NextAI+ Praxis--:----UTCThe five developments observed in this issue come from the United States, the European Union, the G7, Japan, and China. Although their regulatory approaches differ, they collectively point to several key capabilities required for enterprise AI implementation: model safety auditing, platform systemic risk assessment, protection of minors, sovereign model supply, and governance of AI agent identity, permissions, and tool use.
In the United States, Illinois SB315 pushes frontier model developers toward third-party audits. In the European Union, the Temu fine shows that platforms must prove that their recommendation systems and operational designs do not systematically amplify risk. The G7 has incorporated generative AI into its framework for protecting minors online. Japan is using Government AI to promote a unified public-sector AI operating environment and domestic model supply. China, meanwhile, has entered AI agent governance through standards for intelligent agent interconnection, covering identity, permissions, interaction, and tool invocation.
Overall, AI compliance is no longer merely a matter of legal documentation. It is becoming part of enterprise deployment architecture. Enterprises need to build capabilities in model vendor due diligence, risk assessment, log retention, permission control, tool invocation auditing, and multi-region governance in parallel in order to deploy AI safely, compliantly, and sustainably across different regulatory environments.
On May 27, 2026, the Illinois House of Representatives passed SB315, Senate Bill 315, formally titled the Artificial Intelligence Safety Measures Act. The bill has now been sent to the Governor for review and signature. The core of SB315 is to require large frontier AI developers to establish safety frameworks, publish transparency reports, and undergo annual independent third-party audits.
SB315 primarily targets large frontier model developers, namely companies with significant revenue scale that develop highly capable foundation models, rather than ordinary enterprises that use generative AI tools. Its regulatory focus is placed on the model supply side, especially frontier AI models that are most likely to generate systemic safety risks.
This indicates that AI governance in Illinois is shifting from “application-side risk” to “supply-side risk.” Regulators are not only concerned with how a given company uses AI, but also whether the most powerful models themselves have sufficient safety governance, capability evaluation, and risk control mechanisms.
SB315 requires large frontier AI developers to establish, implement, publish, and annually update a frontier AI framework. This framework must cover catastrophic risk assessment, risk mitigation measures, cybersecurity, internal governance, third-party evaluation, and risks arising from developers’ internal use of frontier models.
The focus here is not on “publishing a set of AI principles,” but on forming an executable governance system. Enterprises need to explain how they identify high-impact risks, how they test model capabilities, how they reduce the possibility of model misuse, how they protect model weights and infrastructure, how they handle safety incidents, and who within the organization is responsible for approval, escalation, and response.
Before releasing a new or materially modified frontier model, developers need to disclose a transparency report and explain the high-impact risks that the model may pose, as well as the corresponding mitigation measures.
This indicates that model launch is no longer merely a product decision. It is also beginning to be incorporated into a framework of risk disclosure and regulatory visibility.
Large frontier AI developers must not only claim that they comply with their safety framework, but also have a third-party audit organization confirm whether they have actually implemented it. Capitol News Illinois reported that the bill requires developers to hire third-party auditors to verify their compliance, while amendments further clarify third-party qualifications, audit content, and mechanisms for protecting trade secrets.
This means AI safety is entering a stage of “external verification.” In the past, the logic was that companies disclosed their own safety measures. The logic of SB315 is that companies cannot simply “grade themselves.” Instead, an external party must examine whether their safety commitments have been genuinely implemented. This imposes higher requirements on model developers’ testing records, logs, approval materials, red-team testing, incident response, and internal controls.
On May 28, 2026, the European Commission imposed a EUR 200 million fine on Temu under the Digital Services Act, or DSA, the EU’s core regulation governing online platforms, search engines, online marketplaces, and other digital intermediary services. The reason was that Temu failed to adequately identify, analyze, and assess the systemic risks arising from the sale of illegal products on its platform and the resulting harm to EU consumers.
