AI News Roundup – US election results could impact AI data center buildout, Europe’s AI watermarking rules run into technical challenges, China’s Tencent releases new flagship AI model, and more

To help you stay on top of the latest news, our AI practice group has compiled a roundup of the developments we are following.

• The results of this past week’s midterm elections in the United States could greatly impact the rapid AI data center buildout, according to the Associated Press. Sweeping Democratic victories across congressional races and state legislatures have empowered a new generation of public officials who campaigned against data centers, raising the likelihood of tighter regulations and local construction moratoriums on the facilities. Democrats flipped governorships in several states, including Ohio, Georgia, and the major data center market of Texas, while securing full state government control in six states, including Wisconsin and Pennsylvania. In Texas, Democratic governor-elect Gina Hinojosa ran on blocking data center permits that disadvantage local communities, and agricultural commissioner-elect Clayton Tucker attributed his historic victory to rural anger over allegations that data centers deplete scarce groundwater supplies. Data Center Watch, a project of the consulting firm 10a Labs, found that 32 of the 34 newly elected or reelected state governors favor stricter oversight of data centers, with 12 endorsing moratoria, while local referendums in 12 states passed data center restrictions in 44 out of 45 instances. An AP survey revealed that approximately three-quarters of American voters oppose data center construction in their communities, and those voters appeared to have contributed to the so-called “blue wave” last week. This political backlash, compounded by high borrowing costs and bubble concerns, triggered a global tech sell-off early in the week. Though stocks later rebounded, the sharp volatility underscored growing investor skepticism that the AI buildout will continue unabated. Despite these possible state actions, President Donald Trump, a proponent of data centers, could still block any national restrictions on data centers proposed by a Democratic-controlled Congress.

• IEEE Spectrum covers new technical challenges that have complicated the European Union’s rollout of AI watermarking rules. Under Article 50 of the EU’s AI Act, developers of synthetic media must embed machine-readable markers to identify machine-generated text, as this AI Roundup covered this past August. However, some current watermarking techniques have struggled with short texts and those with simple human-made edits, making detection much more difficult. OpenAI recently introduced TextGrain, a watermarking system that treats entropy, the unpredictability of a model’s word choices, as a budget to balance watermark detectability against output quality. Internal evaluations have shown that while TextGrain achieves a 95% detection rate on passages of around 300 words, its accuracy drops to roughly 80% on 150-word passages, and falls below reliability thresholds for short answers. Furthermore, minor human revisions to the text severely degrade performance, as replacing 10% of words with synonyms lowers detection to 66%, and altering 25% of words causes detection rates to plummet to 17%. Because the EU has not established objective criteria for what constitutes robust detection as required by the AI Act, compliance remains uncertain. If a company’s watermarking process is noncompliant, it could be exposed to fines of up to 7% of its global annual revenue. Outside researchers also note that TextGrain cannot yet be validated independently because OpenAI has not released an open protocol or peer-reviewed documentation for TextGrain, unlike competing open frameworks such as Google DeepMind’s SynthID.

• The Chinese conglomerate Tencent has released its latest flagship AI model, according to Bloomberg. The open-weight system, named Hy4, introduces computer-use and visual perception capabilities across desktop interfaces, competing with domestic rivals such as Moonshot AI, DeepSeek, and Z.ai in coding and browser tasks. Overseen by former OpenAI researcher Yao Shunyu, Tencent designed Hy4 specifically for practical enterprise workflows, including financial analysis and game development, while filtering out offensive cybersecurity data to curb malicious misuse. Despite these domestic strides, the development team acknowledged that Hy4 still trails top proprietary Western models like those from OpenAI and Anthropic in complex mathematics and advanced software engineering. The rollout comes amid a broader reckoning in the Chinese AI industry, as the tech-heavy STAR50 stock index has dropped 40% and erased $480 billion in value in recent months. Tencent itself has seen its market valuation slide 35% to $460 billion this year as investors question the company’s mounting capital expenditure on chips and data centers, while facing potential export restrictions (as this AI Roundup has covered earlier this year) and brutal price wars across the Chinese technology sector. As firms race to release their latest and most advanced models to attract customers, it is likely that the market will continue to be incredibly volatile.

• The Washington Post reports on the backlash to AI-generated images on restaurant websites and menus. Several independent restaurants in Washington, D.C., have triggered consumer revolts after introducing unappetizing AI-generated photos in their businesses, such as those depicting warty fried chicken, leaking burritos, and anatomically incorrect human hands holding food. Diners have taken to social media to criticize these representations as misleading marketing that violates consumer trust and degrades neighborhood dining culture. Small business owners counter that severe economic pressures and plunging sales have forced them to seek affordable alternatives to professional photography, with one deli owner saying to The Washington Post that traditional photo shoots cost $200 per sandwich, making free or low-cost generative AI systems an attractive option. Local restaurants often rely heavily on authenticity and personal relationships, meaning that AI-generated photos can prompt immediate visceral rejection by their customer base that larger corporate chains rarely experience. While some businesses featured in The Washington Post have retracted the AI images and posted handwritten apologies to recover goodwill, others intend to keep using artificial intelligence tools to continue to save on marketing expenses.

• Nikkei Asia reports on a new AI-driven tunnel boring machine being tested in Japan. The construction company Taisei recently demonstrated the autonomous system on an 11-meter-tall, 1,400-metric-ton shield excavator during construction of a highway tunnel connecting inland roads with the Kagoshima port on the Japanese island of Kyushu. Steering the massive machine requires adjusting 30 peripheral hydraulic jacks to counteract unpredictable ground pressure and prevent surface collapses or water leakage. To replicate the intuition of veteran operators, Taisei trained the AI models on 33 kilometers of historical construction data gathered from 20 previous projects, producing 800 specialized operational models for specific projects. In a trial, the system dug a 4.5-meter segment within the planned 2.3-kilometer tunnel with a directional deviation of less than 2 millimeters, easily outperforming the 10-millimeter tolerance standard set for experienced human operators. Taisei has developed the automated technology over the past two decades to address critical labor shortages in the construction industry, mostly stemming from Japan’s rapidly aging workforce. The company plans to deploy the platform in future underground infrastructure projects, including stormwater storage facilities.