Anthropic published a blog post Friday seeking to answer some basic questions about how it will watermark the text generated by its chatbot Claude. Such as: How will the watermarking actually work? Can it be hidden with editing? And how does this affect code?

Claude users have been debating the move since the company revealed earlier this week that it would be doing this watermarking to comply with the EU AI Act’s Transparency Code, which requires AI companies to use systems that make it possible to identify AI-generated content.

On Reddit, for example, one poster characterized this as a conspiracy against innocent Claude users, while another claimed, “The only reason you wouldn’t want this is to lie to people.” And Business Insider reports that “dozens” of users on X have claimed to cancel their Claude subscriptions as a result.

Anthropic’s new post starts with a general overview of the watermarking concept, explaining that when making “low-stakes choices” — like choosing between the words “overcast” and “grey” to describe the weather — Claude can create a pattern in its responses that is “undetectable to the reader, but is detectable to anyone who has a key that encodes it.”

“Watermarking does not impact the quality of Claude’s output,” the company said. “To a reader, a watermarked response is indistinguishable from an unwatermarked one.”

More specifically, Anthropic said it will be using the SynthID-Text approach that the Google DeepMind team outlined in 2024, and that it plans to release a watermark detection API. It also noted that watermarking is distinct from the AI detection approaches offered by companies like Pangram that look for “tells” in the writing (like the construction “his isn’t [X], it’s [Y]”) to reveal AI usage: “Picking up on these patterns is fundamentally different from checking for a watermark.”

Could someone just rewrite the text to hide the watermark? Anthropic said it’s possible, but “light editing probably won’t remove the watermark completely,” while “a complete rewrite where every word is replaced will.”

“In the latter case, of course, it’s arguable whether the text can any longer be described as AI-generated,” the company said.

As for whether the watermark will be detectable in text that was only proofread or edited by Claude, Anthropic said that will depend on “the length of the text and how heavily Claude has edited it.” If it’s only been lightly edited, “nearly all the words” will have been written by the human author and “there’s very little (if anything) for the watermark to attach to.”

Code, meanwhile, should have less of a watermark than other text, because the model will need to create working code and won’t have the freedom to choose between a variety of equally valid options. 

“Having said that, in areas where there is an arbitrary choice between particular words or terms within the code, the watermark can be used, such as comments within code,” Anthropic said. “But by definition, it will have a negligible effect on the actual code produced.”

Anthropic also said that Claude won’t be the only AI chatbot to generate watermarked text, as “other major model developers have signed the same Code of Practice and will be implementing their own watermarks.”

Chinese memory giant CXMT has overtaken Tencent Holdings as the most valuable public company in China, underlining the speculative frenzy around memory companies at the moment.

The chipmaker debuted on the Shanghai Stock Exchange on July 27, and its shares surged 467 percent on their first day of trading. CXMT’s market capitalization later reached $524 billion, according to Bloomberg.

CXMT’s market capitalization reached approximately $524 billion 17 days after its IPO, surpassing Tencent’s roughly $511 billion valuation. The crossover came as CXMT remained near its post-IPO valuation and Tencent shares declined following a sharp increase in the company’s AI infrastructure spending.

Strategically important position in China

While CXMT is the fourth largest memory manufacturer worldwide, it holds a strategically important position as the only one of the top four based in China. Foreign semiconductor and memory manufacturers are still able to do business in the country, but the escalating trade war between the US and China has forced some to limit sales of their most advanced chips, while others have been restricted by the Chinese government in favor of homegrown manufacturers.

In the event of China or the US imposing further restrictions on foreign memory chipmakers, CXMT would become the largest domestic supplier, with a near-monopoly position in China’s DRAM market, which supplies AI data centers and smartphones.

It is not just Chinese companies interested in CXMT either, with Apple reportedly testing the company’s DRAM for its MacBook and iPhone lines. Apple has, like the rest of the industry, suffered from the memory supply crunch, which has forced the iPhone maker to increase the prices of almost all of its hardware.

According to Counterpoint Research, Samsung is the leading manufacturer of DRAM with 39 percent market share, followed by SK Hynix and Micron with 29 percent and 25 percent, respectively. CXMT has seen its market share increase over the past few years, but it still accounts for less than 10 percent of the total.

More must-read AI coverage

Data centers insatiable appetite for memory chips

Huge demand for memory chips has sent the stock prices of the major DRAM suppliers through the roof. SK Hynix, which was valued at $135 billion this time 12 months ago, surpassed a $1 trillion market cap in May, a huge increase for a company focused primarily on memory. Even Samsung, better known for its smartphones and TVs, is now seeing the majority of its revenue come from its memory chip and foundry businesses.

Data centers, especially those linked to AI training and operations, have a seemingly insatiable appetite for these chips, reshaping the entire consumer electronics market. Smartphone manufacturers, which were among memory chipmakers’ biggest customers a few years ago, are now struggling to secure favorable contracts for memory.

This has also affected the video gaming market, with Xbox, Sony, Nintendo, and Valve all having to either increase the price of their consoles or delay the launch of newer hardware. As AI infrastructure absorbs an ever-larger share of global memory supply, consumers are increasingly feeling the impact through higher prices and delayed product launches.

Read more: CXMT’s Shanghai IPO sent its market value soaring as investors backed China’s growing role in the AI memory-chip market.

Singapore has decided that one way to teach people how not to fall for scams is to let them experience one first. A simulated one, thankfully.