Temu had already been designated by the European Commission as a Very Large Online Platform, or VLOP, on May 31, 2024, because it had more than 45 million monthly active users in the EU. Once designated as a VLOP, the platform no longer bears only general e-commerce compliance obligations. It must also fulfill higher-intensity systemic risk assessment and risk mitigation obligations under the DSA.
The key point is that the EU is not only looking at whether the platform “sold illegal goods,” but also whether the platform has the capability to identify and control the systemic risks of large-scale dissemination of illegal products. For cross-border e-commerce platforms, content platforms, social platforms, and app stores, once user scale reaches the VLOP threshold, platform governance upgrades from ordinary compliance to systemic risk governance.
The European Commission stated that Temu’s 2024 risk assessment relied mainly on general risk information about the e-commerce industry as a whole, rather than specific evidence from Temu’s own services, including public reports, tests, and investigation findings.
This is one of the most important compliance signals in the case: risk assessments under the DSA cannot be templated, industry-level, or paper-based documents. They must be based on the platform’s own data and actual operational evidence. A platform needs to explain its own merchant structure, product categories, recommendation mechanisms, complaint data, sampling inspection results, violation recurrence rates, and cases of consumer harm. It cannot simply state that “counterfeit goods, unsafe products, and fraud risks are common across the e-commerce industry.”
The European Commission specifically pointed out that Temu had not adequately assessed how the design of its services affected the risk of illegal product dissemination, including its recommendation systems and product promotion programs involving affiliated influencers.
This means EU regulation is pushing platform governance from outcome-based regulation toward mechanism-based regulation. The issue is not only whether illegal products appear on the platform, but also whether the platform’s ranking algorithms, recommendation logic, promotional resources, influencer cooperation, and traffic distribution mechanisms give greater exposure to high-risk products. For enterprises, this brings recommendation governance to the forefront of compliance.
The DSA requires large platforms to identify, analyze, and mitigate systemic risks. In the European Commission’s official DSA explanations, it clearly states that large platforms must identify and analyze broad risks and take measures to reduce them. For marketplaces, the DSA also requires greater merchant transparency and reasonable efforts to randomly check products sold on the platform or to adopt new technologies to improve product traceability.
The core change reflected in this penalty is therefore that platform compliance is no longer merely about “removing content or products after receiving complaints.” It must be moved upstream to product admission, merchant verification, risk sampling, recommendation system design, influencer promotion management, and evidence-based risk assessment. This is particularly important for platforms that rely on algorithmic recommendation, low-priced products, high-frequency product launches, and cross-border merchant networks.
On May 29, 2026, digital and technology ministers from the Group of Seven, or G7, reached a set of common principles for protecting minors online. These principles require digital service providers to incorporate child safety, privacy protection, age assurance, recommendation system risk control, and generative AI risk governance into product design and operational processes.
The G7 common principles emphasize that digital service providers need to implement effective risk management, risk assessment, and risk mitigation measures and embed them into a safety-by-design approach. Therefore, child safety should not be a patch added after a product is launched. It should be incorporated at the stages of product design, feature development, interface interaction, and default settings.
This means enterprises cannot rely solely on reporting, complaints, and post-event takedown mechanisms. For products aimed at minors or likely to be used by minors, protection mechanisms need to be designed in advance in areas such as default privacy settings, content visibility, interaction permissions, contact from strangers, push mechanisms, addiction control, and risk warnings.
The G7 principles propose the adoption of robust, reliable, and privacy-protective age assurance solutions to support age-appropriate online experiences and prevent minors from accessing age-restricted digital services.
The implication is that age identification is no longer merely a compliance gateway. It is a prerequisite for differentiated product governance. If a platform cannot determine whether a user is a minor, it will be difficult to implement tiered governance for content recommendation, advertising display, private messaging, AI chat, parental controls, and sensitive functions. At the same time, the G7 emphasizes privacy protection: age assurance must not become a new source of risk through excessive collection of identity documents, facial information, or children’s data.
