According to iThome, Anthropic announced its AI-generated content labeling plan: new Claude models launched in the EU starting August 2, 2026 will support machine-readable watermarks from day one (E1).
The text watermark itself is invisible to the human eye. For supported file types such as SVG, PNG, and JPEG, Claude also attaches C2PA-standard, digitally signed provenance metadata to record content origin (E2).
Per iThome, any text produced through a watermark-supporting Claude model carries the marking regardless of entry point — Claude, Claude Platform API, Claude Code, Claude Cowork, and Claude Tag are all listed as covered (E6).
Inside.com.tw's technical breakdown describes the underlying mechanism: the watermark is embedded by slightly adjusting the probability distribution the model uses when choosing words, creating a statistically identifiable signal. The signal is subtle enough to not affect reading quality — but that same subtlety makes it fragile. Running the text through another AI for rewriting, or substantial human editing, can dilute or erase the signal (E15).
Inside.com.tw states this is the first time a major model maker has simultaneously covered two layers of content labeling — text and file — with the rollout timed to land on the effective date of EU AI Act Article 50 (E10).
That dual structure matches what iThome described: an invisible text watermark for generated text, paired with C2PA-signed metadata for supported image and file formats like SVG, PNG, and JPEG (E2).
Anthropic told iThome that the watermark doesn't display on-screen and doesn't affect a text's meaning or readability; when text is copied and pasted elsewhere, the marking may still persist (E7).
C2PA file signatures behave differently, according to Inside.com.tw: cryptography guarantees the metadata hasn't been tampered with, but the metadata itself can be stripped — downloading an image and re-saving it is enough to erase the signature (E16).
iThome reports that neither the presence nor the absence of a marking can fully confirm whether content is AI-generated: heavy rewriting, translation, very short passages, file format conversion, re-saving, or even screenshots can strip or corrupt the source information needed for detection (E8).
That limitation traces back to how the text watermark works. As Inside.com.tw explains, the embedded signal is deliberately subtle so it doesn't degrade readability — which is also why rewriting or substantial edits can dilute or erase it (E15).
| System | Watermark Type | Scale / Status | Date |
|---|---|---|---|
| Google SynthID | Image + Text | 100+ billion images marked; text watermark deployed to Gemini API | October 2024 |
| OpenAI | Image only (via SynthID) | Adopted SynthID for ChatGPT images; text watermark not yet launched | Announced in May |
| Claude (Anthropic) | Text + C2PA file | New EU-released models watermark from first release | August 2, 2026 |
Per Inside.com.tw, Google's SynthID has already marked more than 100 billion images, and its text watermark was deployed to the Gemini API back in October 2024 — currently, per the same report, the largest-scale text-marking system in operation (E12).
OpenAI announced in May that it would adopt SynthID to watermark ChatGPT-generated images, but a text watermark for OpenAI's own outputs has still not launched, according to Inside.com.tw (E13).
Academic research cited by Inside.com.tw found that removal tools can strip Google SynthID's image watermark with a 79% success rate, a data point the outlet uses to illustrate that even the most widely deployed marking system isn't tamper-proof (E14).
Anthropic's rollout begins with new Claude models released in the EU from August 2, 2026 (E1). But iThome reports the mechanism isn't confined to the EU — it will also apply to every Claude service that supported models offer globally (E3).
iThome reports Anthropic has signed the Code of Practice on transparency for AI-generated content tied to Article 50 of the EU AI Act (E4).
The Act requires AI systems that generate text, images, audio, or video to mark AI-generated or modified content in a machine-readable way, and iThome confirms these transparency rules took effect starting August 2, 2026 (E5).
Inside.com.tw reports that non-compliance carries penalties of up to €15 million or 3% of global annual revenue, whichever is higher (E11).
iThome reports Anthropic is still working to add watermark support to Claude models released before August 2, 2026, and has not yet announced a completion date (E9).
The technical details of the text-watermarking method and its detection process also remain undisclosed; Anthropic told iThome it will publish technical documentation later (E9).
By contrast, watermarking already applies across Claude, Claude Platform API, Claude Code, Claude Cowork, and Claude Tag for models that support it from launch (E6) — meaning coverage today depends entirely on which model version a user is on.
Taken together, the evidence points to a gap between rollout and reliability. Anthropic's marking begins with new EU-released models on August 2, 2026 and extends globally (E1, E3), landing the same day the EU AI Act's Article 50 transparency rules — and their penalties of up to €15 million or 3% of global revenue — took effect (E5, E11). Yet by Anthropic's own account to iThome and Inside.com.tw's technical analysis, the underlying signal is fragile: rewriting, translation, short passages, file conversion, or a simple re-save can strip both the text watermark and the C2PA metadata (E8, E15, E16). Google's SynthID offers a scale comparison — over 100 billion images marked since October 2024 — but even that system saw a 79% watermark-removal success rate in academic testing (E12, E14), while OpenAI has adopted image marking without yet shipping a text watermark of its own (E13). The result is a compliance deadline met with a labeling layer that, by its own maker's account, cannot on its own definitively confirm whether content is AI-generated (E8).
Not necessarily. Anthropic told iThome the watermark doesn't display on-screen or affect readability, and the marking may still persist after text is copied and pasted elsewhere (E7).
Yes. Inside.com.tw reports that downloading and re-saving an image is enough to erase its C2PA signature, even though the metadata itself is cryptographically tamper-evident while intact (E16).
Up to €15 million or 3% of global annual revenue, whichever is higher, according to Inside.com.tw (E11).
Not yet. Inside.com.tw reports OpenAI adopted SynthID for image watermarking in May but has not launched a text watermark (E13).
According to a report from money.udn.com, TSMC's board approved a Q2 2026 cash dividend of NT$7 per share, unchanged from the prior quarter, and for the first time will let foreign shareholders elect to receive payouts in US dollars under revised regulations, cnyes.com reported.
According to Technews and CNA, France's APIG press alliance — representing more than 300 outlets — filed a complaint on August 11 over Google's AI Overviews, launched in France in July 2026. APIG cites 70% traffic dependency on Google, a Pew Research Center finding that AI summaries cut click-through rates from 15% to 8%, and a Le Figaro executive's report of 20%-50%-plus traffic declines abroad, while asking regulators to help reset negotiating terms rather than shut the tool down.
According to Cnyes News, Supermicro (SMCI) will report fiscal Q4 2026 results after the U.S. market close on Tuesday, August 11. The company raised its own gross margin guidance to 15%-17%, up from a prior 8.2%-8.4% range, backed by more than $60 billion in new orders. Bloomberg consensus projects EPS of $1.33 and revenue of $11.2 billion, versus $0.41 and $5.7 billion a year earlier.