New Claude Models Will Embed Invisible Watermarks in Generated Text

cryptonews.ruPublished on 2026-08-11Last updated on 2026-08-11

Abstract

Anthropic has announced that starting August 2, 2026, new Claude models deployed in the EU will incorporate machine-readable, invisible watermarks into all AI-generated text. This watermark will be woven directly into the model's output, persisting through copy-paste and some editing, without altering readability or meaning. The rule will apply globally across all Claude platforms, including API, apps, and partner services like AWS and Google Cloud. For generated files (e.g., PNG, JPG), a separate, existing system using C2PA metadata signatures will continue to indicate Claude's involvement. Models released before August 2, 2026, including the current Claude Opus 5, are in a transition period and do not yet embed text watermarks, though Anthropic is working to add support. Currently, no public tool exists to detect these text watermarks; Anthropic says it is developing separate detection tools and will release technical details later. The company cautions that even with detection, a watermark does not conclusively prove Claude authored the text, as it is often used to edit or summarize existing content. The timing aligns with the EU AI Act's Article 50, which from August 2, 2026, mandates clear labeling of AI-generated content, with potential fines for non-compliance. A key technical question remains the watermark's resilience to paraphrasing or translation by other AI models, a challenge noted in previous industry experiments.

Anthropic has revealed a previously unpublished detail: starting August 2, 2026, new Claude models launched in the European Union will support machine-readable watermarking of generated text from their release. This refers to future releases—no model has been released after that date yet, the latest currently being Claude Opus 5, released on July 24. This is stated in an article in the Help Center, updated on August 10, 2026. The rule will apply not only in the EU, but wherever Claude is offered—in the API, the Claude app, Claude Code, Claude Cowork, Claude Tag, as well as with partners like AWS, Google Cloud, and Microsoft Foundry.

Text and Files—Two Different Mechanisms

For text, Anthropic will use invisible watermarks that will be woven directly into the model's response itself, without changing its meaning, quality, or readability. Since the mark will become part of the text itself, it will be preserved during copy-paste and, partially, even after editing. This mechanism has no relation to the C2PA standard.

For files—for example, .svg, .png, and .jpg—a different, already operational mechanism applies: Claude attaches signed provenance metadata using the open C2PA (Coalition for Content Provenance and Authenticity) standard. Such a signature indicates that the file was processed by Claude and allows for the detection of tampering with the metadata after the fact. Support for this feature depends on the specific platform.

Old Models Without Text Watermarks

The rule regarding text concerns only new versions of Claude. Models released before August 2, 2026—meaning all currently available models, including Claude Opus 5—are in a transitional period and do not embed text watermarks. Anthropic says it is working on adding support for watermarking for them as well, and the documentation will be updated as readiness is achieved.

Cannot Check Text Yet

Anthropic states that it is working on separate tools that will allow users and third parties to detect text watermarks. The company promises to reveal details of the detection mechanism in technical documentation later—no such tool exists at the moment. Sites like c2pa.org or contentcredentials.org are not suitable for this: they check metadata in files, not text watermarks.

An important nuance: even when detection becomes available, the presence of a mark will not be definitive proof that the author of the text is specifically Claude, since the model is often used for editing, translating, or summarizing others' content. The absence of a mark will also prove nothing.

AI Opinion

From the perspective of macroeconomic regulation, the coincidence of dates does not appear accidental. The European Union, starting August 2, 2026, introduced mandatory requirements of Article 50 of the AI Act regarding the labeling of AI content, and the European Commission had already approved the final clarifications for this norm on July 20, providing for fines of up to 15 million euros or 3% of the company's global turnover for violating transparency. The launch of Claude's text watermark precisely on this day is, apparently, not a technical coincidence but a compliance reaction to a specific regulatory deadline.

The technical aspect that the article does not reveal is the resilience of such watermarks to paraphrasing and translation via another model. Industry experiments with text marks have shown that paraphrasing noticeably reduces detection accuracy, and translating text into another language often removes the mark completely. Whether Claude's invisible mark will remain detectable after processing by a third-party AI is a question whose answer only the company itself currently knows.

Related Questions

QWhat is the new requirement for Claude models launched in the European Union starting August 2, 2026?

AStarting August 2, 2026, new Claude models launched in the European Union will be required to embed machine-readable watermarks into generated text at the time of release. This rule applies wherever Claude is offered, including API, the Claude app, and through partners like AWS and Google Cloud.