The G7 principles explicitly state that minors’ accounts should have a high level of privacy and safety protection, and that recommendation system design should reduce excessive online engagement.
This shows that regulatory attention is expanding from “what content appears on the platform” to “how the platform designs attention mechanisms.” Infinite scrolling, autoplay, aggressive push notifications, personalized recommendations, reward mechanisms, and interaction incentives were previously often treated as growth tools. In scenarios involving minors, however, these designs may be reassessed as risk factors affecting physical and mental health, sleep, attention, self-esteem, and cognitive development.
The G7 ministerial statement specifically notes that generative AI, especially chatbots, may exacerbate risks faced by minors. It specifically mentions issues such as AI-generated child sexual abuse material, non-consensual intimate images, deepfakes, manipulative simulated interactions, and sexualized or exploitative content.
This indicates that online child protection is no longer only a traditional platform content governance issue. It has entered the stage of AI interaction governance. For AI companions, educational chatbots, customer service bots, virtual characters, game NPCs, and social AI products, the key issue is not merely blocking sensitive words, but controlling whether AI may form inappropriate dependence, inducement, deception, sexualized interaction, or dangerous advice in relation to minors.
On May 29, 2026, Japan’s Digital Agency issued an advance notice that it plans to solicit Japanese domestic large language models for use in fiscal year 2027 for Government AI “Gennai” in November 2026, and it disclosed the evaluation and testing methods in advance. This shows that Japan’s government AI strategy is moving beyond “internal government use of generative AI” toward “supporting domestic foundation model supply capacity.”
On May 28, 2026, Japan’s Digital Agency had just announced a large-scale demonstration of Government AI “Gennai” for approximately 180,000 government employees across all ministries and agencies. On May 29, 2026, the Digital Agency further announced the upcoming public solicitation of domestic foundation models for fiscal year 2027. This timing indicates that the Japanese government is not merely promoting an AI tool. It is simultaneously building a model supply system for Government AI.
For the public sector, the source of the model itself is a governance issue, because it relates to data security, language adaptation, administrative trustworthiness, supply chain resilience, and long-term bargaining power.
Japan’s Digital Agency clearly stated that Government AI particularly requires domestic foundation models suitable for Japanese vocabulary and expression and respectful of Japanese culture and values. At the same time, active government use of models developed by domestic companies and research institutions helps support domestic development of trustworthy AI.
This indicates that the solicitation is not simply a matter of “preference for domestic procurement.” Rather, it links language adaptation, cultural adaptation, trustworthy AI, and public-sector trust together. In government scenarios, models need to understand Japan’s administrative context, policy expression, legal texts, honorific systems, and public service communication practices.
The Digital Agency stated that the previous round of solicitation from December 2025 to January 2026 received 15 applications and that evaluation and validation contracts for fiscal year 2026 have already been signed with five companies. For fiscal year 2027, the government plans to conduct paid government procurement and will take into account the evaluation and validation results of the five companies’ models from fiscal year 2026.
The Japanese government is creating stable demand for domestic AI through real administrative scenarios and government procurement. It is not only supporting model development through subsidies. Instead, it is bringing models into the government use environment, subjecting them to feedback from administrative settings, and then incorporating evaluation results into subsequent procurement. This path connects model development, scenario validation, government demand, and industrial cultivation.
The Government AI page of Japan’s Digital Agency shows that “Gennai” has, on the one hand, been deployed as a government AI use environment across ministries and agencies. On the other hand, it is also advancing support for domestic large language model development, preparation of common government datasets, and development of advanced AI applications. The same page also lists existing information on cooperation with OpenAI to materialize Government AI, as well as arrangements for domestic LLM solicitation and domestic model trials.
This shows that Japan’s approach does not completely exclude overseas models. It is closer to multi-model governance, combining different models across different tasks, risk levels, and data scenarios. The key point is that the government may use globally leading model capabilities while simultaneously cultivating alternative, evaluable, and procurable domestic model options.
