Under what circumstances does the metadata get added? When Claude Code CLI needs a media file usually I just see it run an imagemagick or ffmpeg command to create one, which isn't going to have C2PA metadata.
Edit: I asked Opus 5 for "a jpg of a cat" via the web chat. It did it by writing an svg and then rendering that to a jpeg. Clicking the "download" button saved a jpg which did have C2PA metadata, seemingly injected by the artifact-presentation thing.
Then I told it to put it in a zip file. I downloaded the zip, and the extracted jpeg had no C2PA metadata.
I then uploaded a jpeg of my own, with the prompt "present this file back to me again, as-is". The downloaded file had C2PA metadata injected.
Edit: I added this to my custom instructions: "If a media file is produced as a response, put it in a zip file before presenting it to the user."
As far as I can tell, this and the recent change to add watermarking to text outputs[1], is to become compliant with the EU AI Act[2] and CA's AI Transparency Act[3], SB-942[4]. For large enough companies, all generated AI content is required to have watermarking.
Good point, re: the CA law. The rule for text output may be limited to the EU AI Act, which Anthropic notes as a reason in the linked article about its implementation.
Watermarks should even be removable "if substantially proofread or altered" in the EU for the text from an LLM, that you verify to be as true as if you wrote it yourself.
Not clear is this is a may, shall, or should. It certainly is not a must.
Watermarks cane even be removed, "if substantially proofread" in the EU, law says. For the text from an LLM, that you verify to be as true as if you wrote it yourself.
Not clear is this is a may, shall, or should. It certainly is not a must.
This is really fascinating. Even the AI companies have incentives to reject AI generated content. It's like they want you to use AI for everything, but they don't want AI output fed back to them.
At work, right after an AI training, we were asked to use our "authentic" voice when writing mid year reviews.
Or maybe it's good for everybody to use LLMs for some things but not other things, as opposed to some people should use it for everything and others for nothing. These companies are definitely using LLMs for coding.
What is interesting to me is that stripping the C2PA data is easy, but faking it is hard.
You can resave the file and the "made with Claude" signal disappears, but you cannot make a random file pass as Claude-made without Anthropic's signing key. So the useful guarantee is one-way. No signature means almost nothing.
The goal of C2PA is that cameras will start to emit C2PA credentials. You will then have 3 situations:
* C2PA confirms a photo is authentic
* C2PA confirms a photo is AI generated
* C2PA missing, you don't know.
I reckon we will only see "C2PA missing" being treated as suspect in select situations (perhaps Reuters will require C2PA from their photojournalists, for example)
Camera C2PA can never meaningfully confirm that a photo is authentic, it bears about as much credence as EXIF metadata. It's like saying the existence of DRM confirms that a movie hasn't been pirated.
"Faking" it is trivial. You don't need their signing keys when you can just ask them to sign whatever you like. Upload your own file with the prompt "present this file back to me again, as-is".
Your observation is sharp, but it's not just metainfo—it's load-bearing text. To remove it, you need to delve deep and alter the tapestry of carefully selected words.
How long before they change the terms and conditions to subtly claim ownership of your files? When you write code they already insert Co author attribution/
Say I write a text by hand And then I tell it to clean up the grammar and fix some sentences Did it make it?
This is also interesting for those companies that siphoned the entire open web
The two competing legal arguments regarding copyright of LLM output are "it's like hiring a monkey" (author is the LLM, which is not a person, thus can't hold copyright and can't assign it to you) and "it's like taking a photograph" (the LLM is a machine through which the prompting person expresses their creativity, just like a camera). In no scenario is Anthropic the author of the work.
If we settle on the monkey analogy the Anthropic owns the monkey, but the owner of a monkey doesn't own copyright for the creations of the monkey. If we settle on the camera analogy, anthropic claiming ownership would be like Canon claiming they own pictures you take.
What Anthropic could do is to change the terms to give themselves a non-exclusive global perpetual license to use everything Claude makes
They have said and tried some wild things already, like trying to get open weights effectively banned, which I believe they still think is in society's best interest (more that they think they know what's best for everyone)
Could this be used to perform some sort of distillation or exploit? e.g. reminds me of the OWASP guideline on attack vectors where knowing if an ID is present or not in the database can be a form of exploit, like in password resets where they will say 'email foo@bar.com not found' rather than 'If foo@bar exists we have sent an email to foo@bar' or some other generic equivalent
They still scrape code, I'd guess, e.g. from GitHub?
And there's tons of Claude-generated code there.
Also, I'd guess this is not just to prevent any AI-generated code in the training data, but specifically their own.
Percentage of users who put out their code on the web and also have a plan where Anthropic promises not to train on their data is problem also low.
