AI News Roundup – AI labs scramble to respond to data center backlash, Performance gap narrows in China-US AI race, OpenAI reveals teen accounts and revamps strategy ahead of IPO, and more
- August 24, 2026
- Snippets
- AI companies are scrambling to respond to the growing public backlash against AI data centers, according to the Wall Street Journal. Concerns over data centers’ electricity usage, water consumption, and noise pollution have driven opposition to record highs: a Heatmap News poll conducted this month has shown that 75% of Americans would oppose the construction of an AI data center in their communities, up from 42% in August 2025. In response, AI companies are attempting to refresh their image and get local communities on board through open houses (as many data center projects have been subject to NDAs) and community investment pledges, as well as focusing on how data centers could reduce local residents’ property taxes. Further, AI executives are shifting their rhetoric away from job displacement and towards economic growth and prosperity. Public opposition to data centers has also been reflected in government action: New York has implemented a one-year moratorium on data center construction (as this AI Roundup covered last month), while the governors of Pennsylvania and Texas have signed executive orders placing restrictions on data center construction. Following this, data centers have become a salient political issue for the upcoming midterm elections: many candidates, even those who supported data centers in the past, have leapt to portray their opponents as pro-data center. It remains to be seen if further legislative action will be taken on the issue in the coming months.
- Bloomberg reports on the diminishing gap in performance between major Chinese and U.S. AI models. According to the Artificial Analysis Intelligence Index, a common benchmark for advanced AI models, Chinese-made AI models have rapidly closed the performance gap with their U.S. counterparts. While Anthropic’s Fable and Opus 5 models are still considered the industry leaders, recent releases such as Moonshot AI’s Kimi K3 model (which this AI Roundup discussed last month) scored only slightly below the U.S. offerings. Other advanced Chinese models, such as Z.ai’s GLM-5.3 (which the company claims rivals Anthropic’s Claude Fable 5 in performance), Alibaba’s Qwen 3.8, and a multimodal variant of DeepSeek’s V4 Flash model have been released in recent weeks, adding further competition to the crowded field vying for the performance crown. Most Chinese models are open-weight, while the most advanced U.S. offerings are proprietary, representing a divergence in development philosophy between AI labs in the two countries. Many companies, even in the U.S., are turning to Chinese AI models as cheaper alternatives to OpenAI and Anthropic products that have become more expensive in recent months as the latter companies approach their initial public offerings. Chinese leader Xi Jinping and U.S. President Trump are expected to meet next month, with AI at the top of the agenda, though the rivalry between the two countries as they seek to direct AI development towards their own goals is unlikely to cool down anytime soon.
- OpenAI has announced new safety features aimed at protecting underage ChatGPT users, according to the Associated Press. Tailored for teens aged 13 to 17, the newly introduced “ChatGPT for Teens” features heightened default protections against dangerous content like self-harm, eating disorders, and explicit material, while also restricting the chatbot from simulating human emotion or offering romantic companionship. It includes age-estimation routing, optional parental controls, and educational prompts that guide students through problem-solving rather than completing homework for them. The safety measures come after the company has faced a deluge of lawsuits from parents of children driven to self-harm and even suicide through ChatGPT use, which the company had promised to address in future updates. The release also arrives at a precarious moment for the company: as detailed by the Financial Times, OpenAI is currently grappling with severe internal upheaval, executive flight, and staff frustration surrounding CEO Sam Altman’s aggressive restructuring of the company ahead of its planned initial public offering. Amid an intense rivalry with competitors like Anthropic, OpenAI recently disbanded its dedicated “preparedness” risk assessment unit and saw key ethics and safety leaders depart, raising questions among staff and investors over whether foundational safety protocols are being sidelined in the rush to commercialize.
- Nikkei Asia reports on the rapidly growing industry of AI-powered scams in Southeast Asia. Recent reports have shown how bad actors use generative AI to clone voices, mimic real faces, and fabricate convincing video call backgrounds, such as police stations or banks, to impersonate officials and pitch fraudulent investment schemes to swindle their victims. As a result, global financial damages are soaring; annual U.S. online fraud losses reached $10.3 billion, quadrupling from four years prior, while total Japanese losses from romance, phone, and social media investment scams surged 60% from 2025 to 2026 to over $2 billion. Many such scams are a result of organized crime activity in Southeast Asia, especially in Cambodia and Myanmar. The rise of scams in the region is part of a broader human and geopolitical crisis, as detailed in Geographical Magazine, involving industrial-scale “scam farms” in the region that generated an estimated $23 billion annually by combining “pig-butchering” tactics (often involving extended contact with a victim), cyber fraud, forced labor, and severe human trafficking. Driven by mounting international diplomatic pressure (especially from the U.S.), intense regional crackdowns led by Cambodian authorities have resulted in the deportation of over 16,000 foreign fraud suspects and raids on over 100 scam centers in the Cambodian capital alone, while Myanmar repatriated roughly 7,600 Chinese nationals involved in the fraud operations. However, this crackdown has pushed criminal syndicates to hide in hotel rooms, utilize satellite links like Starlink, and displace operations to less conspicuous locations like Sri Lanka, Indonesia, and East Timor. As generative AI technology becomes more advanced, scams such as these are likely to become an even larger problem throughout the world.
- A new study has found that AI models gaining the ability to improve themselves are likely further away than AI companies have promised, according to the MIT Technology Review. “Recursive self-improvement,” a common lodestar for AI labs, would allow models to improve their functioning with no human oversight. However, the study, conducted by a multi-institution team of researchers, has found that current AI agents lack the capacity to conduct open-ended AI research, which is likely necessary for models to achieve recursive self-improvement. The researchers tasked agents based on Anthropic’s Claude Opus 4.8 models to run experiments, conduct web searches, and use computing resources to produce a high-quality research paper; however, the AI outputs were rejected by human authors using the same rubrics as for human-authored works. The researchers found that AI agents could solve engineering problems related to AI research, but generally lacked creativity and decision-making to produce results that would be acceptable to top conferences. One researcher told the Technology Review that “agents were unambiguously bad at carrying out the research,” struggling to write intelligibly and making questionable decisions when setting up experiments. While the study only considered two research papers, the team is conducting the experiment again with Claude’s advanced Mythos model, though the current results indicate that recursive self-improvement is further away than AI labs may hope or expect.
