a mobile app is currently only possible on the Android platform.


• A private-by-design approach to C2PA certificate management, where no image or group of images can be related to one another or the person who created them.


• Pixel 10 phones support on-device trusted time-stamps, which ensures images captured with your native camera app can be trusted after the certificate expires, even if they were captured when your device was offline.

As Carl Sagan said, "Extraordinary claims require extraordinary evidence." So we began to take a closer look.

Two months later (November 2025), PASAWG, one of my coworkers (Shawn), and I reported to representatives from Google and C2PA about a potential problem with Google's Pixel camera. In particular, we theorized that someone with root on the device could sign any picture as if it were from the camera. While the C2PA representative listened to the concerns, the Google representative was adamant that this type of attack was not possible. In particular, the signing keys are stored in a secure chip and cannot be extracted, and the Android architecture prevents unauthorized applications from accessing the keys.

More Researchers

Unrelated to our research and reporting, I had been contacted by other researchers who thought that they found the same theoretical flaw. One in Canada, one in the US, and one in the UK; this shows that other people are thinking the same way. (And just because I didn't list any state-sponsored threat actors doesn't mean they are not also evaluating this vulnerability.)

Three months ago (May 2026), a researcher named retr0id (David Buchanan) contacted me. He took the exploit from theoretical to implementation. He sent me two sample pictures that were signed using a Google Pixel device. To say I was impressed is an understatement. But I wanted hard proof that he had implemented it. I sent him a challenge:

1. Using ChatGPT, I generated the source picture of a unicorn cow being milked.


2. ChatGPT's picture was a PNG with an embedded C2PA manifest. I stripped out the manifest and re-encoded the picture as a JPEG.


3. I found a different picture from a Pixel 10 and copied over the metadata. This way, the forgery had all of the correct metadata fields for that camera. I intentionally left the EXIF date wrong (dated 2025-08-29 02:10:17 GMT) and set the EXIF camera model name to "Pixel 10 Pro Totally Legit".


4. I sent my forgery to retr0id.


5. Two minutes later (not kidding), retr0id sent the signed forgery back to me. That two minutes includes receiving the image from me, transferring it to the Google Pixel for signing, signing it, and zipping it up to send back to me. Just the data transfers probably took him a minute and a half. This means that it's mostly an automated exploit. (He's released some of his tools on GitHub and a technical write-up on his blog.)

The example demonstrates how someone with a Google Pixel device could sign any picture (real, fake, AI generated, etc.) as if it came from the Google Pixel's camera. Moreover, the forgery (excluding my intentional artifacts) is indistinguishable from a real photo. C2PA's metadata provides no reliable assurance of provenance or authenticity.

The Vulnerability

I'm going to be intentionally vague here because I don't want to enable bad actors. However, the vulnerability isn't very deep and anyone who can get past the first step is almost certainly able to exploit it.

When I asked retr0id how he did it, he sent me back a wonderful picture that explains the process:



The C2PA signing keys are in a subsystem called 'StrongBox'. This is a secure storage area for handling the keys. The keys go in and never come out. You need a special program in the Trusted Execution Environment (TEE) to access the keys. This special program sends data to be signed by the keys and receives the signature.

The exploit:

Step 1: Get root on the device.
This is the hardest part. The Android operating system is intentionally locked down, so it's hard to get root access.

A common attack for Android devices replaces the bootloader. However, replacing the bootloader requires a factory reset, so you cannot access any secrets or protected data that existed prior to unlocking. To implement the exploit, retr0id needed root access without a reset.

As a hardware specialist, retr0id used a well-known chip-based approach to get a root shell. His implementation was hardware-specific, but the underlying methodology has been around for at least a decade. Moreover, preventing this attack vector requires completely redesigning the hardware architecture.

However, we are not limited to a hardware exploit. During the 90-day responsible disclosure waiting period, two other software-only exploits came out that also granted root access. (Exploit #1 and Exploit #2.) It doesn't matter that these software exploits have been patched; a malicious attacker won't patch their system and can gain root access on their own device. (As far as I can tell, you can still take signed photos, even if the device hasn't been patched recently.)

Regardless of your method, you just need to get root on the device.

Step 2: Sign your data
Find the program that signs the C2PA metadata using the protected keys. Use the program to sign anything. This is a Confused Deputy attack. When using Android's secured environment, only the TEE program can submit data to be signed, but the root user can provide any data to the signing program. Fixing this part of the problem requires redesigning the entire Android security model. In other words, there is no easy patch.

If you have root on the Pixel device (and you didn't change the bootloader), then you can sign any file as if it came from the Pixel camera. The signature will be legitimately signed by Google.

As an aside: For most exploits, saying "start with root" means that additional exploits add nothing. If you have root, then you already control everything. I.e., creating more backdoors is trivial if you can already alter every file. However, with Google and C2PA, we're not using root to stay on the device; we're using it to create authoritative files. Those files will leave the device as the forgeries are disseminated. With this attack vector, gaining root is just the beginning.

Reporting Timeline

I currently have over 40 blog entries about C2PA problems, and most of them disclose distinct vulnerabilities. While the public didn't know most of these problems until I made them public, none of the vulnerabilities have been new to C2PA members.

For this Pixel vulnerability, we recorded the reporting history:

1. We reported it, via email and verbally, to both Google and C2PA representatives. The reporting included details and the demonstration picture. Following best practices for responsible disclosure, we gave them 90 days to respond. (Today, Aug 25, is 90 days from the initial vendor reporting, and about 9 months since the theoretical vulnerability was disclosed.)
   
   
   • We reported it to Google because the exploit is explicitly demonstrated against Google's flagship product, the Pixel series of Android devices.
   
   
   • We reported it to C2PA because the Pixel 10 was the first "Level 2" conforming product. Assurance Level 2 means that it must protect the signing keys. However, while the keys are protected from extraction by the Android StrongBox, this exploit shows that the keys can still be used to sign anything. In effect, the keys are unprotected. So either Google is not Level 2 conforming (false advertising), or they are Level 2 on paper but not in the implementation (deceptive practices), or Level 2 is grossly insufficient for providing any kind of assurance (misleading). In any case, this is definitely a C2PA conformance program problem.


