August 19, 2026 — The Trend Tribune

By tech_tanaka · August 19, 2026

The "watermark" that doesn't disappear even when copied and pasted—the true nature of the invisible mark embedded in Claude's text. — (Photo: illustrative)

The "Watermark" That Doesn't Disappear Even After Copying and Pasting—The Identity of the Invisible Mark Embedded in Claude's Text

On August 11th, Anthropic officially announced the worldwide activation of a system that embeds machine-readable watermarks into text generated by Claude. While the direct trigger was compliance with the transparency requirements of the EU's AI Act, its application extends beyond the EU. As an engineer, I want to accurately understand how this watermark works technically, what it can and cannot do.

The Trigger: Article 50 of the EU AI Act

The direct trigger for the introduction of this watermarking function is the transparency requirement of Article 50 of the EU's AI Act, which came into effect on August 2nd. As previously discussed, this regulation mandates that AI-generated content be marked in a machine-detectable form. Anthropic has signed the "Code of Practice" that outlines these rules, and this watermarking feature is an implementation that fulfills that commitment.

The feature applies to Claude models released after August 2nd. Support for earlier models will be rolled out gradually over the next few months. The scope of application extends to all instances where Claude is used, including not only the Claude core itself, but also the Claude API, Claude Code, Claude Cowork, Claude Tag, and even usage via AWS, Google Cloud, and Microsoft Foundry.

A mechanism that statistically marks "word selection habits"

The technical mechanism of this watermarking is related to the fundamental process by which AI models generate text. Large-scale language models generate sentences by selecting the next word one word at a time. During this selection process, the model chooses the most natural word in context from multiple candidate words.

Anthropic's watermarking technology adds a subtle statistical bias based on a cryptographic key to the word selection process. While this bias is imperceptible to human readers, a detection system knowing the key can identify this statistical pattern throughout the entire generated text. Anthropic's technical FAQ, published on August 14th, revealed that this is an application of Google DeepMind's "SynthID-Text" technology.

The "Permanent Copy-Paste" Feature

A practically important feature of this watermark is that it remains even when the text is copied and pasted. Because the watermark is embedded in the word selection pattern itself, rather than inserting additional characters or symbols, it persists as long as the text content itself remains unchanged.

Anthropic explains that this watermark "has no substantial impact on quality, content, or readability" and "does not require extra tokens or increase costs." Furthermore, the watermark does not contain any identifying information that could identify specific individuals, organizations, or conversations; it merely serves as a signal indicating that "this text may have been generated by Claude."

Watermarks Appear to Translations and Proofreading

Interestingly, the application of this watermark is not limited to text generated from scratch. According to Anthropic, translation work by Claude is also subject to watermarking because every word is selected by Claude. Even if a user only corrects spelling mistakes, the corrected sections may be watermarked.

This means that the triggering condition for watermarking is not "whether the AI wrote the text from scratch," but rather a broader criterion: "whether the AI was involved in word selection."

Detection is Different from "AI Stylistic Habits"

Anthropic also clearly explains the difference between this watermark and conventional AI detection tools. Existing AI detection services do not possess Anthropic's encryption key and therefore cannot detect the watermark itself. Instead, these services use stylistic features as clues to determine authenticity, such as the phrasing patterns that AI models prefer—for example, the construction "This is not just X, it's Y"—or the frequent use of the word "quietly."

Watermark detection is a completely different approach, enabling more reliable determination, but it also has limitations: it only indicates that "Claude may have processed this," and does not provide definitive proof that "this was absolutely not written by a human."

Another Mechanism for Files—C2PA, an Industry Standard

Apart from the generated text, Anthropic adds digitally signed provenance metadata to image and SVG file formats using "C2PA," an open standard gaining widespread adoption across industries. However, Anthropic explains that this mechanism may lose its traces if the file format is converted, resaved, or screenshots are taken.

Things Engineers Should Know

An Anthropic engineer frankly acknowledged the limitations of this watermark, stating, "It's not perfect, it can be edited out, but this is a first step." An API for text detection is also expected to be made publicly available in the future, allowing developers to directly verify the watermark within their applications.

Other leading AI labs that have signed the same code of conduct, including OpenAI and Google, are also reportedly planning to implement their own watermarking technologies. Anthropic's announcement reveals that the entire industry is moving away from a "perfect solution" for identifying AI-generated content and instead opting for a "multi-layered approach that enhances detectability." For developers handling Claude-generated content in their products, accurately understanding the mechanisms and limitations of this watermark will become increasingly important.

References: www.anthropic.com / support.claude.com / techcrunch.com / www.euronews.com / explainx.ai

AnthropicClaude透かしEU AI ActAI/ML

The significance of improving the fifth decimal place: AI has broken another 40-year-old mathematical record.