QHow does Anthropic's text watermarking for Claude differ from its file watermarking mechanism?

AFor text, Anthropic will use invisible watermarks woven directly into the model's response, which persist through copy-pasting and partially through editing. For files like .png and .jpg, Claude attaches signed metadata of origin using the open C2PA standard, which indicates Claude processed the file and can detect subsequent tampering.

QWill Claude models released before August 2, 2026, have text watermarking?

ANo, Claude models released before August 2, 2026, including the current Claude Opus 5, are in a transitional period and do not embed text watermarks. Anthropic is working to add support for them later.

QCan users currently detect text watermarks in Claude's output?

ANo, users currently cannot detect text watermarks. Anthropic states it is developing separate tools for users and third parties to detect these watermarks, but such tools do not exist yet. Websites like c2pa.org are not suitable as they check file metadata, not text watermarks.

QWhat is the likely reason for Claude's text watermark launch coinciding with August 2, 2026, according to the article's analysis?

AThe launch date likely aligns with compliance to the EU's AI Act. Starting August 2, 2026, the EU enforces mandatory AI content labeling requirements under Article 50, with potential fines for non-compliance. The timing appears to be a regulatory compliance reaction rather than a technical coincidence.

Related Reads

History Repeating? Bitcoin Returns to the Digital Gold Era

Is Bitcoin history repeating? The article examines its shifting correlations with traditional assets like gold and tech stocks. Currently, Bitcoin's 90-day correlation with gold has surged to +0.56, its highest since 2020, while its link to the Nasdaq 100 and the US Dollar has faded. This suggests Bitcoin is behaving more as a scarce "digital gold" store of value, driven by macro concerns over currency debasement and sovereign debt, rather than as a high-beta risk asset. The analysis highlights two opposing macro forces: U.S. Treasury actions supporting long-term bonds (which weaken the dollar and benefit scarce assets) versus the Federal Reserve's ongoing inflation fight keeping real yields high (which pressures non-yielding assets like Bitcoin). The market remains highly sensitive to data like Non-Farm Payrolls and CPI, which rapidly shift Fed policy expectations and trigger significant Bitcoin volatility. While Bitcoin remains a core barometer for crypto market risk sentiment, the article concludes that the broader digital asset ecosystem—spanning decentralized exchanges, tokenization, and prediction markets—is developing independent growth drivers. This suggests the industry has the potential to evolve across various macro cycles, regardless of Bitcoin's short-term price movements influenced by interest rates and inflation data.

marsbit32m ago

History Repeating? Bitcoin Returns to the Digital Gold Era

marsbit32m ago

Liquid Attacker Keeps $47 Million for Themselves: Is It a White Hat Bounty or Disguised Extortion?

On September 6th, an attacker exploited a vulnerability in the Elements software, the foundation of Blockstream's Liquid Network, to generate approximately 4,000 fake LBTC tokens. Using the normal withdrawal channels of the SideSwap service, they exchanged these for roughly 3,998.5 real BTC (worth about $320 million) from Liquid's multi-signature wallet. The attacker then left a message in a Bitcoin transaction claiming to be "whitehats" and requested on-chain contact. After Blockstream patched the vulnerability upon contact, the attacker returned 3,400 BTC but kept approximately 598.5 BTC (worth around $47 million) in an address under their control. Blockstream is currently in communication to recover these remaining funds and has not officially recognized them as a bug bounty payment. This incident has sparked debate within the crypto community. Supporters argue the attacker acted as a white hat by exposing a critical flaw, securing the fix, and returning most funds, thus preventing a total loss. Some suggest a 15% retention could set a precedent for incentivizing ethical disclosures in major hacks. Critics, however, contend that exploiting the bug first to extract funds and then unilaterally deciding on a reward resembles extortion rather than standard white-hat procedure, where vulnerabilities are typically reported first for a negotiated bounty. Blockstream's ongoing recovery efforts indicate they view the withheld sum as assets to be retrieved, not an agreed-upon bounty.

marsbit32m ago

Liquid Attacker Keeps $47 Million for Themselves: Is It a White Hat Bounty or Disguised Extortion?