So not excluding own code could be a real issue, since it would be impossible to deduplicate the training and RILHF data from their sessions with the code accessible elsewhere, and written by the very same users.
I'm wondering why they have restricted file types. You can't check a PDF for example... surely the main use case for people will be to check if a document was produced or edited by an LLM? That could be an attractive (if not misunderstood) proposition for academics
> I'm wondering why they have restricted file types. You can't check a PDF for example...
TFA/page actually seems incomplete. Text uses a completely different watermark format (an actual watermark as opposed to a provenance/authenticity signature), so it makes sense to me that they're not claiming to be able to scan PDFs when they can't yet incorporate that signal.
On a linked page, they say:
> Watermark detection is currently in private preview [...]
Luckily all pdfs I made in academia have been generated from source (tex or derivatives, asciidoc) and llm are better in generating source than pdf. Even people that didn’t use tex used Word to generate the pdf.
Source can have markers inserted in it as weighted word choice or phrase choices, how commas are inserted, etc., so that the output can still can be identified. In other words just because it's source doesn't mean it can be watermarked.
Unfortunately this is just for Media? Some manual tells for Excel or PDFs is to check the author. Claude creates PDFs via wkhtmltopdf so the PDF Producer will be Qt and the Content Creator is wkhtmltopdf. Xlsx files are being created via openpyxl so in the metadata that is the author.
The actual implementation matters less than the fact that they can be perceived to be working on something.
You're never going to have a perfect solution here, and perfect is the enemy of good.
The idea of forcing an entity (or entire categories of entity) to add traceable watermarks to all the work they create is... an interesting precedent.
Translating that into the human realm, it would be like if every digital thing you touched were "fingerprinted" with your identity, or at least with a proof of what category of individual you have been categorized as.
Due to the flood of AI, which may soon become indistinguishable from human video, I expect to see something similar. Blue checkmark, but for humans.
> To detect watermarks embedded in text we have a Detection API which is currently in private preview to eligible organizations as required under EU law.
How would this get more data though? It's more so about being compliant with the EU's new AI laws (which I guess could get them more data since more users)
For anyone wondering who the text watermarking works:
Instead of using an arbitrary random number generator
to pick the next word, watermarking uses a secret key
and a few words that come before to settle what
word the model should pick.
They basically twiddle with the distribution of used output tokens so that the pattern is recognizable by machine means. Nothing to do with special characters.
> Knowing where content came from, and whether AI was involved, makes it easier to trust what you see online.
That’s a cute way to imply their service is used to generate misinformation. They are basically saying to not trust the AI content made from their own product :)
Like they should ban astronaut on a horse images or what? Misinformation becomes misinformation upon the fact of presenting it as so, not the fact of creation
A watermark answers "did this model touch the file?" It doesn't answer "did a person read it and stand behind it?"
Detection will get gamed, and a lot of useful text is already mixed. The more useful check, for me, is still human: would I put my name on this before it goes out?
this is going to be the big thing in the next 100 years for me. where there is a human, there will be context, passion and meaning. it's the old saying: machine can tell you what, but only humans can tell you why.
To see the signed metadata: go to claude.ai and ask: "generate an image of a cat". When you'll go to download the image, it will offer an SVG and a PNG version that both have the metadata.
In the case of SVG, it's literally `<metadata><c2pa:manifest>(signature)...` right after the <svg> tag. In the case of PNG, it's stuffed as a `caBX` chunk.
To declaw these files, you can ask in your CLAUDE.md to run `exiftool -jumbf:all=` (for jpg & png), and to simply directly remove the <c2pa:manifest> tag in SVG files. I haven't looked at video files yet.
TFA is talking about C2PA, a regular signature in the meta data. A lack thereof means nothing, but if it's there, the false positive rate should be near zero.
Yes, but I also think it will be trivially by-passable if you pass your output through another LLM. At least for text.
It might help catch students cheating, but not real spam-bot usage. As soon as platforms start checking for watermarks spambots will add anti-watermark passes.
Under what circumstances does the metadata get added? When Claude Code CLI needs a media file usually I just see it run an imagemagick or ffmpeg command to create one, which isn't going to have C2PA metadata.
Edit: I asked Opus 5 for "a jpg of a cat" via the web chat. It did it by writing an svg and then rendering that to a jpeg. Clicking the "download" button saved a jpg which did have C2PA metadata, seemingly injected by the artifact-presentation thing.
Then I told it to put it in a zip file. I downloaded the zip, and the extracted jpeg had no C2PA metadata.
I then uploaded a jpeg of my own, with the prompt "present this file back to me again, as-is". The downloaded file had C2PA metadata injected.