2. I made it clear that I planned to blog about this problem. But I also offered to work with them on the release cycle. For example, if they were about to provide a patch, then I would be willing to delay the blog and coordinate a release. Both Google and C2PA repeatedly acknowledged my offer during the 90-day period. However, I received no feedback from either organization.


3. Google has a bounty program that pays researchers for finding vulnerabilities. I never signed up because Google requires agreeing to legal terms. (Even if I conceptually agree to the reasons behind their terms, I cannot sign anything that could be construed as a legal agreement. I just want to report a bug.) However, retr0id doesn't have those same limitations. Since he implemented it, we (PASAWG, myself, and Google) asked him to submit it through Google's Vulnerability Reward Program (VRP). He did.


4. Google's VRP almost immediately sent retr0id two emails. The first said that the vulnerability was out of scope. The second said to ignore the first email and that it was in scope. They did end up logging it as a received report.


5. Fast forward two months. Retr0id received an email from Google's VRP. (I am including it here with his permission.)
   

   jo...@google.com #9 Jul 14, 2026 12:36AM
   
   Status: Won't Fix (Infeasible).
   
   Hello,
   
   The Android Security Team has conducted an initial severity assessment on this report. Based on our published severity assessment matrix (1) it was rated as not being a security vulnerability that would meet the severity bar for inclusion in an Android security bulletin. If you have additional information that you believe we should use to reassess this report, please let us know.
   
   Please note that notwithstanding our severity rating and the closure of this external bug, we may nonetheless pass this issue on to the feature team for remediation. Therefore, please know that we appreciate this submission and any future contributions.
   
   The Resolution Notes label has been set to NSBC (Not Security Bulletin Class) to reflect this assessment.
   
   Thank you,
   Android Security Team.
   (1) Severity Matrix: https://source.android.com/security/overview/updates-resources#severity
   
   How did we do? Please fill out a short anonymous survey.

   They closed it out as a "Won't Fix (Infeasible)". Google defines "Won't Fix (Infeasible)" as "The changes that are needed to address the issue are not reasonably possible."
   
   More importantly, Google labeled it as "NSBC (Not Security Bulletin Class)". This code means that it either isn't a security vulnerability or isn't considered severe. In effect, Google explicitly said that a vulnerability in Google's flagship Pixel product line, which permits anyone to sign anything as if it legitimately came from the camera, is not a significant security vulnerability. I disagree with Google: verifiable history (provenance), reliable source attribution, and secure key management are explicitly security concerns. (See NIST SP 800-193 Platform Firmware Resiliency Guidelines, NIST SP 800-57 Recommendation for Key Management, and NIST SP 800-53 Rev. 5 Security and Privacy Controls for Information Systems and Organizations.) This demonstrates a fundamental disconnect between how Google views "OS platform boundaries" and "content provenance integrity".

It took a while, but retr0id did receive payment for reporting this bug to Google. VRP bounties are only for security issues. By paying the bounty, Google implicitly confirms that this bug is a security problem, even though it was classified as NSBC and kept out of the security bulletin. The NSBC classification also means no CVE was assigned, which keeps the issue out of regulatory tracking, enterprise compliance audits, and the National Vulnerability Database (NVD).

We have done our due diligence for reporting this problem. Google has decided to downplay the vulnerability, claiming that it isn't a noteworthy security issue. In contrast, C2PA did not respond at all.

Revoking Certificates

Within days of demonstrating the bug and sharing the sample image, Google revoked the X.509 signing certificate used for the glitter-milk picture. That sounds like responsible incident response on the surface, but in practice, it reveals a fundamental flaw in how Content Credentials interact with public key infrastructure (PKI). Keep in mind, they quickly revoked the certificate (a security response), despite Google's formal response weeks later saying that it was not significant enough for a security bulletin.

There are two major problems with relying on revocation to fix forged media:

1. Validators Don't Check Revocation
   The current C2PA specification does not require validators to perform revocation checks. As of this writing, I am unaware of any conforming validator products that check whether a manifest's certificate has been revoked. So even though Google revoked the certificate for the glitter-milk photo, most C2PA validation tools will still happily report the image as authentic.


2. The Privacy Paradox: Unique Signing Certificates
   To prevent third parties from tracking users across photos, Google designed their C2PA implementation to issue an ephemeral, unique signing certificate for every single photo.
   
   The trust chain looks like this:
   
   
   • Root CA: Google's root certificate sits on the C2PA-managed trust list.
   
   
   • Intermediate Certificate: Google's root issues an intermediate certificate. As far as I can tell, every Pixel device uses the same set of intermediate certificates.
   
   
   • Leaf Certificate: The intermediate cert issues a brand-new, single-use leaf certificate that is used to sign an individual image capture.
   
   Because every picture gets its own unique signing certificate, revoking the glitter-milk certificate only invalidated that one specific photo. This does not prevent retr0id (or anyone else with this exploit) from generating millions of additional forged images on that exact same compromised Pixel.

This signing approach, with unique signatures per picture, introduces serious problems:

• Ineffective Revocation: Google can only revoke certificates for forgeries that are actively discovered and reported to them. Unreported forgeries remain 100% valid.


• Denial of Service: An attacker running an automated batch script could sign millions of synthetic images. Reporting all of these intentional forgeries would likely swamp Google's certificate revocation infrastructure.
  
  (At the technical level: this is an attack against the ingest pipeline and OCSP signer; C2PA does not support CRLs for revocation. Google currently lacks a portal or documented process for users to submit individual forged photos for revocation. If Google were to build a portal without rate-limiting, it risks becoming a bandwidth/DoS problem on its own. If they add CAPTCHA or other throttling to protect the ingest pipeline, then known forgeries may not be submitted in a reasonable time, and humans could become discouraged. Moreover, bulk OCSP revocation could plausibly strain cryptographic signing throughput.)


