Farmed AI content has been a problem for years now. AI text watermarking is about to change all that. Largely thanks to the EU’s new AI-generated content policies and additional mandates from India, California, China and beyond, all text produced by online content writers using Claude, Gemini and other LLMs is about to become more transparent. Anthropic, Google and other AI powerhouses are implementing invisible AI watermarks. In this blog, I’ll show you what this means in practice.
A Long-Awaited Change in the Rules
Few real writers or website operators will disagree that AI-generated content needs regulation. Enter the EU AI Act’s Code of Practice on Transparency of AI-Generated Content, otherwise known as Article 50(2).
Elsewhere, California has passed SB 942, requiring AI providers to make invisible disclosures about all AI-generated content, and China has strict watermarking tools running under the Cyberspace Administration of China (CAC) rules.
While not issued worldwide, these new legal frameworks effectively make it mandatory to be able to identify AI-generated content everywhere in the world immediately. Recently, Anthropic revealed how its AI text watermarking procedure will work, and it isn’t the only company engaged in it. Allow me to explain the basics.
How Will Anthropic Implement Watermarks in Claude?
Anthropic has opted to pinch Google DeepMind’s SynthID-Text technique for watermarking AI-generated content. In practice, this is what you can expect:
- Claude makes choices to create statistical word patterns when several words can be used identically in a sentence, based on which words that precede or follow another. This is a concept known as Low-Stakes Choices. While invisible to readers, the pattern will be identifiable to anyone with a key or the soon-to-be-released API.
- This same process is then repeated many times throughout an entire generated text. It leaves behind a trail or pattern, in effect, an invisible watermark.
- Those with keys or access to the API can then scan documents to check if the word choice used is similar to decisions that Claude would make using its Low-Stakes Choice pattern.
Crucially, this procedure doesn’t impact the quality of an AI-generated article. Nor is it visible to readers. In fact, there are no hidden characters, nothing added to the text, and no discernible differences between watermarked and unwatermarked content.
Anthropic recently provided an example of how the process works in their blog on the subject:
- “The weather today was cold and…”
Anthropic is obviously correct in stating that the next word in the sequence is unlikely to be “sugary”. It is more likely to be “overcast” or “grey”. Ultimately, Claude will pick a word, both of which are suitable. The way it chooses these words is what forms the basis for the watermark.
Editing Documents to Remove Watermarks: Is It Possible?
Not everyone uses Claude or LLMs to generate text. At WritingKosmos, we rely on human writers. However, some may use AI to edit text. I wanted to know: how does the new AI text-generating watermark apply to them?
In theory, uploading a human document and asking AI to edit it might lead to a watermark if substantial edits are made to the text. However, if just a few edits are made, even if a watermark is practically there, it may not be visible enough to flag attention when checked.
So, what about editing AI-generated content to remove a watermark? We’ve only got Claude’s explanation to go on here. Light editing of AI-generated text isn’t likely to remove the watermark, but going all-out and editing every word very much will.
Anthropic believes this is fine, as the content would then be edited enough to be deemed practically human at that point. Of course, that doesn’t necessarily mean it will escape other AI checkers, such as Originality.AI, which looks for different patterns and tells.
Spotting Claude-Written Content: Weaknesses Explained
Having access to an API will make life a lot simpler for webmasters and genuine writers. While AI checkers are currently notoriously unreliable and often taken (erroneously) as gospel truth, the new AI text watermarking systems seem more reliable. There are flaws already emerging.
- Firstly, Claude’s watermark check can’t guarantee that content is human, or that it hasn’t been written by another AI tool. They may use different patterns.
- Secondly, an open-source tool is available on GitHub today that actively strips away watermarks and tracking markers. Called Anthropies, it is an open-source project from Cardano’s founder, Charles Hoskinson. There is no guarantee of success, though.
- Watermarking doesn’t work well on small pieces, as the pool of words AI can choose from is limited. The same is true for predominantly factual content, where specific terms and data must be used.
On the upside, the watermark is not physical. Anyone thinking of opening a new Word document and typing out the AI output word for word in a bid to remove the watermark will be disappointed. The watermark is based on the order of words, so it doesn’t matter how the text is entered into a file. It doesn’t have to be copy-and-pasted to flag.
How Watermarking Affects Images and Code
While code has, for now, been largely excluded from watermarking, content that features imagery, including screenshots, will be subject to watermarking. Each AI-generated image will carry cryptographic signatures in its metadata, in line with other C2PA standards commonly found in photo-editing software. Any AI-generated screenshots generated by AI for iGaming content will be visible.
Support and Criticism from the Community
The emergence of AI text watermarking has caused a division along commonly established battle lines in the iGaming content community. Firstly, genuine writers and webmasters are praising it. Their argument is straightforward:
- If you’re not using AI to generate content, what do you have to fear?
Their argument is that only those known to farm content and produce budget articles through AI should have anything to worry about. Those on the opposite side of the fence argue that AI is a valuable tool and have threatened to close their Claude accounts.
Other AI Models Are Following Suit
Anthropic’s Claude is not the only show in town producing AI watermarks. In a bid to be compliant with new rules and regulations, others have followed suit, but not everybody is willing to play ball:
- Google Gemini already uses its SynthID text in Gemini, across text, images and audio.
- OpenAI’s ChatGPT has not implemented text watermarking, and there’s no surprise there. The controversial AI company is afraid of losing users at a time when it’s burning through cash and its dominance has been challenged and lost.
- Meta AI takes a similar line to ChatGPT on this stance, although it does (like ChatGPT, to be fair) implement watermarks on images and audio.
- Grok’s xAI is arguably the worst offender, shunning most watermarks, and when they do appear, they are limited to editable metadata, which can be stripped with no additional software.
Better, More Original Content for Everyone
What does this mean for AI writing? AI text watermarking is already active and has been since August 2 across Gemini and Claude. While it cannot guarantee originality or prove farmed content any more than current AI checkers, and while Anthropies cannot guarantee removal of watermarks, I still view this as a step in the right direction for the casino writing space, in particular, which is often flooded with farmed content..
Current AI checkers are often far off the mark. AI text watermarking looks as though it will separate true writers from content spinners and spammers. It could lead to a place where webmasters hire human writers once more, which is great for our iGaming industry.
While AI text watermarking won’t get rid of poor, farmed content, it may give those who produce it food for thought about continuing. Furthermore, if used in conjunction with harsher Google penalties, this could be the turning point the iGaming content industry has been looking for.