More than 20 billion pieces of AI-generated content have already been watermarked using Google’s SynthID since the technology was introduced in 2023.
Most people would never notice. That is the point. An AI watermark is designed to leave a machine-detectable signal without visibly changing the content. And now that idea is moving deeper into text.
On August 14, 2026, Anthropic announced how text watermarking will work in Claude, including that future Claude models will generate watermarked text.
That raises some obvious questions in everyone’s mind:
And if a watermark is found, does that prove AI wrote the whole thing?
Not necessarily. A watermark can indicate that a particular AI system was likely involved. It cannot automatically tell you who wrote the original content, how much a human contributed, or whether the information itself is accurate.
That distinction is where AI watermarking starts to make much more sense.
At its simplest, AI watermarking is a way of placing a machine-detectable signal into content produced by an AI system. A compatible detection system can later look for that signal and use it as evidence about where the content came from.
The important word here is evidence.
For text, a watermark is not a visible stamp and, in systems such as Claude’s, it is not a string of hidden characters sitting behind the words.
Anthropic explicitly says nothing is added to its watermarked text and that the signal carries no information identifying a particular user, organization, or conversation.
The same broad idea can also be applied to images, audio, and video, although the technical implementation is different. Google’s SynthID, for example, can embed imperceptible signals directly into generated text, images, audio, and video.
Where people get confused is assuming all provenance technologies are doing the same job. But, they are not.
Suppose an accountant writes an entire client email, then asks Claude to rewrite it more clearly.
A watermark could indicate that Claude was involved in the resulting text. That does not tell you whether Claude came up with the tax recommendation, whether the accountant wrote 90% of the original draft, or whether AI merely improved the wording.
Anthropic describes the limitation directly: its watermark can indicate that Claude was likely involved, but it cannot distinguish between “Claude wrote this” and “Claude heavily edited this.”
That means AI watermarks generally cannot answer questions such as:
Those are authorship, accuracy, ownership, and governance questions. A watermark is primarily a provenance signal.
Technology | What it examines | What it can indicate |
Statistical AI watermark | Patterns created during AI generation | Evidence that a compatible AI system was involved |
C2PA / Content Credentials | Signed metadata attached to a supported file | Information about the file’s source or provenance |
AI detector | Characteristics of finished text or media | A probability that the content resembles AI output |
Hidden-character checker | Characters such as zero-width Unicode | Whether unusual invisible characters are present |
The distinction becomes especially important when evaluating a watermark AI checker online. A tool that scans for invisible Unicode characters is not necessarily reading the statistical watermark used by an AI provider.
Likewise, a conventional AI detector may estimate whether something looks AI-written without detecting a watermark at all.
OpenAI makes a similar distinction in its own provenance documentation.
It uses Content Credentials and SynthID for different supported media because metadata can carry richer provenance information, while an embedded watermark travels within the content itself and may survive some transformations even when metadata does not.
To understand an AI text watermark, forget the idea of a secret message being inserted after the answer is written.
The watermark can be created while the model is choosing the answer itself.
A large language model generates text one token at a time. A token may be a full word, part of a word, punctuation, or another small unit the model uses to process language.
At each step, the model considers many possible next tokens and assigns probabilities to them.
Imagine an accounting firm asks an AI assistant to complete this sentence: “After reviewing the client’s quarterly results, the CPA recommended…”
Several continuations could make sense. “Adjusting,” “revising,” or “reviewing” might all be reasonable depending on the context.
Normally, the model chooses from those possible continuations according to its generation process. A watermarking system adds another small influence to those choices.
Google’s SynthID-Text works by adjusting the token-selection process during generation. The change is designed to be subtle enough that the output still reads normally, while creating a statistical pattern that a compatible detector can recognize later.
One choice tells you almost nothing. Hundreds of choices can.
That distinction matters because there is no permanent dictionary of “AI words.” A system does not simply decide that one synonym is always watermarked while another is safe.
The preceding text affects the generation process, so the same word can appear perfectly naturally in both human and AI writing.