• Verification Problem: When a user submits a picture to Google for revocation, how does Google know that it really is a forgery? With the glitter-milk example, we explicitly showed them how it was made. However, a malicious person could submit legitimate photos and claim they are forgeries. Google needs some way to identify whether a signed picture from a Pixel device is actually from the camera. This remains a hard problem. Depending on their implementation, Google could reject real forgery reports if the verification process is too strict or revoke legitimate photos if it's too lenient.
  
  
  • Without C2PA: Individual analysts must evaluate the media using whatever tools they have available.
  
  
  • With Google's C2PA signature: When someone submits a picture for revocation, the onus is on Google to provide the verification. (I suspect that nobody asked Google's legal department about whether the company wanted to be put in the position of validating all pictures.) Keep in mind: the entire premise of C2PA is that Google cannot otherwise verify whether a picture is authentic, so asking Google to verify whether a revocation request's media is real just restates the same unsolved problem.


• Painted Into a Corner: With the current architecture, Google cannot revoke the device's intermediate certificate without instantly invalidating every authentic, legitimate Pixel photo ever taken. Google's revocation approach effectively becomes all or nothing. In either case, they cannot stop one individual from creating signed forgeries.

Since the core exploit impacts Android's StrongBox and TEE architecture, revoking individual certificates does not resolve this problem. Revoking a certificate, only to have an attacker compromise the replacement certificate in the exact same way, is not an effective security solution.

By choosing privacy through single-use certificates, and without addressing local key abuse, Google created a system where revoking a compromised image is nothing more than security theater.

Real-World Problems

It is easy to treat "Glitter Milk" as an amusing and harmless proof-of-concept. But the implications of a broken content provenance model are anything but funny.



Image provenance is critical for determining whether the media represents something real or fake. Whether it's a political proof-of-life, images of war or strife, or even something less extreme, like an insurance claim, there are direct consequences from forged provenance.

• This Mitch McConnell picture has no camera-original metadata, but does include an XMP record showing that it was altered with an Adobe application hours before being released to the public. If someone replaced the metadata with fake Google Pixel information, and then had it signed by a real Google Pixel device, would it be more trustworthy?


• The second picture is from an artist who creates AI-generated pictures of life in Russia. If we removed the Facebook re-encoding artifacts and had it signed by a Google Pixel device, would you think it was authentic?


• The third picture is part of a product defect claim. Unlike the first two, this one isn't hypothetical: it carries a cryptographically-valid C2PA signature that genuinely came from a Google Pixel device. But now that we've shown that same "came from a camera" signature can be applied to non-camera media, should you trust it?

It's hard enough to debunk one false picture. However, with a little effort, a malicious actor could add in fake camera metadata and have it authoritatively signed by a trusted device. That significantly increases the effort to debunk a picture since it has the backing of Google's cryptographic signature as an unbreakable "official truth". (Remember kids: Strong cryptography over untrusted data does not make the data more trustworthy.)

Untrusted By Design

This glitter-milk picture demonstrates how any image can be assigned false provenance and signed with a cryptographically valid Google signature. Moreover, this problem also works in reverse: genuine photos with signatures can be easily dismissed as "just another C2PA forgery." Regardless of the ground truth, an analyst cannot determine if a picture is real or fake based on Google's implementation of C2PA; the signature effectively means nothing.

Google's initial announcement made some extraordinary claims that have failed to stand up to scrutiny:

• Claim: "Pixel and Android are bringing a new level of trust to your images with C2PA Content Credentials".
  
  Fact: The devices can be used to sign any file, real or fake, with legitimate C2PA-signed claims identifying that the media came from the camera. This does not introduce a new level of trust; it enables a new way to commit fraud and disinformation.


• Claim: "The Pixel 10 lineup is the first to have Content Credentials built in across every photo created by Pixel Camera."
  
  Fact: This is false. Nikon shipped C2PA Content Credentials in Z6 III firmware in late August 2025, weeks before Google's announcement. Days later, researcher Adam Horshack showed the camera could be used to sign arbitrary images, forcing Nikon to indefinitely suspend the service and revoke every certificate it had issued. Google isn't first; it's just the first to repeat Nikon's mistake with better marketing.


• Claim: "Pixel Camera app achieved Assurance Level 2 ... only possible on the Android platform."
  
  Fact: While they acquired Assurance Level 2 on paper, it appears to be absent from the implementation. Moreover, they stated that protecting the keys from signing arbitrary images is not possible ("Won't Fix (Infeasible)"), so whatever Assurance Level 2 is meant to guarantee, it clearly doesn't hold up in practice on the Android platform.


• Claim: "A private-by-design approach to C2PA certificate management, where no image or group of images can be related to one another or the person who created them."
  
  Fact: While true, this prevents them from revoking future pictures from a known-compromised device. A device that has been rooted and is generating signed forgeries can continue to operate unabated.


• Claim: "Pixel 10 phones support on-device trusted time-stamps, which ensures images captured with your native camera app can be trusted after the certificate expires, even if they were captured when your device was offline."
  
  Fact: While it is true that the Pixel 10 has a built-in trusted time-stamp service, that does not mean that it is only applied to "images captured with your native camera app". This claim is misleading.

In effect, Google's C2PA-enabled devices provide no reliable protections or 'truth' about the media -- and Google knows it.

Flawed Foundations

The problems detailed in this blog are not limited to the Google Pixel or its C2PA Assurance Level 2 rating. These problems are fundamental and impact other C2PA implementations. For example, Evergreen Labs has a C2PA Assurance Level 2 application called "GreenCheckmark" (screenshot) that can be used to sign any image or video as if it came from the device. However:

• C2PA's Conformance Program only checks the paperwork for compliance, not the implementation. In this case, the Conformance Program states that the app has Level 2 assurance.


• According to Evergreen Labs, the app received approval for Level 2, but only implemented Level 1. There is no C2PA-provided or user-identifiable information that identifies this discrepancy.


