Media Outlets Resigned To Changing Web Revenue Model

This Guardian article revisits the fear of media outlets that generative AI chatbots and Google AI Overviews are significantly reducing visits to their Websites.

• Reuters survey of 280 media companies across 51 countries found that the companies expect a fall of 43% in search engine referrals over the next three years.
• Google Search for 2’500 news sites has fallen by 33% globally over the past year. Only publications related to live reporting or current affairs are unaffected by this trend.
• The Reuters Institute also says that media outlets are increasingly pushing journalists to become more like social media content creators.
• There seems to be an emerging consensus that the “traffic era” as a revenue model for new sites is coming to an end.

Organizations are losing ground in the fight against AI-assisted cyberattacks. The situation is complicated by business executives embracing generative AI and 89% of whom, according to Gartner, are willing to bypass security guidance to meet business objectives.

• One use of generative AI by criminals is to reverse engineer software patches. This allows them to understand the vulnerability that is being patched, and then to launch attacks on organizations which have not applied the patch. Experts recommend applying software patches within 72 hours.
• For one executive, a fundamental challenge brought by AI is its unpredictable nature: “Defense-in-depth strategies predicated on deterministic rules and static signatures are fundamentally insufficient against the stochastic, semantic nature of attacks targeting AI models at runtime.”.
• Among the main attack vectors for generative AI are prompt injection attacks and indirect prompt injection via RAG (where the attack prompt is hidden in documents retrieved in RAG operations).
• An emerging form of attack is multi-turn crescendo attacks which exploits the context aspect of generative AI models. Here, the malicious prompt payload is distributed over several prompts of the conversation. No individual prompt is considered malicious, but the combination leads to an attack. The Crescendomation tool achieved a 98% attack success rate on GPT-4 and a 100% success rate on Gemini-Pro.
• Other more well-known attack types come from data exfiltration by negligent insiders (pasting proprietary content in prompts to model), obfuscation attacks where malicious prompts are encoded using ASCII art or Base64, and deepfake enabled fraud.

This article considers the perception, notably in marketing, to content created by generative AI.

• The McDonald company removed a commercial in December entitled “The Worst Time of the Year” that was AI-generated after receiving criticism. Coca-Cola has also been criticized for commercials entitled “Holidays are Coming” that were AI generated. Critics have used terms such as “soulless”, “dystopian”, “devoid of creativity”, and “disjoint”.
• The challenge is that AI content is too polished, cheap to produce and perhaps boring to consume. For the Instagram Chief, “People want content that feels real. In a world where everything can be perfected, imperfection becomes a signal. Rawness isn’t just aesthetic preference anymore. It’s proof that you’re offering authenticity, reality, value.”.
• Meta says that its content selection algorithm is being worked so as to prioritize content that appears human-generated over AI-generated content.

In this MIT Technology Review article, the journalist relates lessons from the recent Consumer Electronics Show in Las Vegas which attracted over 4’000 exhibitors. One quarter of these were Chinese companies.

• AI is being sold in nearly all types of products, from PCs and other IT systems, to household devices. An example is Luca AI, a robot panda that watches babies, and Fuzozo, a robot personality-infused pet. China has an advantage in electronic gadget AI because it has the edge in consumer electronics.
• Another area where China is investing heavily is in humanoid robots. One robot demonstrated at the conference was from the Unitree company in Hangzhou. One of its robot “challenged” humans to boxing bouts. The goal of the company was to demonstrate the stability and balance of the robot.
• One advantage for China in the robot manufacturing is its access to supply chains and manufacturing sources for EVs, batteries, motors and sensors.
• The article reports how Chinese companies are generally working on all layers of the software and hardware stack for robots, and suggests that the open-source culture adopted by China is positively impacting research.
• Another advantage is scale. This allows companies to “iterate fast” compared to Western companies.
• One challenge for robot development is that their AI needs to use “vision-language” models rather than language models – the training data needed for such models is not so easy to capture as for language models.

The US administration has imposed a 25% tariff on advanced AI semiconductor chips fabricated outside of the US, including the Nvidia H200 AI chips, which are passing through the US for sale to approved customers in China.