Singapore’s Cyber Security Agency is running a six-month National Simulated Scams Exercise, in which volunteers receive simulated robocalls designed to mimic government impersonation scams. The pilot runs through Aug. 31 and is part of Singapore’s broader push to make its population more resilient to increasingly sophisticated digital fraud.

The experiment raises a useful question for businesses, too: If employees can be fooled by increasingly polished calls, messages, deepfakes, and impersonation attempts, is an annual security-awareness course really enough?

Singapore wants people to experience a scam before the real one arrives

Singapore’s Cyber Security Agency, with support from the Ministry of Home Affairs, launched the exercise on March 1. Participation is voluntary, but those who sign up don’t know exactly when the simulated scam will arrive.

At some point during the six-month exercise, participants receive robocalls that mimic Government Official Impersonation Scams (GOIS). According to CSA’s description of the exercise, the controlled simulation is intended to let people experience scammers’ techniques firsthand and learn what to do when they encounter similar tactics outside the exercise.

The experiment isn’t limited to yesterday’s scam tactics. In a July update on Singapore’s AI-driven threat landscape, CSA said the pilot includes AI-enabled government impersonation scam calls. That matters because the scam itself increasingly resembles a conversation rather than a suspicious link.

Voice-based social engineering can put victims under pressure in real time, exploiting authority, urgency, and fear before they have time to verify what they’re hearing. Security researchers have also found increasingly sophisticated tooling designed specifically for these attacks.

For example, TechRepublic previously covered phishing kits built for voice-based scammers that can provide attackers with real-time information as they try to persuade victims to approve multifactor authentication requests. Singapore’s exercise effectively gives participants a fire drill for that moment.

The scam problem is still expensive

Singapore has good reason to experiment.

According to Singapore Police’s 2025 scam and cybercrime figures, the country recorded 37,308 scam cases in 2025, with victims losing approximately S$913.1 million. Both figures declined from the previous year, but the losses still illustrate the enormous financial consequences of successful scams.

Government impersonation scams moved in the opposite direction.

Singapore’s Annual Scams and Cybercrime Brief 2025 shows GOIS cases more than doubled from 1,504 in 2024 to 3,363 in 2025, while reported losses climbed from S$151.3 million to S$242.9 million. The basic technique exploits something no software patch can completely remove: trust.

Scammers may pose as banks, government agencies, police officers, regulators, or other seemingly authoritative organizations. Their goal is often to create enough urgency or fear that the victim acts before independently checking the story.

That’s why technical defenses alone haven’t made the problem disappear. As TechRepublic previously examined in its analysis of Singapore’s S$913 million scam problem, even longstanding identity requirements for SIM registration haven’t eliminated telecom-enabled fraud. Attackers adapt around controls rather than politely crashing into them.

Must-read security coverage

What IT leaders can borrow from Singapore’s experiment

Businesses don’t need to start prank-calling their entire workforce tomorrow morning.

But Singapore’s experiment points toward a useful principle for security teams: People may learn more from safely experiencing an attack than from being told what one looks like. Traditional phishing simulations already use that idea. The difference is that the threat surface has expanded well beyond the inbox.

An employee might now receive a WhatsApp message supposedly from an executive, a convincing phone call from “IT,” a video call featuring a digitally manipulated face, or a request presented as a confidential assignment from senior management.

That’s not hypothetical. In a 2026 advisory on executive impersonation scams, Singapore Police warned that criminals had impersonated company executives on WhatsApp and, in some cases, used digitally altered appearances during video calls. Victims were told they were working on confidential projects and instructed not to discuss them with colleagues, cutting off one of the easiest ways to discover the deception.

For IT and security leaders, that suggests simulation programs should test more than whether an employee clicks a suspicious email.

Teams could practice scenarios involving:

• unexpected calls from supposed IT staff asking for credentials or MFA approval;
• urgent messages from executives requesting payments or sensitive information;
• requests to move a conversation from an official channel to WhatsApp or another messaging service;
• supposed regulators or law-enforcement officials demanding immediate action;
• voice or video impersonation intended to override an employee’s normal verification process.

The objective isn’t to catch employees making mistakes. It’s to build a reflex: Stop, verify, and use a second channel before acting.

That becomes more important as social engineering grows more interactive. TechRepublic has reported on the surge in social engineering attacks, including attackers posing as help-desk or IT personnel to exploit trust and urgency and persuade employees to weaken authentication controls.

Security training may need to feel more like a fire drill

There’s an obvious limitation to Singapore’s approach: Participants volunteered.

Employees in the real world don’t get to opt in before a criminal targets them. Nor does recognizing one simulated government scam guarantee that someone will spot the next fake CEO, supplier, help desk worker, recruiter, or bank representative. But the underlying idea is harder to dismiss.

Organizations regularly rehearse fires, evacuations, outages, incident-response procedures, and disaster-recovery plans because knowing a procedure isn’t the same as executing it under pressure.

Social engineering may deserve the same treatment. Instead of asking whether employees completed their annual cybersecurity module, security leaders may increasingly need to ask whether employees have practiced responding to the kinds of attacks they’re actually likely to encounter.

Singapore is betting that experiencing the trick once, in a controlled environment, can make the real trick easier to recognize.

For businesses facing AI-assisted impersonation, voice phishing, deepfakes, and increasingly personalized fraud, that may be the more useful lesson: Don’t just teach employees what a scam looks like. Give them practice saying no to one.

Related reading: For another sign of how technology is reshaping Singapore, the country recently raised its 2026 growth forecast as booming AI demand fuels electronics exports and manufacturing.