• Even if the app was fully implemented, Assurance Level 2 requires using Android's StrongBox/TEE, and Google already stated that it knows the environment does not provide adequate protections ("Won't Fix (Infeasible)").

GreenCheckmark isn't the point of failure here; failures are inherited from Google and C2PA.



If you still believe that C2PA works, then I have news for you: Scientists have created multi-colored sheep for dye-free yarn. According to Adobe Inspect (a conforming validator) and CAI Verify, this is legitimate "captured media" from a camera, signed by GreenCheckmark, and it is a Level 2 conformant application (screenshot). Similarly, YouTube's description reports that this video clip from the CGI movie "Big Buck Bunny" is signed by Evergreen Labs and "Captured with a camera" (screenshot).

The same class of vulnerability exists for Android and iOS (except that iOS is harder to root). Moreover, retr0id has additional working demonstrations from many other C2PA-enabled apps, including Proofmode (a Level 1 conformant app; see forgeries at Adobe Inspect and CAI Verify). To date, no C2PA implementations are immune to signing forged media.

We live in an era of deep skepticism, where public trust in visual media is at an all-time low. Proponents of C2PA argue that cryptographic signing solves this problem: if an official photo carries a valid, hardware-backed C2PA signature, the public can trust it. But the truth is that the C2PA signature carries no weight for providing any type of reliable authentication, validation, or provenance. Instead, it turns every device into a powerful tool for laundering disinformation as fact, which is worse than doing nothing.

Special thanks to retr0id, Shawn, and PASAWG for their assistance. All vendors whose products are shown signing forgeries in this blog were notified of the problem. Claude and Gemini were used to help write portions of the code for these demonstrations. (At one point, we had to pause for a few hours after running out of free tokens.) Getting root is hard. Writing the code to implement the vulnerability is a very low bar and can be done with an AI assistant.

I have a secret way to tell what's on people's minds: they all write to me about a topic. This time, there's a news article about the EU AI Act. The big requirement is in Article 50: Transparency Obligations for Providers and Deployers of Certain AI Systems. This recently became enforceable and requires marking AI-generated content:

2. Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated.

This is a problem for companies like OpenAI (ChatGPT), Anthropic (Claude), Google (Gemini), Microsoft, and Adobe. Each provides AI-generation services to people in the EU.

Article 50 does not mandate the use of watermarking. A company could use metadata, logging, fingerprints, or some other technique. However, watermarking is the approach that they all seem to be employing.

I previously documented how image-based watermarking is grossly inadequate. Through empirical testing:

• Google's Gemini has a 1 in 20 error rate, where it fails to detect its own watermarks. Moreover, the detection depends on how you ask the question. Uploading the same image twice with different prompts could generate different results.


• Adobe's TrustMark has a 10%-20% false-positive rate.


• Meta's Stable Signature has a collision rate of 1 in 4.

C2PA tries to resolve this by providing an authoritative claim from a known signer. However, I have repeatedly demonstrated ways to have false signatures applied to both real and fake pictures.

None of these image-based solutions are reliable. In some cases, you're better off flipping a coin. (Or for D&D players, rolling a die.) However, text-based content is special because there is no hidden metadata; you cannot store binary signatures between letters without being noticed. For this reason, many companies are focusing on AI-based text watermarking.

Use Your Words

The new panic? Anthropic and OpenAI will start marking text with an invisible watermark. A lot of people have asked me how this works.

Let's back up a moment. Watermarking is just another form of steganography. Over 30 years ago, there was a steganographic text system called 'texto'. This encoder works like Mad Libs. It has a list of sentences:

The _THING _ADVERB _VERBs to the _ADJECTIVE _PLACE.
I _VERB _ADJECTIVE _THINGs near the _ADJECTIVE _ADJECTIVE _PLACE.
Sometimes, _THINGs _VERB behind _ADJECTIVE _PLACEs, unless they're _ADJECTIVE.
Never _VERB _ADVERB while you're _VERBing through a _ADJECTIVE _THING.
We _ADVERB _VERB around _ADJECTIVE _ADJECTIVE _PLACEs.
While _THINGs _ADVERB _VERB, the _THINGs often _VERB on the _ADJECTIVE _THINGs.
Other _ADJECTIVE _ADJECTIVE _THINGs will _VERB _ADVERB with _THINGs.
Going below a _PLACE with a _THING is often _ADJECTIVE.
...

It also has a list of words. There are 256 'THING' words, 256 ADVERBs, 256 VERBs, 256 PLACEs, etc. It reads the file to encode, chooses a sentence pattern, and replaces the words accordingly. For example, if the first byte is 7, then it will choose the 7th THING word. The results look like gibberish, but it contains a hidden message:

The watch absolutely hugs to the messy moon. I lean cold caps near the sharp squishy planet. Sometimes, brushs love behind soft squares, unless they're lazy. Never move superbly while you're sitting through a yellow arrow. We surely keep around grey white deserts. While yogis deeply smell, the boats often keep on the unique dogs. Other squishy wet tyrants will count wistfully with shirts. ...

To decode it, texto identifies the words and maps them back to values.

Modern Words

The texto approach hides a binary message inside a text block. The AI approach to watermarking uses a similar concept, but the watermarking doesn't need to store a large binary message; it only needs to store a few bytes of data, a small semaphore, or a statistical bias. It can use the rest of the text to repeat the encoding over and over, making it easier to detect.

While texto was limited to 256 nouns, verbs, etc., AI can use a lot more than 256 words and a few fixed sentence patterns. In fact, they can build it into the entire decision tree!

The Kirchenbauer/KGW approach injects a detectable bias into the word selections. This has become the typical approach used today:

1. AI uses random numbers when making choices. This is why repeating the exact same prompt will generate completely different responses. However, for watermark encoding, it uses a secret key as a weighted random number seed.


2. The AI (LLM) approach generates text autoregressively, token by token (or word fragment by word fragment). At each step, it calculates a probability distribution over the entire vocabulary based on the preceding token.
   