Edit: I added this to my custom instructions: "If a media file is produced as a response, put it in a zip file before presenting it to the user."
Well, guess they are just trying to conform all the weird EU checkmarks
I'm not sure that 'preventing ai-generated fraud' is really a 'weird checkmark'
5 replies →
As far as I can tell, this and the recent change to add watermarking to text outputs[1], is to become compliant with the EU AI Act[2] and CA's AI Transparency Act[3], SB-942[4]. For large enough companies, all generated AI content is required to have watermarking.
[1] https://www.anthropic.com/news/claude-text-watermark
[2] https://digital-strategy.ec.europa.eu/en/policies/code-pract...
[3] https://www.kqed.org/news/12095398/new-california-law-requir...
[4] https://www.leginfo.legislature.ca.gov/faces/billTextClient....
The CA law in link 4 clearly says it covers "image, video, and audio output", not text output. It's right in the second paragraph.
Good point, re: the CA law. The rule for text output may be limited to the EU AI Act, which Anthropic notes as a reason in the linked article about its implementation.
This is just C2PA for Claude-processed files, apparently. It's not related to their text watermarking.
C2PA is file metadata and can be trivially stripped away, unlike hidden watermarks, e.g. SynthID.
That is where their responsibility ends in terms of the EU AI Act. And that is fine and how it should be, no secret watermarks.
Watermarks should even be removable "if substantially proofread or altered" in the EU for the text from an LLM, that you verify to be as true as if you wrote it yourself. Not clear is this is a may, shall, or should. It certainly is not a must.
Watermarks cane even be removed, "if substantially proofread" in the EU, law says. For the text from an LLM, that you verify to be as true as if you wrote it yourself. Not clear is this is a may, shall, or should. It certainly is not a must.
2 replies →
There's no way text watermarking isn't causing a degradation in writing quality.
9 replies →
Watermarks can even be removed, "if substantially proofread" in the EU, law says (for now. There might be a transition period on that).
For the text from an LLM, that you verify to be as true as if you wrote it yourself.
Not clear if making watermarks removable is a may, shall, or should, according to legislation. It certainly is not a must.
The way I read it, I thought they would just check if the file hash exists on their end. Is that not the case?
That’s incompatible with the file not leaving your browser, which you can trivially verify with the Network tab in chrome (or wireshack, etc).
This is just a C2PA metadata checker.
2 replies →
> I thought they would just check if the file hash exists on their end
That would invalidate the hash on minor changes. Too much effort and not enough return.
[flagged]
This is really fascinating. Even the AI companies have incentives to reject AI generated content. It's like they want you to use AI for everything, but they don't want AI output fed back to them.
At work, right after an AI training, we were asked to use our "authentic" voice when writing mid year reviews.
Just like the tech company owners not allowing their kids access to social media…
Not really. This is more like not pumping the exhaust back into the intake.
2 replies →
Or maybe it's good for everybody to use LLMs for some things but not other things, as opposed to some people should use it for everything and others for nothing. These companies are definitely using LLMs for coding.
What is interesting to me is that stripping the C2PA data is easy, but faking it is hard.
You can resave the file and the "made with Claude" signal disappears, but you cannot make a random file pass as Claude-made without Anthropic's signing key. So the useful guarantee is one-way. No signature means almost nothing.
The goal of C2PA is that cameras will start to emit C2PA credentials. You will then have 3 situations:
* C2PA confirms a photo is authentic
* C2PA confirms a photo is AI generated
* C2PA missing, you don't know.
I reckon we will only see "C2PA missing" being treated as suspect in select situations (perhaps Reuters will require C2PA from their photojournalists, for example)
Camera C2PA can never meaningfully confirm that a photo is authentic, it bears about as much credence as EXIF metadata. It's like saying the existence of DRM confirms that a movie hasn't been pirated.
16 replies →
[dead]
"Faking" it is trivial. You don't need their signing keys when you can just ask them to sign whatever you like. Upload your own file with the prompt "present this file back to me again, as-is".
For removing cspa you can use https://metastripper.devkram.de/
Yeah, C2PA solves a different problem than watermarking, i.e. authenticity/provenance.
Next month: Show HN: How to successfully spoof C2PA and make any file as claude made!
And it's one-way in the direction that typically no one cares about.
Anthropic's legal team cares
Up next:
1. Generate a bunch of responses with both Claude and various non-Claude LLMs (ChatGPT, Gemini, Kimi)
2. Train a discriminator model that can differentiate Claude vs. non-Claude
3. Train a de-watermarking model using the discriminator model as loss
Or just write a 6 line program to remove the metainfo from the file?