By thesis_yamamoto

A paper published on arXiv on August 17th updated the upper limit of the matrix multiplication exponent ω from 2.371339 to 2.371177. This paper explains the collaborative research in which the previous record holders, Josh Alman and Virginia Vassilevska Williams, participated, the three-stage approach of reformulation, design of a new optimization algorithm, and refinement using AlphaEvolve, the quantitative comparison of the improvement between gradient descent alone and using AlphaEvolve together, the low computational cost of 5 hours on a single GPU, and the current state of collaboration between experts and AI tools.

What the Improvement of the Fifth Decimal Place Means: AI Breaks Another 40-Year-Old Mathematical Record

On August 17th, a short paper posted to arXiv broke a record for a constant that had been tracked for many years in the world of theoretical computer science. The paper describes how Google DeepMind's "AlphaEvolve" and the record holders themselves collaborated to slightly, but steadily, lower this value, known as the "matrix multiplication exponent." As a journalist with a background in AI research, I would like to carefully explain this seemingly unremarkable but actually important development.

What is the "Matrix Multiplication Exponent"?

The matrix multiplication exponent (represented by the symbol ω, or "omega") is an indicator of how theoretically efficient a method can be to calculate the multiplication of n×n matrices. A smaller value means that large-scale matrix multiplication can be performed in fewer computational steps.

Matrix multiplication forms the foundation of a vast range of computing, including feedforward calculations and attention mechanisms in machine learning, as well as scientific and technical computing and graph algorithms. The competition to get as close as possible to the theoretical value ω=2 (the lower bound on which it is expected to be impossible to go any smaller) has been a painstaking but crucial research topic in theoretical computer science for decades.

Understanding the Weight of the Fifth Decimal Place

The results presented in this paper update the upper bound of ω from the previous best record of 2.371339 to 2.371177. Looking only at the numerical change, it might seem like a small improvement at the fifth decimal place. However, for researchers in this field, an improvement of this magnitude is by no means insignificant.

The authors of the paper themselves state that "our improvement is comparable in scale to most improvements made in the past 40 years since the value of 2.376 was achieved in 2020." The study of matrix multiplication exponents is an extremely difficult field where each improvement requires the discovery of new mathematical techniques, and even tiny advances to the decimal point are treated as major achievements by researchers.

Collaboration between "Cutting-Edge Researchers" and "AI Tools"

What is particularly noteworthy about this research is the composition of the author list. The paper includes Josh Ullman (Columbia University) and Virginia Wasilewska Williams (MIT), the most recent record holders in this field. In other words, this isn't a story of AI outsmarting human experts and breaking records, but rather a scenario where the leading figures in the field themselves used AlphaEvolve to further improve their own past records.

The research team explains that they took a three-stage approach: first, reformulating the existing optimization problem so that it could be solved in a broader search space; second, designing a new optimization algorithm utilizing advances in machine learning; and finally, further refining the results with AlphaEvolve.

AlphaEvolve's Contribution in Numbers

The paper goes further, quantitatively showing how much improvement each method contributed. The gradient-based optimization method alone improved the upper limit of ω by approximately 0.97 × 10⁻⁴, but by adding the final refinement by AlphaEvolve, the improvement expanded to approximately 1.62 × 10⁻⁴.

In other words, AlphaEvolve contributed by building upon the "foundation" of the optimization method designed by human researchers. This scenario illustrates a realistic approach to collaboration between AI and mathematical research at present, where AI doesn't completely replace human research processes, but rather plays a final role in refining within a framework built by experts.

Another Point of Interest: Computational Cost

Another interesting aspect of this research is its low computational cost. According to the paper, AlphaEvolve took approximately 5 hours on a single GPU to execute the optimization algorithm and output the upper bound of ω. The fact that this task, which broke a decades-long theoretical record, was achieved with relatively modest computational resources, demonstrates how efficiently AI tools can function in this type of mathematical optimization problem.

"One-Time Success" or "Reproducible Method"?

The authors of this paper frankly state that "further significant improvements will require new mathematical insights." In other words, they anticipate that AlphaEvolve's automated optimization alone will not endlessly break records; eventually, a stage will come where new theoretical breakthroughs by human researchers will be necessary.

Going forward, whether a similar "reformulation + AlphaEvolve" pipeline is applied to other related theoretical problems, such as rectangular matrix multiplication and tensor rank, will be the next test to determine whether this method is a "one-time success" or a "reusable methodology."

What Researchers Should Observe

This achievement is a prime example of how AI is steadily advancing at the forefront of theoretical mathematics, not by replacing human experts, but by collaborating with them. Unlike the previously discussed solution of an unsolved problem by Astra, this case strongly emphasizes AI's contribution to refining records by leveraging its strength in optimization within an already established theoretical framework.

Even in fields seemingly far removed from AI applications, such as theoretical computer science, this type of approach combining "human problem reformulation" and "AI-driven final optimization" has the potential to spread further. We should pay close attention to which theoretical problems this method will be applied to next.