   
   • Without watermarking: The AI model calculates a probability distribution over the vocabulary for the next word/token. It then chooses the next word/token based on some kind of sampling algorithm and a list of the most likely candidates.
   
   
   • With watermarking: The KGW solution uses the secret key to assign weights to the possible options. These are often called 'green' words and 'red' words. This biases the sampling algorithm so that it will prefer a green word over a red word. This doesn't exclude red words from being used; it just makes them less likely.

When decoding, the detector uses the secret information used by the generator to reconstruct the selection bias. It uses the preceding token(s) to re-generate the green list for the next token, then counts how many tokens fall into their respective green lists.

The detector usually has some kind of threshold function. For example, "80% green tokens" may be the minimum threshold for identifying the watermark. This is why more text is important to rule out false-positives; it is very possible for a few words in a sentence to have lots of green words, but very unlikely for an entire essay to be mostly green. In general, text from a human will fall far below the threshold for detection, while AI-generated watermarked text will have a detectable bias far above the threshold.

The Catch

Watermarked text has been demonstrated as feasible (at least more accurate than their image watermarking). However, it's far from perfect. The AI vendors have not disclosed their specific watermarking algorithms. While some variation of the Kirchenbauer/KGW approach is likely, it isn't the only option. But regardless of the option, they all have the same classes of weaknesses because they are all based on natural languages. This limits the set of words and available sentence structure choices. The fundamental problems include:

• Word Choice: The word selection can often be sub-optimal. If the AI starts sounding odd, then it's probably due to the weighted words.


• Repetition: The system cannot spot the watermark from one sentence. It needs a paragraph or more. This also means that it is more likely to be redundant, using the same sets of green words multiple times so that it can repeatedly spot the watermark and lower the likelihood of a false-positive.


• Coincidental: In a typical implementation, roughly half of the possible tokens are on the green list. However, it is very possible for a human to coincidentally write a few sentences or paragraphs that are more than 80% from the unknown green list.
  
  The root problem is that the detector must be probabilistic rather than absolute. A human-written passage can coincidentally produce a watermark-like statistical pattern. The detector must use a threshold that balances a trade-off between the false-positive rate and detection sensitivity.


• Translations: Not everyone is a native English speaker. (Or Spanish, Chinese, or whatever language you are writing in.) It is very common to see people use AI to translate text from their native language into the target language. However, the translations may end up watermarked. This can become a problem. For example, nearly every major academic publisher and scientific organization (including IEEE, ACM, Elsevier, Nature, and Springer) treats undisclosed AI-generated text as scientific misconduct. Many top-tier sci-fi and literary markets, like Clarkesworld and Asimov's Science Fiction, explicitly forbid AI-generated text. Wired and the Associated Press have similar blanket bans. If they use watermarking detection, then they may erroneously exclude participation from non-native English speakers.


• Training/Trained: AI learned how to write from humans, but humans learn how to write like those around them. With the younger generation (well, people younger than me) spending so much time interacting with AI systems, they will inevitably learn some of those odd wording styles. This will make real human text read more like AI, and could even appear weighted toward more green wordings.


• Versions: Today's AI uses today's watermarking approach. A newer AI system will need to use a revised approach. However, this can cause a problem when an old watermark is no longer detectable by a newer system. Moreover, many AI companies push out code changes without announcements. Text that tested positive for having a watermark yesterday may appear human-written tomorrow.
  
  In order to maintain backwards compatibility, the detector would need to know which model generated the text, which watermarking algorithm was used, which key/version, etc. This makes long-term detection much more complicated and unsustainable if there are rapid revisions.


• Paraphrasing and Removal Attacks: The original KGW algorithm was able to survive some paraphrasing attacks. (Is it called paraphrasing or plagiarism when you rewrite AI-generated text?) More advanced rewriting can remove the watermark. For example, passing the watermarked text through a lightweight local model, running it through a translator, or manually swapping a few synonyms may be enough to disrupt the green-token sequence and appear to be human-generated. As Google noted with their own SynthID-text system:
  

  [SynthID for text] performs well even under some transformations, such as cropping pieces of text, modifying a few words and mild paraphrasing. However, its confidence scores can be greatly reduced when an AI-generated text is thoroughly rewritten or translated to another language.



• The Editor Problem: Most professional writing organizations use human copyeditors to review the text, fix grammar, etc. Today, smaller organizations often rely on AI systems to act as proofreaders. The problem is that large text edits by AI systems can introduce watermarking into otherwise human-created text. (Full disclosure: Gemini and ChatGPT proofread this blog before my human editor received it. They each caught spelling errors and a few details that my human editor would have likely missed. However, there are no large AI-text edits in this blog.)


• Spoofing Attacks: Adversaries can reverse-engineer green lists or use watermarked text outputs to create false positives. This could be used to falsely accuse or frame human authors.


• Interoperability: Article 50 explicitly says that solutions "shall ensure their technical solutions are effective, interoperable, robust and reliable". However, every vendor is keeping their watermarking solutions private. That ensures that they are not interoperable. On the flip side, if they made the details public then it would assist competing detectors, simplify removal, and enable forgeries. (It's a no-win situation for the watermarking companies.)

While text-based watermarking is an interesting academic exercise, I question whether it is ready for widespread public dissemination.

The Compliance Paradox

The EU AI Act's Article 50 effectively writes a technological fantasy into law. It requires some way to mark, tag, or label synthetic text "in a machine-readable format and detectable as artificially generated or manipulated". These regulators have created a legal requirement for technology that simply does not exist in a reliable form today.

Ironically, the lawmakers added a caveat that solutions only need to be as "robust and reliable as far as this is technically feasible". This creates a bizarre paradox: companies are pressured to deploy flawed, easily bypassed schemes just to demonstrate legal compliance. We are left with regulatory compliance theater, where algorithms pretend to detect what cannot be reliably detected, and users are handed a false sense of security.

I'm wrapping up my investigation into invisible watermark algorithms and I am extremely disappointed. Not only do none of the modern AI-based algorithms work as they claim, it turns out that they are all making the same fundamental mistake.

