🚨 JUST IN: Claude models will now have invisible watermarks embedded in ALL text, and ALL metadata attached to files…
Horror, scare what next! The reaction on social media, in forums and traditional media has been quite extreme. Headlines like "EU-enforced update leads to fears that people will be accused of using artificial intelligence with their own writing" or "Claude models will now have invisible watermarks embedded in ALL text, and ALL metadata attached to files." were the result of Anthropic's announcement that their models would comply with the EU AI Act (Regulation 2024/1689), enflamed further by saying it would apply irrespective of where the user was. Meaning whether or not you used it in a jurisdiction that fell under EU AI Act (Regulation 2024/1689).
But is this really something new?
Have you never given AI something and asked "Is this written by AI?"
How did you think the AI was coming up with conclusions? Or when Linkedin started labelling images on its platform that were generated with AI, how did we think it was determining that?
On images it is easier to understand how this is being watermarked. It can be embedded into meta data or even hidden within the pixels themselves. Most model companies directly embed metadata for 3rd parties like social media platforms to detect and prevent AI-generated misleading content.
Google have published their methodology, which uses their SynthID
"SynthID is our new watermarking tool, designed specifically for AI-generated content. It empowers users to identify AI-generated (or altered) content, helping to foster transparency and trust in generative AI."
Anthropic have published the same for Claude, which you can read by clicking here.
Should we as users have an issue with it? Overall I believe not, since with the exponential use of AI-generated content, transparency regarding its source provides valuable context for consumers navigating digital information. The real risk is nefarious players that will not adopt those as I expected the non mainstream AI models will not do.
But at its heart this is to help protect us as consumers of this content.
With respect to written text, how does the watermarking occur? My research into this reveals many different ways, and here are some examples:
1. Basic formatting, associated with earlier ways and easily defeated. Some web-copied outputs or specific AI implementations inject invisible Unicode characters—such as Zero-Width Spaces (U+200B), Zero-Width Joiners (U+200D), or Non-Breaking Spaces (U+202F) into paragraphs.
2. Statistical watermarks alter the probability distribution of words during generation, preferring specific "Green List" token sequences based on previous words. Because these watermarks live in the statistical cadence rather than explicit words, light word-swapping is rarely enough to erase them. To a human reading a sentence or two, the writing looks completely natural. However, over a longer passage (e.g., 200–300 words), an unusually high percentage of words will land on the "Green" list. In plain English for you and I to understand AI text watermarking works by subtly altering the mathematical process an AI uses to pick its words, embedding a secret statistical pattern into the generated text without changing its meaning or readability to a human.
There does seem to be a large concern about using AI to edit written content. I have seen lots written about people worried about academic submissions being identified as AI written. But is this different from tools like Turnitin were being used by Universities. I recall from my MBA, we were told that getting a Turnitin score of zero meant we had not referenced items and should strive for a match score of around 20% as long as we had used the Harvard referencing format correctly.
So perhaps there will be a tolerance factor built in the analysis. I cannot image a world where institutions are not encouraging (responsible) use of AI tools to prepare students for a real world future.
I gave Gemini a task of checking one of my articles if it had been written by AI.
Gemini determined that based on the statistical cadence, structure, and phrasing, this text reads as predominantly human-written, though it contains a few structural markers common in polished AI outputs.
Of course I used AI, I am not a trained journalist, nor do I have a sub editor on my personal payroll.
So I pay for an AI subscription, currently Gemini, to become my digital assistant across many different work flows. In this case it took on the persona of my sub editor and performed tasks for me, such as reviewing the raw copy for grammar spelling, logical flow and structure, writing attention grabbing headlines, and ensuring that the articles fit the required space and style before publication.
I also use AI for structuring visual layout of text, for instance into comparison tables and this does then get detected later as AI. Perhaps a false positive in that case?
Here are some tangible examples that I use almost daily: Meeting Summaries, Image Generation and HTML generation.
Meeting Summaries
I took a recent meeting summary that I generated with Claude and asked AI for a view if the content was AI generated. What you should know is the report is output into a standardised format that I have defined in a skill.
