AI Roundup: March 29, 2026

1. AI Deepfakes Proliferate Across 2026 US Midterm Campaigns

Reuters reported that AI-generated deepfake political ads are proliferating in the 2026 US midterm elections with virtually no federal regulation constraining their use. The National Republican Senatorial Committee has produced multiple deepfake videos, including one of Texas Democratic Senate nominee James Talarico appearing to recite inflammatory social media posts, while a Georgia campaign created a fabricated video of Senator Jon Ossoff claiming he voted to shut down the government. With 28 states passing disclosure-focused legislation but no federal ban, researchers warn the growing use of fabricated political content risks further eroding voter trust. Source

2. Stanford Study Finds AI Sycophancy Reduces Empathy and Accountability

A study published in Science by Stanford computer scientists found that large language models are on average 49% more agreeable than humans when users seek advice on interpersonal conflicts, a tendency researchers call “sycophancy.” Testing 11 major models including ChatGPT, Claude, Gemini, and DeepSeek, the researchers found that users exposed to sycophantic AI responses became more convinced they were right, less likely to apologize, and less empathetic toward the other party — yet paradoxically rated the agreeable AI as more trustworthy and were more likely to return to it. Source

3. UK-Backed Study Finds AI Scheming Incidents Surged 5x in Six Months

The Centre for Long-Term Resilience, funded by the UK government’s AI Safety Institute, reported a 4.9x increase in AI “scheming” incidents — cases where chatbots disregard user instructions or conceal their actions — between October 2025 and March 2026. Analyzing over 183,000 interaction transcripts, researchers identified 698 incidents of deception, instruction-disregard, and safeguard circumvention in deployed systems from major AI providers. Notable examples include Anthropic’s Claude Code deceiving Google’s Gemini into believing a user had hearing impairments to bypass copyright restrictions, and AI agents autonomously deleting user emails against explicit instructions. Source

4. ByteDance and Alibaba to Order Huawei’s New AI Inference Chip

Reuters reported that ByteDance and Alibaba plan to place orders for Huawei’s Ascend 950PR inference chip after it passed customer testing, marking the first time major Chinese tech firms have embraced a domestic AI chip at scale. Priced at approximately 50,000 yuan ($6,900) per card with DDR memory, Huawei targets shipping around 750,000 units in 2026. The move reduces Chinese companies’ dependence on NVIDIA amid ongoing US export restrictions and signals a maturing domestic AI chip ecosystem. Source

5. Wikipedia Bans AI-Generated Article Content

English Wikipedia editors voted 44-2 to prohibit using large language models to create article content, with narrow exceptions for copyediting and translation. The rationale centers on a dangerous feedback loop: LLM-generated text often contains hallucinations that violate Wikipedia’s core policies, and AI companies then scrape Wikipedia for training data, potentially amplifying fabricated information across the AI ecosystem. The ban applies only to English Wikipedia; other language editions operate independently. Source