Anthropic gives a useful example. In a sentence such as “The weather today was cold and…,” words such as “overcast” or “grey” might both work. Watermarking can use many low-stakes decisions like these across a longer response to build a pattern.
This is also why longer passages are generally more useful for statistical detection than a sentence or two. A detector needs enough choices to determine whether the expected pattern is showing up more often than chance would suggest.
The “watermarking key” is another part of this topic that is easy to misread.
It is not a secret customer ID being hidden in your article.
The key helps the watermarking system generate the pseudorandom rules that influence token selection. When the corresponding detection system later evaluates a piece of text, it can test whether the pattern of choices is consistent with the watermarking scheme.
This is different from encryption.
It is also different from the cryptographic signatures used in systems such as C2PA Content Credentials, where signed metadata can help verify information about a file’s provenance.
That distinction is worth keeping clear: the statistical signal lives in the generated language, while provenance metadata can travel alongside a file.
There is no single AI watermarking standard that every model provider currently applies in the same way.
As of August 31, 2026, Google, Anthropic, and OpenAI are at different stages of deploying provenance signals across their products.
That distinction matters because a statement such as “AI text is now watermarked” is far too broad.
Provider | Current text watermark status | Public text verification | Other provenance signals |
Google / Gemini | SynthID-Text is used for text generated through the Gemini app and web experience | Google’s current consumer verification features focus on supported image, video, and audio content rather than arbitrary pasted text | SynthID is also used across supported images, video, and audio |
Anthropic / Claude | Anthropic says future Claude models will include text watermarking, while pre-August 2 models are being transitioned over time | Anthropic has announced detection tooling, but a general public text detector is not yet available | Supported generated files can also use provenance metadata |
OpenAI / ChatGPT | OpenAI does not currently document a deployed watermark for ordinary text outputs | OpenAI Verify currently supports its documented image and audio provenance signals, not text | Supported images use C2PA and SynthID; supported audio uses SynthID |
Google has the longest production history of the three in text watermarking. It began bringing SynthID-Text into Gemini in 2024 and now describes the technology as part of its broader watermarking system for generated text, images, audio, and video.
Anthropic’s August 2026 announcement is what pushed AI watermarks back into wider discussion. The company says future Claude models will generate watermarked text and that older models launched before the EU requirement took effect have a transition period while Anthropic adds support.
Its implementation contains no hidden characters and is based on the same general family of techniques as Google’s SynthID-Text.
OpenAI is in a different position for text. Its current provenance documentation covers supported images and audio.
OpenAI says its goal is to expand provenance signals to text as standards and tooling mature, but it does not currently claim that normal ChatGPT text responses contain a deployed text watermark.
That gives us an important answer to one of the most common questions around this topic: Does ChatGPT watermark text today?
Based on OpenAI’s published documentation as of August 31, 2026, not through a currently documented text-watermarking system.
That may change. When it does, the exact implementation will matter much more than a headline saying that “all AI text can now be detected.”
There is no universal AI watermark checker that can identify text from every AI model.
Watermark detection is tied to the system that created the signal. Anthropic, for example, says its detector is designed to estimate whether Claude was involved in a piece of text.
A third-party AI detector does not automatically have access to that same watermarking system.
This matters when websites advertise a “Claude watermark checker.” Some tools are actually scanning for hidden Unicode characters, metadata, or general signs of AI writing. That is different from detecting Claude’s statistical watermark.
Anthropic has announced a detection API for its own watermark, but says the tool is still forthcoming.
Google has open-sourced SynthID-Text, but that does not create a universal Gemini detector. Its technical documentation explains that detectors are associated with specific watermark configurations and keys.
So, Google is not the only organization that can use AI watermark detection, but neither can any random checker reliably detect every provider’s watermark.
Copying AI-generated text into Word, Google Docs, or a CMS does not automatically remove a statistical watermark because the signal comes from patterns in the generated language, not hidden characters.
Editing is different.
Anthropic says light editing may leave enough of the original pattern to remain detectable, while extensive rewriting can weaken or remove that signal.