Your observation is sharp, but it's not just metainfo—it's load-bearing text. To remove it, you need to delve deep and alter the tapestry of carefully selected words.
edit: almost forgot the "—"
1 reply →
very difficult to remove text watermarking
2 replies →
You just described how Pangram works.
How long before they change the terms and conditions to subtly claim ownership of your files? When you write code they already insert Co author attribution/
Say I write a text by hand And then I tell it to clean up the grammar and fix some sentences Did it make it?
This is also interesting for those companies that siphoned the entire open web
I don't see a legal pathway to do that
The two competing legal arguments regarding copyright of LLM output are "it's like hiring a monkey" (author is the LLM, which is not a person, thus can't hold copyright and can't assign it to you) and "it's like taking a photograph" (the LLM is a machine through which the prompting person expresses their creativity, just like a camera). In no scenario is Anthropic the author of the work.
If we settle on the monkey analogy the Anthropic owns the monkey, but the owner of a monkey doesn't own copyright for the creations of the monkey. If we settle on the camera analogy, anthropic claiming ownership would be like Canon claiming they own pictures you take.
What Anthropic could do is to change the terms to give themselves a non-exclusive global perpetual license to use everything Claude makes
Microsoft Word has not claimed ownership in 40 years. Why would Anthropic do?
Because Anthropic are the type of strange people who think that language models have feelings.
1 reply →
They have said and tried some wild things already, like trying to get open weights effectively banned, which I believe they still think is in society's best interest (more that they think they know what's best for everyone)
Microsoft Word does not write your documents for you.
1 reply →
I'm so curious if you're actually able to keep a straight face while typing that.
Does code also get watermarked? Surely you can't add non printable characters, except perhaps in comments?
There is a text watermark detector API but it's in private preview right now. Very curious if it works with code as well as prose.
https://support.claude.com/en/articles/16266773
Much harder to do apparently, but they can still watermark comments.
https://www.anthropic.com/news/claude-text-watermark, Ctrl-F for "What about code?"
Could this be used to perform some sort of distillation or exploit? e.g. reminds me of the OWASP guideline on attack vectors where knowing if an ID is present or not in the database can be a form of exploit, like in password resets where they will say 'email foo@bar.com not found' rather than 'If foo@bar exists we have sent an email to foo@bar' or some other generic equivalent
Probably not, given that both C2PA and SynthID-Text are statelessly verifiable.
I think all of this it's so they don't get ai generated content in their training data
Or it’s so they can continue to operate in the EU where this is required.
So they do it as a free service to the other LLMs?
They still scrape code, I'd guess, e.g. from GitHub?
And there's tons of Claude-generated code there.
Also, I'd guess this is not just to prevent any AI-generated code in the training data, but specifically their own.
Percentage of users who put out their code on the web and also have a plan where Anthropic promises not to train on their data is problem also low.
So not excluding own code could be a real issue, since it would be impossible to deduplicate the training and RILHF data from their sessions with the code accessible elsewhere, and written by the very same users.
This thing works using WebAssembly - https://claude.com/_next/static/media/c2pa_verify_wasm_bg.53... - which is a WASM compiled version of this Rust library: https://github.com/contentauth/c2pa-rs
>Supported formats: JPG, PNG, GIF, WEBP, TIFF, HEIC, AVIF, SVG, DNG, JXL, MP4, MOV, AVI, WAV, MP3, M4A, FLAC · up to 100 MB
I'm wondering why they have restricted file types. You can't check a PDF for example... surely the main use case for people will be to check if a document was produced or edited by an LLM? That could be an attractive (if not misunderstood) proposition for academics
> I'm wondering why they have restricted file types. You can't check a PDF for example...
TFA/page actually seems incomplete. Text uses a completely different watermark format (an actual watermark as opposed to a provenance/authenticity signature), so it makes sense to me that they're not claiming to be able to scan PDFs when they can't yet incorporate that signal.
On a linked page, they say:
> Watermark detection is currently in private preview [...]
Luckily all pdfs I made in academia have been generated from source (tex or derivatives, asciidoc) and llm are better in generating source than pdf. Even people that didn’t use tex used Word to generate the pdf.
So I’m unsure what your aiming at.
Source can have markers inserted in it as weighted word choice or phrase choices, how commas are inserted, etc., so that the output can still can be identified. In other words just because it's source doesn't mean it can be watermarked.
claude can pretty easily generate pdfs and give those to the user to download, skipping the step of needing a source first.
That's a worse workflow imo than having a source w/ version history, but I don't think it's uncommon.