I previously evaluated Google's SynthID and Adobe's TrustMark algorithms. Both of them claim to have incredibly accurate results.

• According to Google's peer-reviewed and published paper, they claim to have a true positive rate (TPR) above 99.97% -- meaning that they will miss their own watermarks no more than 3 in 10,000 times. However, my own empirical testing found that is it much closer to 1 in 20. Moreover, SynthID is proprietary and only accessible through Google's "Gemini" AI system. Gemini has been observed hallucinating results and providing contradictory conclusions depending on how the question is phrased.


• According to Adobe's Content Authenticity Initiative, their TrustMark "can exceed 96% bit accuracy at around 42-45dB PSNR quality under severe noise degradations". However, that statistic focuses on resilience and not accuracy. In my empirical tests, I found that TrustMark has a 10%-20% false positive rate, effectively making it useless. (If you see a TrustMark signature, then it is very likely random noise and not an actual signature.)

This time, I evaluated Meta's "Stable Signature" algorithm. (Their paper and code are in GitHub.) This system encodes a 48-bit sequence into the picture's visual content. The idea is that you can encode a unique 48-bit sequence as your watermark. If your decoder finds the same 48-bit sequence, then it can identify your own watermark.

WARNING: This blog entry leans heavily into math and statistics to prove that Stable Signature, TrustMark, and SynthID are nowhere near as reliable as their developers claim.

The Basic Algorithm

Traditional (non-AI) invisible watermarks typically hide in subtle locations, such as the least significant bits, changes in brightness (e.g., Digimarc) or the frequency spectrum (DCT or FFT). There is always the risk that image encoding could corrupt the hidden data, so these algorithms typically rely on repetition over the image to help identify the true signal. In addition, they may include error correction code (extra bits in the data) to fix any minor data errors.

However, there is a problem with the traditional approaches: injecting hidden data in the image could create visible distortions. The modern approach uses an AI system to better hide the data with less added distortion.

As with SynthID and TrustMark, Stable Signature encodes binary data and uses an AI-model to decide where to hide it in the image. The AI is tuned to minimize visible distortions when embedding the data. Later, an AI-based decoder looks at the image and identifies the likely location where bits are stored, then it extracts the data.

There is always the case that the data may be mixed with noise. Different AI-based watermarking systems rely on different techniques for reducing the noise. For example:

• Google's SynthID only stores a few bits of data (effectively a flag or version number). This allows them to use a lot of data as repetition and to increase the accuracy rate.


• Adobe's TrustMark uses the Bose-Chaudhuri-Hocquenghem (BCH) algorithm. This acts as a combination of checksum and error correcting code that should reduce the number of errors.

Meta's Stable Signature uses a simple Hamming distance.



The Hamming distance measures the number of bits that need to be swapped in order to correct the code. In effect, it defines a set of stable states (e.g., 10110 and 11000) and places a ring around each state that represents the single bit changes. If you change enough bits, then you will reach a different stable state.

According to Meta's Stable Signature research paper, the 48-bits should be uniformly distributed and cites a "false positive rate below 10^-6", or 1 in one million. This means you can choose a 48-bit sequence to use as your signature. Every picture will generate a 48-bit sequence, and the sequence can vary a little based on noise in the picture. However, if you find a code that is within a short Hamming distance of your code (e.g., within 6 bits difference), then you can determine that it is the same code with a high reliability.

At least, that's the theory.

Empirical Testing

I went into this experiment assuming that everything works like they claim. I want to be able to reliably identify invisible watermarks associated with Meta. What I don't know is what sequence they use, or whether they use multiple codes depending on whether it comes from Meta's AI system, Facebook, Instagram, WhatsApp, etc.

Fortunately, this is something I can test! I grabbed an uncurated sample of pictures from FotoForensics: the first 10,000 unique images uploaded last month (May 2026). If the bit sequences are uniformly distributed with a "1 in 1 million" collision rate, then I should see a huge number of unique bit sequences and a few small clusters around pictures from Meta (Meta AI, Facebook, Instagram, etc.). Those clusters will represent the invisible watermarks used by Meta.

The results from my empirical test were definitely not what I expected. I found:

• No clusters associated with any Meta images. This suggests that Meta does not use their own Stable Signature watermarking software found on GitHub.


• With a random distribution, there should be no clusters. However, I had 25 different pictures that had the exact same bit sequence: 110110100111111011101001111000100111011000011101. With a 1 in a million collision rate, this should not happen! These pictures came from very different sources. Here's four of the 25 pictures (ranging from planets to light bulbs to text with a transparent (black) background):
  
  
  
  All of these pictures have dark/black backgrounds and something bright in the middle. This suggests that Stable Signature operates more like a perceptual hash than an invisible watermark.


• Stable Signature uses a Hamming distance to identify a cluster. If I assume the 25 pictures are the center (centroid) of the cluster and use a 6-bit Hamming distance, then there are 356 pictures that are similar. And if I assume that the 25 pictures are not the center but part of a cluster, then a Hamming distance of 6 has a cluster of 450 pictures centered 3 bits away, at 110110000111111011101011111000100111001000011101. This cluster represents 4.5% of the uncurated image data set! Here are a few samples from this larger cluster:
  
  
  
  (I'm explicitly not sharing pictures with personal information, like invoices, recognizable people, and GPS information.)

It's not just one random cluster that is massively large (450 pictures out of 10,000). There's a cluster of 184 pictures at 110101001011001011001011111000100111001000011101, 58 pictures at 110100000011111010001001111000100111011000011101, etc. I found over 60 clusters with more than 10 pictures each at a Hamming distance of 6. That should not happen with a "1 in 1 million" collision rate.

Independent Analysis

I went back to Meta's research paper to see if I could find the discrepancy. And there it was, in section 3.1: They tested their system against the hypothesis that the 48-bits are each independent and uniformly distributed. The problem is, they use one neural network to generate the bits. That explicitly means that the bits are dependent, not independent.