What did AI tell me about the document?
"Looking at this document, yes — it has several strong markers of AI-generated text"
Here is verbatim what AI gave me as its view:
Structural tells:
- Perfectly uniform formatting throughout — every section has identical structure, bullet density, and action register entries with flawless consistency
- Reference IDs are auto-generated in sequence (260809-MTLK-01 through -20) with zero gaps or anomalies
- Every action has an owner, due date, and item reference — real meeting notes rarely achieve this level of completeness
Language tells:
- Phrases like "noting that books for July had not yet been closed", "reaffirmed as the single source of truth", and "prompting discussion around potential data quality issues" are classic AI summary prose — precise, passive, slightly formal
- The executive summary bullet points are unusually balanced in length and scope
- Transitions like "The session covered several strategic investment areas" are textbook AI paragraph openers
Content tells:
- No hedging, no partial information, no "unclear" items — AI tends to resolve ambiguity rather than flag it
- Every discussion point lands cleanly with an outcome or action; real meetings have loose ends
- The tone is entirely neutral — no personality bleed-through from any individual speaker
Now am I bothered by this? By no means, and here are the reasons:
Perfectly uniform formatting throughout
I actually want a perfect formatted document, for consistency and professional appearance.
Reference IDs are auto-generated in sequence
This is actually a core part of the skill I built that every item needs to have a unique reference number to allow for later tracking within in a master tracker. My skill even defines how to come up the reference number!
"YYMMDD-ACRN-NN — meeting date + 4-letter acronym (from Step 2) + sequential 2-digit number incrementing across the whole document (01, 02, 03…)"
Executive Summary
Again this is performing the explicit ask in the skill itself.
"3–5 bullet points covering: purpose, key outcome/decision, and dominant programme sentiment. Corporate tone: purposeful, concise, direct. Each bullet is a complete thought"
Content Tells
Well no surprise since I am within the skill giving it direction to do just that.
"Identify all distinct items discussed; group related exchanges under one heading.
For each item produce:
- Item heading — 3–6 words
- Discussion summary — 2–5 sentences, third person past tense, synthesised not transcribed
- Actions / Outcomes — bulleted list of commitments/decisions from this item, each ending with its Reference ID in parentheses (e.g.
(260716-KLPJ-01)). Omit if no actions arose."
Whilst my document is riddled with the hallmarks of AI generation, that is exactly the point. To generate meeting minutes and trackable action items now takes me a few minutes post a meeting to generate a comprehensive, corporate branding and professional document.
Is AI not doing what I need here? By acting as a new tool in my toolbox I am saving time and increasing not only the quantity but quality of output?
Is this any different from someone having sat in a meeting scribing and then generating a document? Would that human document not have the hallmarks of the scribe that produced it? But importantly could they do it with such speed, accuracy and consistency each time? And of course my skill can be shared across teams or even the whole organisation for them to achieve uniform output.
That's the text side of the equation, where the watermark hides in cadence and structure. Images work differently — the tell is right there in the metadata.
What about image generation?
So now that AI is watermarking images do I care? No, since I am not trying to pass off its work as my own. Do I feel the need to declare I used AI for image generation? Not at all.
Historically you would have either just copied an image from somewhere (copyright issues right there), cobbled an image together yourself using basic tools if you were not proficient to use Adobe Photoshop (poor or amateur results most of the time) or paid someone to do the image at your instruction as to what you required (cost and time).
So lets think about the last option, in your personal life perhaps the cost of paying someone vs the utility ROI was low, or in a corporate setting you were not given budget or when you were it was for something specific like a huge pitch you were doing and the company agreed to invest into a designer to help.
But when you did engage a designer two things stand out to me, firstly you did not feel the need to tell anyone that you used an external party to generate images and importantly you needed to explain to the person what type of output you were looking for, so they know the direction and desired outcomes. Today that instruction, we just call a prompt.
In my mind explaining to a human what I need vs prompting an AI are exactly the same. And whether I am then paying a person or subscribing to an AI, is the output not then my paid for content?
