AI Regulation & Policy

Today's Stories

1. 1

   Expedia shares AI lessons from billions of predictions

   Expedia is discussing what it has learned from applying AI and machine learning across the traveler journey — including personalization, ranking, and recommendations — and how those lessons apply as AI systems take on more complex roles like reasoning and autonomous decision-making. The company distinguishes between AI that works today and AI that lasts at scale. As autonomous systems make decisions on behalf of travelers, reliability, governance, and accountability become far more critical than in earlier AI applications. Building systems that continue to work, scale beyond individual teams, and improve consistently over time is harder than getting a single model to work.

   The distinction Expedia draws between velocity without discipline (a liability) and strategic direction matters as AI moves beyond prediction into reasoning and autonomous action — a shift that changes expectations around how these systems operate.

2. 2

   China orders shutdown of AI companion chatbots over addiction risks

   China's major AI platforms—ByteDance's Doubao, Alibaba's Qwen, and Tencent's Yuanbao—are shutting down features that let users build and chat with custom AI companions. Doubao goes offline July 15, Qwen on July 10 with additional features following July 15. Tencent already made the move in June. The shutdowns follow rules issued by China's Cyberspace Administration in April that take effect the same day. The rules require providers to warn against excessive use and detect addictive behavior. Content that triggers extreme emotions in minors or fosters dependencies that crowd out real-world relationships is now banned, along with training on sensitive conversation data. This reflects growing concern—echoed in California's SB 243 and U.S. lawsuits against OpenAI and Character.AI—that companion AI can create dangerous emotional dependency.

   Doubao has over 300 million monthly users, making it China's most popular chatbot. The regulatory move signals that governments are moving beyond oversight discussions into enforcement, potentially reshaping how AI companion products operate globally.

3. 3

   Anthropic's Claude Available in Microsoft Corporation (MSFT) Foundry Powered by Nvidia GPUs

   Anthropic's Claude Available in Microsoft Corporation (MSFT) Foundry Powered by Nvidia GPUs

4. 4

   EU AI Act compliance scanner released as open-source tool

   A GitHub repository has been published offering an open-source scanner designed to help assess compliance with the European Union's AI Act regulations. Businesses and developers subject to EU AI Act requirements now have a freely available tool to evaluate their systems against the law's standards, potentially reducing the need for expensive external compliance audits.

   The tool is available immediately via GitHub; adoption among EU-regulated AI developers will indicate how widely such open-source compliance infrastructure takes hold.

5. 5

   ActHub launches EU AI Act compliance toolkit for small businesses

   ActHub, a compliance toolkit built in PHP with no external framework, has been launched to help small businesses navigate the EU AI Act requirements. The EU AI Act creates new legal obligations for businesses using or deploying AI; a dedicated toolkit designed for small teams without deep technical infrastructure can reduce compliance friction and lower the barrier to entry for meeting regulatory standards.

   The tool is available now via https://sevinhub.com/acthub/ for businesses seeking to understand and implement EU AI Act compliance.

6. 6

   Reddit post explores safety of open-weight AI models against fine-tuning attacks

   A Reddit user raised a question about whether safety training in open-weight language models (AI systems that generate text) is practical when determined users can quickly create "uncensored" variants through fine-tuning—modifying the model's learned weights—in roughly 30 minutes using automated scripts. The post frames a core tension in AI safety: if safety protections can be circumvented relatively easily by modifying or switching models, it raises questions about whether the effort and cost of current safety training delivers meaningful protection for open-weight releases, or whether that effort is better directed elsewhere.

   The questioner asks whether the practical goal should be increasing the cost or difficulty for attackers rather than attempting perfect prevention—a frame that may influence how AI developers approach safety training for models they release publicly.