On May 22, 2026, the State Administration for Market Regulation and the Standardization Administration of China approved and issued eight national standardization guiding technical documents, including Artificial Intelligence — Intelligent Agent Interconnection — Part 1: General Architecture. Among them, seven directly focus on AI intelligent agent interconnection, covering general architecture, identity codes, identity management, intelligent agent description, intelligent agent discovery, intelligent agent interaction, and tool invocation. This indicates that China’s AI governance is beginning to extend from model and content governance to governance over AI agents’ identity, permissions, collaboration, and execution behavior.
In the past, enterprise discussions of AI governance focused mainly on whether models generated illegal content, leaked private information, exhibited bias, or produced hallucinations. Intelligent agents are different. They do not merely output text. They can execute tasks based on objectives, invoke tools, connect to external systems, and collaborate with other agents.
The core change brought by these documents is that the governance object is beginning to expand from model outputs to intelligent agents’ identity, interaction, and execution behavior. For enterprises, the future question will not only be “whether the model’s answer is correct,” but also “which agent executed the task, what tools it invoked, whether it exceeded authorization, whether logs were retained, and who bears responsibility.”
The GB/Z 185 series specifically includes two parts on “identity codes” and “identity management.” Once an intelligent agent enters an enterprise system, it cannot remain a vague “AI assistant.” It should have a clear identity, similar to an employee account, service account, or API key.
The governance focus here is that an intelligent agent must have an identity before permissions, auditing, and accountability can be discussed. If an enterprise cannot identify which agent initiated a certain operation, it cannot determine whether that agent had permission to access a database, invoke a business system, or trigger an approval process. Nor can it conduct accountability or traceback after an incident occurs.
This is consistent with the directions proposed in the Implementation Opinions on Standardized Application and Innovative Development of Intelligent Agents, including intelligent agent digital identity management, retrieval and discovery, and capability declaration.
Intelligent agent description and intelligent agent discovery appear on the surface to be technical issues of interconnection and interoperability. In essence, however, they are also governance issues. In a multi-agent ecosystem, before one intelligent agent is invoked by other systems, the invoking party needs to know what it can do, what it cannot do, which scenarios it is suitable for, what its interfaces are, and where its risk boundaries lie.
Therefore, intelligent agents cannot be treated merely as black-box services to be invoked. They need understandable capability descriptions and usage boundaries. For enterprises, this is similar to establishing a “capability profile” for each agent: what data it can access, what tasks it can execute, what tools it can invoke, which business scenarios it serves, whether it involves sensitive permissions, and whether human confirmation is required.
Without intelligent agent description and discovery mechanisms, multi-agent collaboration can easily become an uncontrollable automation network: one agent invokes another agent, which then invokes tools or databases, ultimately leaving the enterprise unable to explain the task chain and accountability boundaries.
GB/Z 185.7-2026 focuses on “intelligent agent tool invocation,” which is one of the most sensitive areas in enterprise deployment. Tool invocation means that an intelligent agent does not merely generate text, but invokes external APIs, databases, software tools, office systems, transaction systems, code execution environments, or other business systems to complete tasks.
The key point is that the main risk of AI agents lies not in what they can say, but in what they can do. Once an intelligent agent can send emails, modify code, query customer data, place orders, generate contracts, invoke payment interfaces, or modify system configurations, the enterprise must establish permission controls, approval mechanisms, invocation logs, anomaly interception, and rollback capabilities.
The Implementation Opinions on Standardized Application and Innovative Development of Intelligent Agents also propose clarifying the boundaries of autonomous decision-making by intelligent agents, ensuring that execution operations do not exceed the scope of user authorization, and studying safety management for permission management, behavior control, model access, application programming interface invocation, and the use of extended tools. This indicates that “tool invocation governance” is becoming a key component of China’s intelligent agent governance system.
Cite as · AI Governance Weekly · 4 June 2026
If you want both columns delivered together, four times a year, in one quiet email — leave an address. Otherwise just bookmark this page.