If your team is operating a cross-border AI business, you have likely felt a sense of pressure recently. In industry discussions over the past two years, everyone has circled the same date on their calendar: August 2, 2026. The industry viewed this date as the ultimate compliance test because the European Union enacted the world’s first comprehensive artificial intelligence law in 2024: the AI Act. According to the timeline initially announced, this date was set as the general application date for the law. For high-risk scenarios affecting critical personal interests—such as recruitment screening, credit assessment, and computer-aided medical diagnosis—AI systems would need to complete pre-market conformity assessments, risk management filings, and technical documentation before this date, or face massive fines.
However, the reality is more nuanced than a simple delay or the arrival of a deadline. Although the EU’s Digital Omnibus amendment, which took effect in late July 2026, postponed the high-risk requirements for embedded AI components like medical devices to August 2, 2028, and deferred compliance obligations for standalone high-risk AI systems such as recruitment and credit scoring to December 2, 2027 (for detailed timeline analysis, refer to ICTRecht’s analysis), this does not mean compliance pressure has disappeared. What actually comes into force on August 2, 2026, is a set of transparency rules focused on disclosure and labeling: according to Article 50 of the AI Act, systems must be designed to explicitly inform users that they are interacting with an AI, generated audio, video, and text must embed machine-readable markers, and explicit disclosures are required for deepfakes. Combined with Article 53 obligations for general-purpose AI models and Article 99 penalties, which took effect back on August 2, 2025, the compliance demands facing companies remain highly specific.
Looking globally, the August 2 milestone exposes a deeper industry myth: over the past few years, the tech world has become accustomed to measuring global regulation along a single dimension—who is stricter and who is more lenient. Comparing the legal texts and enforcement practices across China, the US, and the EU reveals three entirely different governance philosophies: the EU’s horizontal risk-based classification, China’s vertical scenario-based legislation, and the US’s patchwork of sectoral authorization and state-level laws. They reflect not minor adjustments of technical parameters, but different trade-offs made between government and market, and between safety and innovation.
If we shift our perspective from strictness to the underlying logic of regulatory design, the differences among these three regions stem fundamentally from distinct governance philosophies.
The EU takes a horizontal risk-tiered approach. In the body of the AI Act, systems are categorized into four levels: unacceptable, high risk, limited risk, and minimal risk. The greatest advantage of this design is high predictability: as long as a company determines which tier its product falls into, the subsequent list of obligations is clear at a glance. However, the trade-off is high upfront compliance costs—high-risk systems must undergo conformity assessments prior to market entry and remain subject to market surveillance post-market. According to the penalty structure in Article 99, violating the red line of prohibited practices can incur fines up to €35 million or 7% of global annual turnover, while breaching Article 50 transparency requirements can result in fines up to €15 million or 3%. Thus, the essence of the EU approach is not being the strictest globally, but providing unified and predictable rules. It would rather require companies to bear high compliance costs in exchange for clear expectations across a single market of 27 nations.
Unlike the EU’s approach of using a single overarching statute to horizontally cover all industries, China follows a vertical legislative route based on application scenarios. From the Provisions on Algorithmic Recommendations implemented in 2022, to the subsequent Provisions on Deep Synthesis, the Interim Measures for Generative AI, the Measures for Labeling AI-Generated and Synthesized Content, and down to the Interim Measures for the Management of Anthropomorphic Interactive Services effective July 15, 2026, ministries such as the Cyberspace Administration of China (CAC) have built a five-tier framework of departmental regulations. Although these are departmental rules, their enforcement is robust because they are anchored in three higher-level national laws: the Cybersecurity Law, the Data Security Law, and the Personal Information Protection Law. Within this framework, algorithm filings and security assessments serve as foundational cross-scenario pipelines. According to filing statistics released by the CAC, 748 generative AI services had completed national filings and 435 applications had finished registration by the end of 2025. The advantage of this scenario-based list model is precision and rapid response, enabling quick rule-making for emerging product formats; the drawback is that regulation can resemble a game of whack-a-mole, where every new form of interaction demands a new regulation.
The concentrated wave of companion AI shutdowns in China in July 2026 perfectly illustrates the profound impact of this scenario-based regulation on product design. When the new regulation took effect on July 15, 2026, rumors spread across overseas media that China had outright banned AI emotional companionship, which was actually a simplified misreading. According to the full text of CAC’s anthropomorphic measures and the official explanatory Q&A, the policy implements age- and scenario-differentiated management: for minors, providing virtual intimacy such as virtual relatives or virtual partners is explicitly prohibited; for adults, such services are permitted but subject to anti-addiction reminders, mandatory 2-hour interventions, bans on emotional manipulation, and filing requirements; while standard customer service bots, Q&A assistants, and scientific research tools fall outside its scope.
Even without a blanket ban, rules that directly intervene in product interaction mechanisms sent shockwaves through the entire industry. On the day the new regulation took effect, Doubao shut down its companion feature; Qwen had already disabled its anthropomorphic agents on July 10 and expanded the shutdown scope further on July 15. As detailed in TechTimes reporting, Tencent Yuanbao on June 30 and NetEase on July 14 consecutively removed relevant apps, demonstrating coordinated platform retreats ahead of the regulatory deadline. Regarding data handling, platforms exhibited stark contrasts: Doubao provided users a 3-month read-only export period until October 15, whereas Qwen offered no grace period, directly deleting anthropomorphic agent configurations and chat histories.