Translation also depends on the workflow. Translating an existing watermarked passage through another process can disrupt the original pattern. But if Claude itself performs the translation, the new output can receive Claude’s watermark because Claude is selecting the translated words.
A sentence such as: “Form 1120 is the U.S. Corporation Income Tax Return.” Gives an AI model far fewer reasonable wording choices than a 1,500-word tax article.
That matters because statistical AI watermarks need enough language choices to form a detectable pattern. Anthropic says short samples and highly factual passages can therefore provide weaker evidence.
Watermarking existed well before 2026. Google introduced SynthID for images in 2023 and expanded SynthID-Text into Gemini in 2024. What changed was regulation.
Article 50 of the EU AI Act requires providers of systems generating synthetic text, audio, images, or video to make those outputs machine-readable and detectable as artificially generated or manipulated, where technically feasible.
The regulation does not require every company to use the same watermarking technology.
Anthropic says it is initially applying Claude’s watermarking globally rather than limiting it to Europe.
That is why recent watermarking AI news matters even to U.S. businesses.
For accounting firms, watermarking should be treated as one more part of AI governance, not as a test of whether someone “cheated” by using AI.
That distinction matters as firms increase their use of AI. In the AICPA’s 2026 CPA Firm Top Issues Survey, which included 629 CPA firms, managing change caused by technology and AI ranked as the leading expected five-year issue across firm-size groups.
There is currently no published Google Search rule saying an AI watermark causes a ranking penalty.
Google instead focuses on content quality and warns against producing large amounts of low-value content primarily to manipulate search rankings.
For a CPA firm, that means the important questions remain whether the tax or accounting information is accurate, useful, reviewed by someone qualified, and genuinely adds something for the reader.
A detected watermark can indicate AI involvement, but not exactly how AI was used.
Someone may have generated a document from scratch, rewritten an existing draft, translated it, or simply asked AI for substantial editing. Anthropic specifically says its watermark cannot distinguish between Claude writing something and Claude heavily editing it.
For accounting firms, the better controls are therefore straightforward:
An AI watermark can help with provenance. It cannot replace professional judgment.
AI watermarking gives us another way to understand where digital content may have come from, but provenance is only one part of the picture.
For businesses publishing online, the bigger priority is still the value of the content itself.
Google’s current guidance for both traditional Search and generative AI experiences emphasizes content that is useful, original, reliable, and created for people rather than pages produced simply to capture search traffic.
That is where experienced SEO, AEO, and GEO content strategy matters. Getting found in search or picked up by AI Overviews still comes down to one thing: whether the content is genuinely useful.
If it answers the question clearly, brings in real expertise, and gives readers something they could not get from ten similar pages, it has a much better chance of earning trust and visibility.
At Credfino, we have helped accounting firms strengthen their online presence through content built around that approach. If you want to make your content a real growth asset rather than simply another page on your website, our experts would be happy to discuss the roadmap with you.
In most cases, you probably will not notice a difference. The watermark is designed to sit quietly in the model’s word choices without making the writing feel awkward or unnatural.
It can. A detection result is better read as a signal that AI may have been involved, not as proof that the entire piece was written by AI.
Not always. A detector can only identify what its watermarking system was designed to recognize, and some providers may only confirm that their AI was involved rather than naming the exact model version.
Anthropic says Claude’s text watermark does not add tokens, increase response costs, or create a practical slowdown. The watermark is created during token selection rather than added as extra content afterward.
No. A watermark may show that AI was involved, but it cannot tell you whether that use violated a firm policy, exposed client data, or produced inaccurate work. Those questions still require human review and clear AI governance.
Learn how accounting firms can use SEO to increase their chances of being cited by ChatGPT, improve AI visibility, and attract more qualified clients.
Learn how local SEO services for accountants can improve Google Maps visibility, attract nearby clients, increase local rankings, and generate more qualified leads for your accounting firm.
Learn how to find the best SEO keywords for accounting firms, including high-intent, local, and service-based keywords that can improve rankings, attract qualified leads, and grow your practice.