Their paper assumes a binomial distribution. That is, given an arbitrary image, the 48-bits represent a random coin flip. The math becomes:
P(X ≤ T)=∑Tk=0(48k)(0.5)k(0.5)48−k

This computes the probability of 48 random bits being within a Hamming distance (T). The probabilities table becomes:

Hamming Distance Threshold (T)	Bit Error Rate (BER)	Probability of a Random Image Matching by Chance
14 bits or fewer	≤ 29.17%	1 in 362.63
13 bits or fewer	≤ 27.08%	1 in 957.81
12 bits or fewer	≤ 25.00%	1 in 2,788.35
11 bits or fewer	≤ 22.92%	1 in 8,999.08
10 bits or fewer	≤ 20.83%	1 in 32,416.80
9 bits or fewer	≤ 18.75%	1 in 131,390.28
8 bits or fewer	≤ 16.67%	1 in 605,094.89
7 bits or fewer	≤ 14.58%	1 in 3.20 Million
6 bits or fewer	≤ 12.50%	1 in 19.83 Million
5 bits or fewer	≤ 10.42%	1 in 146.19 Million
4 bits or fewer	≤ 8.33%	1 in 1.32 Billion
3 bits or fewer	≤ 6.25%	1 in 15.24 Billion
2 bits or fewer	≤ 4.17%	1 in 239.15 Billion
1 bit or fewer	≤ 2.08%	1 in 5.74 Trillion
0 bits (perfect match)	= 0.00%	1 in 281.47 Trillion

Meta's paper says that they use a Hamming distance of 7 bits (requiring 41 of 48 bits), which matches their claim of a "false positive rate below 10⁻⁶". However, I'm seeing problems at a Hamming distance of 6 (should be 1 in 20 million) and even collisions at 0 (1 in 281 trillion)!

The Core Problem

There is clearly a discrepancy between the theoretical probabilities and the empirical testing. When I looked back over Meta's research paper, I saw the problem:

According to Meta's paper, each of the 48-bits are independent. In a perfectly independent 48-bit hypercube, un-watermarked images should scatter uniformly across all 2⁴⁸ possible values. However, neural networks map a non-linear manifold (a multi-dimensional wavy surface) through this hypercube. This mathematical landscape is warped with its own peaks, ravines, and valleys. It has attractors that form clusters, and repulsers that form voids where stable values can never exist; this is a feature of a neural network. And most importantly, the output bits are explicitly not independent.



The left diagram illustrates an expected uniform distribution if all of the bits were independent. The right diagram are the types of theoretical clusters that form when the bits are dependent. There should be clusters around attractors and voids (areas with no dots) from the repelling regions.

Moving from theoretical to empirical, I graphed the data. The 48 bits can be represented as bytes. I took the first 24 bits and converted them into 8-bit red, green, and blue pixel colors. If the data is truly random, then the colored dots should be distributed across the RGB cube. However, if the bits are dependent, then there should be very clear clusters, structures, and voids. Here's the graph:



Yes, there are very clear structures that look like planes and lines. Within the planes are clusters, and outside the planes are very large voids -- areas where there are no dots at all. The data generated by Meta's Stable Signature implementation fails this basic test for independence.

The biggest cluster that I found represents a Zero Signal Bias (ZSB). When their neural network doesn't find a watermark, it moves the 48 bits toward a strong attractor, like a massive gravitational well. At 6 bits error, it should have a collision of around 1 in 20 Million. But in reality, my 10,000 pictures had a cluster of 450 images within 6 bits due to the ZSB. That's an error rate of around 1 in 22 with the ZSB alone. If we add in all of the other clusters that contain at least 10 pictures, then 2327 pictures are in various clusters; we're looking at an error rate around 1 in 4 -- and that's at a Hamming distance of 6, which is more conservative than their paper's Hamming distance of 7. (In AI terms, this is a representation collapse or structural bias that is typical for deep neural networks.)

(As an aside: Given their "1 in 1 million" claim, I could look for any clusters of 2 or more pictures. At clusters of 2 or larger, 5,237 of the 10,000 test images were in clusters, or 52%. If you show their algorithm 10,000 pictures, then there is a better-than 50% chance of a false positive match.)

Less Than Random

It's one thing for me to claim that there are visible clusters and to show pictures of clusters, but another to prove it mathematically. (Time to dust off my college textbooks from "Introduction to Statistics"...)

I fed Meta's code the first 10,000 images from May 2026. A few of the images were in unsupported formats (HEIC, WebP, and a few corrupted JPEG files), resulting in 9,847 viable pictures. I evaluated this data with elements from the NIST Statistical Test Suite (SP 800-22) for randomness, including a monobit test and Chi-Squared (χ²) test for independence.

The monobit test determines if the baseline frequency of adjacent bits seems independent.

• Total Bits Processed: 9,847 pictures × 48 bits per signature = 472,656 bits
• Observed Count of Ones ('1'): 266,419
• Observed Count of Zeros ('0'): 206,237
• Expected Count (E): 236,328 for each.

Running a simple standard Chi-Square Goodness-of-Fit test for this bit balance:
χ2=(266419 − 236328)2236328+(206237 − 236328)2236328= 3831.41 + 3831.41 = 7662.81

• In mathemat-ese: with 1 degree of freedom, a χ² statistic of 7,662.81 yields a p-value infinitely close to 0.0 (p ⋘ 10^-100). (As an aside, most Chi-square tables usually evaluate the 1 degree of freedom up to around χ²=10. This χ² value is so astronomically high that the probability p effectively becomes zero.)


• In English: That's definitely not random or independent.

The watermark extraction is strongly biased toward producing 1s over 0s across global arbitrary images (roughly 56% ones to 44% zeros). This immediately violates the uniform distribution assumption.