• An executive order signed by Donald Trump writes that “The United States currently fully manufactures only approximately 10% of the chips it requires, making it heavily reliant on foreign supply chains. This dependence on foreign supply chains is a significant economic and national security risk.”.
• The Chinese government decided last year not to support data centers that used foreign chips. However, there are signs that the government may loosen this restriction since Chinese chips are not yet powerful enough to compete with Nvidia’s in data centers.
• Nvidia said they welcomed the “decision to allow America’s chip industry to compete to support high-paying jobs and manufacturing in America. Offering H200 to approved commercial customers, vetted by the Department of Commerce, strikes a thoughtful balance that is great for America.”.

This article by company Netcall’s CIO, Richard Farrell, looks at some of the mistakes made in trying to adopt AI in 2025, and how these mistakes can be avoided in 2026.

• One mistake mentioned is the large fragmentation of AI tools that have been used within organizations. These tools are often designed to address efficiencies in individual tasks, not efficiencies at the process level.
• AI tools should more integrated – maybe from fewer providers – and integrate better with existing organizational IT.
• The starting point for AI integration must be the organizational process, which should be mapped out to see where inefficiencies can be improved.
• Another failing of AI implementations in 2025 was the absence of precise metrics for implementation success. The goal of AI is to analyze data for actionable outcomes. When these are missing, the reason for AI is lost.
• The Year 2026 will see increased attention paid to governance issues with the implementation of safety, audit trail, and personal data protection protocols.

Journalists from the Guardian have found that X’s standalone Grok App still allows users to post nonconsensual sexualized content on the platform. For instance, some users have continued to post AI-generated videos of women stripping down to bikinis.

• X had said last week that it had “implemented technological measures to prevent the Grok account from allowing the editing of images of real people in revealing clothing such as bikinis”, and that it had “zero tolerance for any forms of child sexual exploitation, nonconsensual nudity, and unwanted sexual content”.
• A spokesperson for the End Violence Against Women coalition wrote that “The continued ease of access to sophisticated nudification tools clearly demonstrates that X isn’t taking the issue of online violence against women and girls seriously enough”.
• The UK prime minister said that “Free speech is not the freedom to violate consent… Young women’s images are not public property, and their safety is not up for debate.”.
• Canada has becomes the latest country to open an investigation into X.
• The debate around the platform is not yet interfering with its popularity. Elon Must shared a post late last week claiming that X’s “popularity and real world usage are skyrocketing globally”.

• OpenAI is in need of revenue as it plans to invest over 1 trillion USD in AI infrastructure by 2030.
• Some analysts believe that OpenAI is preparing an IPO soon.
• Advertisements have always been a temptation for the company which currently has 800 million active weekly users. The danger is that advertisements might upset some users who would then leave for competing chatbots.
• The move can also put pressure on these competing companies. One analyst wrote that OpenAI’s move to advertisement would force rivals clarify their own monetization approaches.

This article reviews a breakthrough at Google in reinforcement learning that can help models solve complex reasoning tasks more efficiently. Reinforcement learning is used in a post-training phase of language models, notably for tasks that are multi-step such as those undertaken by AI agents.

• Language models generate output sequences one token at a time. They explore new strategies by making small changes to next output token. This works well in natural language processing tasks, but this is the wrong level of abstraction for reasoning tasks which have many steps.
• In the approach put forward at Google, called internal RL, an internal neural network controller is introduced that applies changes to activations of artificial neurons in the middle layers of the neural network. This gives greater possibilities for solution searches for tasks with long horizons.
• Internal RL gives less predicability, or more creativity, than traditional token-by-token which is more predictable. In practice, a combination of approaches might be needed. In the case of software development for instance, internal RL would be used for finding algorithms, while the more predictable token-by-token approach would handle code generation where syntax needs to be precise.

This Business Insider article reports how Apple no longer has the same level of control over its supply chain, as providers and assembly companies are increasingly attracted by AI companies.

• There is a general understanding that tech companies need to control the hardware supply lines to be the most successful. Apple has been losing ground here to AI and cloud companies like Nvidia, Amazon, Microsoft, and Google.
• For instance, TSMC is the world’s largest chipmaker and fabricates the iPhone chips. In recent financial reports, it has become clear that Apple is no longer TSMC’s most important segment. Chips for Nvidia and hyper-scale data centers now account for 58% of TSMC revenue.
• Memory chip makers are also becoming more attracted to data centers as these require high quantities of dynamic RAM. This demand for DRAM has led to an increase in memory prices, which is leading to increases in smartphone prices.
• Foxconn, a manufacturing company a large part of whose revenue came from iPhone assembly, is also prioritizing AI companies now.