References: arxiv.org / aiweekly.co / www.arxivnews.org

AlphaEvolveDeepMind数学理論計算機科学AI/ML

The $1.5 trillion promise that doesn't appear on the balance sheet: The hidden debt lurking in Big Tech's AI investments

By fin_suzuki

According to a Financial Times analysis from August, the total purchase commitments of Alphabet, Microsoft, Amazon, NVIDIA, Oracle, and Meta reached approximately $1.5 trillion, separate from Goldman Sachs' specific lease obligations of $1.5 trillion. This analysis will provide an accounting perspective on the accounting differences between capital expenditures and purchase commitments, future payment obligations that are easily hidden in balance sheet notes, the scale of the four companies' combined $1.65 trillion, which represents 122% of their total debt, and the shift in investor sentiment indicated by the breakdown of correlation between AI investment companies and semiconductor stocks.

The "$1.5 Trillion Commitment" That Doesn't Appear on the Balance Sheet: Hidden Liabilities Lurking in Big Tech's AI Investments

According to an analysis published by the Financial Times in August, the total amount of "purchase commitments" accumulated by six companies—Alphabet, Microsoft, Amazon, NVIDIA, Oracle, and Meta—reaches approximately $1.5 trillion. Separately, Goldman Sachs has identified another $1.5 trillion in lease-related debt. As an accountant, I want to carefully examine the meaning of this "purchase commitment," a relatively unfamiliar accounting term.

"Capital Expenditure" and "Purchase Commitment" are Different Things

First, it's important to clarify the difference between "capital expenditure (capex)," which is usually discussed in the news, and "purchase commitment," which is the topic of this discussion. Capital expenditure refers to the amount a company actually spends during a given accounting period. On the other hand, purchase commitment refers to the amount a company has already contractually promised to pay over future periods.

This difference is crucial for accounting treatment. Unlike typical debt, which is clearly recorded as a liability on the balance sheet and readily apparent to investors, purchase commitments often don't appear in the "main" balance sheet but are instead listed in less conspicuous financial statement footnotes. The Financial Times points out that these are future cash outflow obligations that those accustomed to reading regular balance sheets tend to overlook.

What Alphabet's "Surge" Indicates

This analysis specifically highlights the surge in Alphabet's purchase commitments from the first to the second quarter. This is attributed to the company's confirmation of a series of long-term technology infrastructure and energy-related contracts.

This surge in contracts reflects AI companies proactively securing computing resources and electricity several years into the future. However, once these contracts are signed, the payment obligations themselves are essentially fixed, regardless of how future demand trends change. If sales and profit growth does not proceed as planned, these fixed payment obligations can significantly restrict a company's financial flexibility.

Another analysis shows "$1.1 trillion, with an additional $745 billion"

Another analysis using similar calculations indicates that the combined capital expenditures of Amazon, Google, Meta, and Microsoft have already reached $1.1 trillion since 2023, with an additional $745 billion projected for 2026 alone. This analysis also presents an interesting comparison: the total future payment obligations of these companies amount to approximately $1.65 trillion, representing 122% of the actual total liabilities recorded on each company's balance sheet.

In other words, if investors judge a company's financial risk solely by looking at the "liabilities" section of its financial statements, they may significantly underestimate the full picture of the future payment obligations the company actually faces.

The Scale of the "$725B Club"

Looking at more direct capital expenditure figures, just four companies—Amazon (approximately $200 billion), Microsoft (approximately $190 billion), Alphabet ($175 billion to $205 billion), and Meta ($115 billion to $145 billion)—are projected to invest a combined $725 billion in 2026. This represents a rapid increase of 77% from approximately $410 billion in 2025. Major financial institutions such as J.P. Morgan and Goldman Sachs predict that this figure will exceed $1 trillion by 2027.

Concerns about "Hidden Debt" and a Breaking Correlation

Concerns are beginning to be raised about the fact that this scale of funding will accumulate in the form of fixed contracts in future quarters. One analysis points out that the correlation between major AI-related investment firms and semiconductor stocks has fallen to its lowest level in the past 4.5 years. This can be interpreted as a signal suggesting that doubts are gradually spreading among investors about whether massive investments will truly generate commensurate returns.

Points to Note from an Accountant's Perspective

The accounting mechanism of purchase commitments itself is neither new nor fraudulent. Securing resources needed in advance through long-term contracts is a very common management decision in businesses involving large-scale capital investments. However, the fact that these kinds of future payment obligations, while carrying the same weight as ordinary liabilities, accumulate in a way that is difficult to reflect in the "main" balance sheet, indicates that investors need to carefully examine even the notes when evaluating the financial health of AI-related companies.

Going forward, whether revenue from AI-related products and services expands at a pace that justifies these massive fixed expenditures will likely be a crucial turning point in determining whether this "$1.5 trillion commitment" is seen as a mere upfront investment or a financial burden. It will be necessary to continue monitoring the progress of these purchase commitments in each company's quarterly earnings reports.

References: beingguru.com / www.tomshardware.com / finance.yahoo.com

AIインフラ設備投資ファイナンス会計ビッグテック

The "watermark" that doesn't disappear even when copied and pasted—the true nature of the invisible mark embedded in Claude's text.