marsbit32m ago

Fed Enters the "Data Game": 0.01 Percentage Point Could Tip the Rate Decision

The Federal Reserve's upcoming interest rate decision hinges on a razor-thin margin, potentially determined by differences as small as 0.01 percentage points in upcoming inflation data. Markets are split between expectations for a rate hike and a pause, with key focus on this week's Producer Price Index (PPI) and Consumer Price Index (CPI) reports for August. Economists note the Fed's primary inflation gauge is the Personal Consumption Expenditures (PCE) price index. Current CPI and PPI data will be used to estimate the upcoming PCE reading. Analysts like Krishna Guha of Evercore ISI suggest that a core PCE monthly increase of around 0.21%-0.22% could lead the Fed to hold rates, while a 0.23%-0.24% rise would "likely tilt towards a hike," highlighting the precarious nature of the decision. Fed Chair Kevin Warsh's role is seen as crucial following his Jackson Hole remarks emphasizing the prolonged miss of the 2% inflation target. Market pricing currently implies about a 60% chance of a hike. Officials are divided, with some advocating for a hike to demonstrate anti-inflation resolve, while others stress data dependence. Further complications include upcoming revisions to historical PCE data and external pressure, such as former President Trump's recent threats related to trade and interest rates. Guha notes that if market expectations solidify strongly in favor of a hike before the meeting, it could make pausing more difficult for Warsh. Ultimately, the September decision depends not just on whether inflation is high, but precisely where the PCE lands and whether Chair Warsh can unite a divided committee.

marsbit46m ago

Fed Enters the "Data Game": 0.01 Percentage Point Could Tip the Rate Decision

marsbit46m ago

SpaceX Weight Increase Triggers $12.4 Billion in Passive Buying; Trial of 2.3 Billion Shares Unlock Still Ahead?

SpaceX's inclusion weight in the Nasdaq 100 index is set to rise from 1.25% to approximately 1.51% in the quarterly rebalancing effective September 21. This increase, triggered by a jump in its free float to nearly 30% following the unlocking of over 1 billion shares, is projected to drive about $12.4 billion in passive fund inflows, according to JPMorgan. However, this mechanical buying pressure coincides with a significant wave of share supply. More than 2.3 billion additional restricted shares are scheduled to become eligible for sale in two batches: over 1 billion before late October and another ~1.3 billion after the Q3 earnings report in mid-November. This creates a core market dynamic for Q4, pitting near-term index-driven demand against potential medium-term selling pressure from these unlocks. While the initial post-earnings unlock in August did not lead to major insider selling, analysts highlight the November unlock as a key test for the stock's true market depth. The price has recently traded sideways around its $135 IPO level in the absence of fundamental catalysts. Beyond the Nasdaq 100, future potential for passive buying remains, as SpaceX is already in the Russell 1000 but must wait until mid-2027 for potential S&P 500 eligibility. The interplay between index fund buying and the impending supply of shares will be crucial for SpaceX's pricing in the coming months.

marsbit46m ago

SpaceX Weight Increase Triggers $12.4 Billion in Passive Buying; Trial of 2.3 Billion Shares Unlock Still Ahead?

marsbit46m ago

CARF: What to Report? An Interpretation of Reportable Information and Local Implementation Differences

The OECD's Crypto-Asset Reporting Framework (CARF) defines the information that Reporting Crypto-Asset Service Providers (RCASPs) must disclose annually. This report consists of three core data types: information about the RCASP itself, details on reportable users (including entity controllers), and aggregated transaction data for relevant crypto-assets (excluding certain types like CBDCs). Transaction categories include acquisitions/disposals for fiat or other crypto-assets, reportable retail payments, transfers, and transfers to unknown wallets. However, as jurisdictions implement CARF into local law, key differences from the OECD standard emerge. Variations include whether domestic tax residents must be reported (e.g., UK and New Zealand require it; Japan and Singapore typically do not), the local currency used for reporting valuations, thresholds for retail payment transactions (e.g., 500,000 yen in Japan), specific reporting field requirements (e.g., birthplaces, Tax Identification Number formats), and rules for nil returns. For compliance, RCASPs must integrate CARF's tax due diligence into existing KYC/AML processes, establish consistent fiat conversion and valuation mechanisms, and meticulously track jurisdictional implementation rules, as local adaptations will dictate the final data submitted to tax authorities.

marsbit57m ago

CARF: What to Report? An Interpretation of Reportable Information and Local Implementation Differences

marsbit57m ago

Trading

Spot
活动图片