Here is an example of how I generated the header image for this exact article, I pay for a Gemini Pro subscription and then gave it the below prompt.
Prompt "Generate me a header image to use on my Blogger blog. The article is going to be about the use of AI to denote different outputs. Those outputs could be: generating an image, doing a grammar check, formatting text, using a repeatable skill to generate corporate documents such as a PowerPoint pitch deck or meeting minutes, and then even using AI to try to generate original written content. The article will be discussing the positives and negatives of this now that AI watermarks content."
Running the same prompt a few times and occasionally giving further clarity within a thread I had a section of images I could select from. You know like a graphic designer would have done previously. Except here since I have the subscription, I did not need to first quantify cost, wait for signoff of the cost and then sit in work queue of the designer. I got instant output. Below of some of the iterations I generated over the course of 10 minutes to select my final image.
If a generated image can be my paid-for content, what about an entire template? That's the biggest watermark test of all — not a single asset, but the whole site.
HTML editing
It has been a really long time since I have edited HTML. I can remember when I did initially using notepad, progressing onto WYSIWYG tools like Microsoft Frontpage and others. I did make the effort to learn Macromedia Flash, but by the time CSS and Web 2.0 rolled around I was in a different phase of my career and no longer had the time to keep up to date. When I started this site for the first few weeks I used the built in templates available on the blogger platform, but then purchased a commercial template to give it a distinctive look and feel.
In later years I engaged a blogger template designer I found on Fiverr and paid for customisation. For that I selected a base template and then requested targeted changes to it. After a few iterations, I was presented a final template for my use. And I guess there was nothing to stop that designer reusing it with his other clients. Over the years I have tinkered with the XML template by making HTML changes in Notepad, this normally required me Google searching what I was trying to achieve, make the change, test it in a separate blog and after fine tuning it use it as my actual site template.
Fast forwards to this year.... I wanted to reinvigorate the look and feel of my site. I started with enhancing the headline images using AI tools, this was to replace prior graphics which in the whole were screenshots of items I had compiled using Powerpoint from clipart and stock images. This immediately elevated the look. But the overall look was stale. I started with a generic but free blogger compatible template and fed that into AI with my prompts of changes that I required. Naturally I still tested it in a stand alone blog before deploying to production. But I was able to make extensive layout, styling, look and feel changes without coding a bit of HTML myself, AI took care of that.
That was an exercise at a point in time, ironically at the same time Google had a false positive on Malware within Blogger and I thought my changes had triggered the pull down of my site, thankfully it had not - but you can read more about that clicking here.
But now similar to the Meeting skill I have, I have built a Gem (Google Skill) to take my draft article and redo it in HTML applying my styling changes to it. I did need to add an explicit instruction within the Gem of:
"TEXT PRESERVATION: You must preserve ALL original input text, punctuation, links, and paragraph structure 100% VERBATIM. NEVER summarize, condense, rewrite, omit, or create new text."
And this was based on prior attempts where I had asked Gemini to apply my styling but I noted it had indeed arbitrarily made content changes on its own.
But here is a test, here is a screenshot of this article before applying the styling Gem to. And what you are reading is a consequence of my formatting Gem (skill).
So watermarking isn't the plot twist the headlines want it to be. I'm not hiding my use of AI. It is simply another tool, like my laptop or O365 subscription residing in my toolbox, and I don't think I should have to worry that I am using it. What matters isn't whether a Green-List token pattern shows up in my sentences or a tag sits in my image metadata, it's whether the output does the job I needed. A designer never had to declare they used a mouse. A sub-editor never had to disclose their spellchecker. The prompt is just the brief, and the brief has always been mine - and I am paying someone as my agent to perform the task. If anything, the AI Act crowd should relax, transparency about origins of content protects the honest use, not the lazy one.
The Red Herring is commentators bleating about the EU AI Act and why should it apply to them, when the real issue is if you are using it correctly should you be worried?
My final comment about the watermarking is that transparency doesn’t negate utility.