Unfortunately this is just for Media? Some manual tells for Excel or PDFs is to check the author. Claude creates PDFs via wkhtmltopdf so the PDF Producer will be Qt and the Content Creator is wkhtmltopdf. Xlsx files are being created via openpyxl so in the metadata that is the author.
I could do without more surveillance.
It's for your safety!
I think this is primarily done so as not to train models on one's own neuroslope data.
The actual implementation matters less than the fact that they can be perceived to be working on something. You're never going to have a perfect solution here, and perfect is the enemy of good.
I just made a figure with Claude and fed it to the tool, which told me there was no evidence it was made with Claude.
So… this utility appears to be pretty worthless.
The idea of forcing an entity (or entire categories of entity) to add traceable watermarks to all the work they create is... an interesting precedent.
Translating that into the human realm, it would be like if every digital thing you touched were "fingerprinted" with your identity, or at least with a proof of what category of individual you have been categorized as.
Due to the flood of AI, which may soon become indistinguishable from human video, I expect to see something similar. Blue checkmark, but for humans.
The primary functionality is that LLM text is watermarked, but this tool doesn't accept any text formats as input...
> How can I tell if Claude created the text?
> To detect watermarks embedded in text we have a Detection API which is currently in private preview to eligible organizations as required under EU law.
Off topic: Is the grammar off here? Shouldn't it be "by Claude" instead of "with Claude"?
Both are fine. Use “by” if you think Claude do all the work. Use “with” if you consider it a tool.
Tried one of simonwillison's pelicans: We didn’t find any signs this file was processed by Claude!
Like so many things with AI, I like this and hate it at the same time.
Can't they just store checksums of all their output
Probably violates privacy laws.
Is this a trick to get more data to train on?
How would this get more data though? It's more so about being compliant with the EU's new AI laws (which I guess could get them more data since more users)
> How would this get more data though?
It works by uploading files to them?
For anyone wondering who the text watermarking works:
https://www.anthropic.com/news/claude-text-watermark
* Edited the above from initially asking if anyone has any ideas how it works. Found the answer, rewrote the comment.
They basically twiddle with the distribution of used output tokens so that the pattern is recognizable by machine means. Nothing to do with special characters.
droped some source files from different languages and i could process non of them
It clearly tells you what files it can process. This is about media files, not source code.
> Knowing where content came from, and whether AI was involved, makes it easier to trust what you see online.
That’s a cute way to imply their service is used to generate misinformation. They are basically saying to not trust the AI content made from their own product :)
Like they should ban astronaut on a horse images or what? Misinformation becomes misinformation upon the fact of presenting it as so, not the fact of creation
I mean, if we're developing an AGI do you expect it not to be able to generate misinformation?
I have no idea how you’re reading that from my comment
1 reply →
grep “load[ -]bearing” file.txt
A watermark answers "did this model touch the file?" It doesn't answer "did a person read it and stand behind it?"
Detection will get gamed, and a lot of useful text is already mixed. The more useful check, for me, is still human: would I put my name on this before it goes out?
this is going to be the big thing in the next 100 years for me. where there is a human, there will be context, passion and meaning. it's the old saying: machine can tell you what, but only humans can tell you why.
To see the signed metadata: go to claude.ai and ask: "generate an image of a cat". When you'll go to download the image, it will offer an SVG and a PNG version that both have the metadata.
In the case of SVG, it's literally `<metadata><c2pa:manifest>(signature)...` right after the <svg> tag. In the case of PNG, it's stuffed as a `caBX` chunk.
To declaw these files, you can ask in your CLAUDE.md to run `exiftool -jumbf:all=` (for jpg & png), and to simply directly remove the <c2pa:manifest> tag in SVG files. I haven't looked at video files yet.
[dead]
TL;DR: Remember to paraphrase Anthropic models through non-Anthropic models.
Claude watermarks are a farce and a waste of time. It's hilarious to me that they burn cash to even entertain the idea.
Any attempts they use are defeated by a text editor and CTRL SHIFT V. Unicode characters are no new thing.
Reminds me of how people tried to argue that NFTs aren't anything more than just jpegs.
You misunderstand how Claude watermarks things.
TFA is talking about C2PA, a regular signature in the meta data. A lack thereof means nothing, but if it's there, the false positive rate should be near zero.
Most critics of watermarking have no idea how it works. Yet there are valid arguments against (and for) it, but they prefer to be misleading
2 replies →
Yes, but I also think it will be trivially by-passable if you pass your output through another LLM. At least for text.
It might help catch students cheating, but not real spam-bot usage. As soon as platforms start checking for watermarks spambots will add anti-watermark passes.
1 reply →
Imagine telling someone they're wrong without providing any evidence or context.
1 reply →