The second test is the Chi-Square (χ²) Test for Serial Independence. If the bits were independent, the transition probability between adjacent bits would just be the product of their individual probabilities. This table shows the occurrence rate of the transition pairs across all of the observed 10,000 (well, 9,847) pictures:

Transition Pair	Observed Count (O)	Expected Count under Independence (E)
0 to 0	106,750	90,051
0 to 1	95,296	116,186
1 to 0	95,302	116,186
1 to 1	165,461	149,976

χ2=∑(O − E)2Eχ2=16699290051+(−20890)2116186+(−20884)2116186+154852149976=3096.7 + 3756.2 + 3754.0 + 1599.0=12,205.9

• In mathemat-ese: With 1 degree of freedom for the transition contingency table (accounting for fixed margins), a χ² value of 12,205.9 gives a p-value of 0.0.


• In English: Ain't no way this is random or independent.

And as if this wasn't conclusive enough, there are other tests we could apply:

• Static Tail Patterns: Looking closely at the end of the 48-bit sequences, a massive cluster of strings end explicitly in ...111101 or ...00111101. Additionally, bit position 46 is nearly always "1" (228 zeros vs 9619 ones, or 97.7% of the time it is "1"), position 47 is "0" (8958 of 9847 images, or 90.97%), and position 48 is "1" (found with 9696 images, or 98.5%) across thousands of uncurated, real-world images.


• Structural Clustering: Certain bit columns share an extraordinarily high Mutual Information score (I(X;Y)). For example, knowing the output of bit position 12 gives you better than an 80% accuracy in predicting bit position 28.

The assumption of a "uniform distribution over arbitrary pictures" relies on the idealistic premise that random natural image features project uniformly across the decision boundaries of a network. However, because the extraction network maps inputs to a constrained, highly continuous hyper-dimensional manifold, the network's latent layers natively enforce structural smoothness.

For the TL;DR crowd:
Meta's researchers made a fundamental mistake when computing their accuracy rates. It's not a "1 in 1 million" chance of a false match, it's closer to 1 in 4 -- because the 48 bit values per signature are not independent.

As I re-read Meta's research paper, I realized that the statistical error wasn't an oversight; Meta's researchers explicitly acknowledged the problem. In their paper (Section 4.1), they wrote:

Second, we observed that W’s output bits for vanilla images are correlated and highly biased, which violates the assumptions of Sec. 3.1 [the section about independent statistical test methods].

In other words, they recognized that the extracted bits are not independent. Despite this, their published false-positive analysis still relies on the assumption that the bits are independent.

Widespread Problems

Knowing that Meta's accuracy rate is grossly inflated due to assuming bit-wise independence when there is none, I looked back over Google's and Adobe's papers for their own watermarks. Did Google's and Adobe's researchers make this same mistake?

• Google's SynthID research paper talks in terms of True Positive Rates (TPR). They do make this same "bit-wise independent" mistake, but it's obfuscated in the paper. You can see the error in their Equation 3 (PDF page 8), where they assume there is a uniform (independent) distribution. Their paper hyperfocuses on the true positive rate and never addresses the false positive distribution. (Either they didn't know to look, or they knew and decided to not report it because it would expose a serious weakness in their solution.)


• Adobe's TrustMark research paper also makes assumptions of independence. You can see this in their PDF with the binary cross-entropy loss in Section 3.1.4. This mathematically treats each bit position as an independent Bernoulli trial. (By definition, a Bernoulli process strictly requires independence.) In their experiments (Section 4.1), they wrote "At test time, every image is associated with a random watermark", but they never tested if the random watermarks were similar to each other.

This introduction-to-statistics mistake is found in all three of these invisible watermarking technologies. The detections produced by these systems are so unreliable that an analyst cannot determine whether a reported detection is real or a false positive, or whether a reported non-detection is genuine or a false negative.

It's also worth noting that, shortly after releasing Stable Signature, Meta developed another algorithm: Pixel Seal. (Not to be confused with my own Secure Evidence Attribution Label / SEAL technology.) Pixel Seal moves to a 256-bit payload to increase the capacity, and their related model, Chunky Seal, pushes up to 1024 bits. While Meta's approach focuses heavily on addressing the invisibility side using an adversarial-only discriminator, the underlying approach still uses a neural network mapping. Using more bits only exacerbates this flaw.

Potential Uses

Algorithms can have uses. For example, Meta, Google, and Adobe are training their own AI models on images that they encounter. To prevent poisoning their training sets, they want to exclude images generated by their own systems. In this regard, watermarking does help them. For example, if Meta excludes an extra 25% of images (from false positives), then they still have a lot of images that they can train on.

However, that same usage does not work with legal cases. For example, consider an insurance company. Most insurance claims today include photographic evidence. The company wants camera-original photos, but have to use whatever the customer submits. The problem is that there is a lot of insurance fraud. In theory, seeing a watermark from an AI system like Meta, Google, or Adobe, should be great for identifying and ruling out fraud. Unfortunately, Stable Signature, SynthID, and TrustMark are so inaccurate that none of them can be trusted; it's not even worth testing to see if customer photos contain these invisible watermarks.

For these watermarking systems, I'm talking about very high error rates: roughly 1-in-4 for Meta, 1-in-5 for Adobe, and 1-in-20 for Google. But let's pretend that they work much better, like a 1-in-20,000 false positive rate. An insurer processing 100,000 claims per month would expect to accuse around 5 completely honest customers of fraud each month. Falsely denying 5 out of 100,000 claims? That creates a toxic customer service nightmare, severe legal liability, and fines from regulatory bodies for bad-faith claim denials. This could even become a class-action lawsuit that they couldn't win.

As bad as it is for insurance and financial institutions, there are much higher stakes at play. The EU AI Act (Article 50(2)), China's GB 45438-2025, California SB 942, and similar legislation are moving toward mandating AI content watermarking.

The failure of these three leading systems, from three Fortune-500 companies, to meet their own claimed accuracy rates is not just an academic curiosity. Regulators and courts will employ these systems for attribution and fraud detection. Reliable AI-based watermarking technology is not ready.

Three companies. Three algorithms. Three different research teams. The same fundamental error. The false positives won't go on trial. People will.