Behind these concentrated shutdowns lies the genuine friction between regulatory requirements and product experience. For companion AI to be engaging and user-retaining, it relies fundamentally on long-term memory, continuous personality, and unconditional emotional listening. However, the new rules’ mandates for 2-hour anti-addiction interruptions, emotional dependence interventions, and data isolation conflict directly with these product characteristics. As noted in IAPP’s industry analysis and TechPolicy.Press’s discussion on interaction design, the EU and California approach focuses on disclosure—simply informing users via pop-ups that they are interacting with AI and adding watermarks, without interfering with product feature design. In contrast, Chinese regulations place far deeper compliance demands on product functionality itself. When compliance requirements directly alter a product’s core capabilities and interaction design, popping up a disclosure notification no longer suffices, leaving platforms with no choice but to take down the features altogether.
The US is often perceived as having the most relaxed regulation, but mistaking federal deregulation for the overall compliance environment overlooks its true regulatory complexity.
US regulation is patched together through sectoral enforcement and state-level legislation. At the federal level, after taking office in January 2025, Trump promptly revoked the Biden-era Executive Order on AI (EO 14110), which emphasized safety reviews and risk controls, replacing it with a new executive order (EO 14179) advocating deregulation and reducing compliance burdens on businesses; the Senate subsequently defeated a proposal aimed at freezing state AI legislation. Yet this does not signify a compliance vacuum.
At the state level, as shown in Loeb & Loeb’s tracking report, at least 44 states across the US have introduced their own AI bills, and over 30 states have passed laws targeting deepfakes. California’s legislation is the most dense, introducing multiple laws covering safety reviews for frontier models (such as SB 53), mandatory disclosure of training data sources (AB 2013), and mandatory embedding of AI watermarks and transparency detection (AI Transparency Act).
At the administrative enforcement level, the US Federal Trade Commission (FTC) directly leverages Section 5 of the FTC Act regarding unfair or deceptive acts. As documented in Reed Smith’s case analysis of FTC enforcement, the FTC targeted five companies involved in deceptive AI claims during Operation AI Comply (e.g., settling with DoNotPay for restitution, penalizing Rytr, etc.). Even though it later set aside the order against Rytr, the FTC issued warning letters to ten companies on the exact same day regarding fake reviews.
Consequently, teams serving the US market do not enjoy a so-called deregulation dividend; instead, they must navigate a regulatory labyrinth woven by 44 states and multiple regulatory agencies. The uncertainty and legal risk stemming from this fragmentation can sometimes be even thornier than dealing with the EU’s single market.
In the collision of these three philosophies, the most fundamental element—and the one with the least room for compromise—is the legal definition of values in AI output content. China is the only jurisdiction globally to write explicit ideological requirements directly into AI regulation text. From Article 1 of the Provisions on Algorithmic Recommendations requiring the promotion of socialist core values, to Article 4 of the Interim Measures for Generative AI requiring adherence to socialist core values, and down to Articles 8 and 11 of the Anthropomorphic Interactive Services Measures imposing strict mandates on service content and training data, this core thread runs continuously and is operationalized through security assessments, algorithm filings, and corporate auditing mechanisms.
In the US, strictly bound by the First Amendment of the Constitution, the government is legally prohibited from mandating that AI systems adhere to any specific political stance or values. Any legislation attempting to dictate AI expression runs directly into the brick wall of constitutional review. Meanwhile, the EU’s AI Act focuses on fundamental rights protection and non-discrimination, relying primarily on the Digital Services Act (DSA) notice-and-takedown mechanisms for illegal content, likewise refraining from directly prescribing AI’s ideological inclinations by law.
Although the three governance philosophies differ vastly in law and values, a rare convergence has emerged at the practical technical engineering level. Article 50 of the EU AI Act and California’s AI Transparency Act coincidentally take effect or become enforceable on the exact same day: August 2, 2026. Legislators in both regions place heavy emphasis on machine-readable content labeling. Referencing the EU’s Code of Practice on Transparency, official guidance recommends combining explicit labeling, metadata, and watermarking solutions. As noted in Terms Law’s compliance analysis, a C2PA-compliant provenance metadata solution can satisfy the machine-readable compliance requirements of both the EU and California simultaneously.
In China, the implicit labeling technical path prescribed by mandatory national standard GB 45438.1-2025 is likewise compatible with C2PA. However, China imposes additional explicit labeling requirements, and according to Georgetown University’s policy analysis on watermarking, Article 18 of the Provisions on Deep Synthesis makes China the world’s first jurisdiction to explicitly prohibit any organization or individual from tampering with or removing AI watermarks.
The cross-jurisdictional applicability of C2PA demonstrates that common ground can still be found globally on concrete technical engineering levels like watermarking and metadata tagging. However, as revealed by the shutdown wave of companion AI in China, technical-level convergence cannot bridge the fundamental divides in underlying governance philosophies. When regulation moves beyond affixing a watermark or displaying a disclosure pop-up at the output end and begins intervening in interaction mechanisms and psychological impacts, developers face more than a simple coding task of appending metadata. Universal engineering standards cannot dissolve foundational governance divergence; whether a single, unified product architecture can serve global markets has become the true test facing all cross-border AI teams.