On July 3, 2026, the National Cyberspace Administration released the "Internet Information Service Management Regulations (Revised Draft for Comments)," which for the first time added a dedicated chapter on "Intelligent Information Services." Compared to the "Interim Measures for the Management of Generative Artificial Intelligence Services" released in 2023, this revised draft requires AI service providers to disclose the basic principles of their technology and the sources of their training data, and clearly prohibits forcing users to use intelligent information services.

When large models casually claim that their pre-training data exceeds 20 trillion tokens yet remain vague about the sources, and when various apps generally exhibit a phenomenon of "forcing everything to be enhanced by AI" with default settings that are hard to turn off, the implementation of the new regulations signifies that our country's AI regulatory logic has fully shifted from "inclusive prudence" to "emphasis on transparency and manipulation prevention." For AI entrepreneurs and developers, this is no longer just a macro document, but a direct challenge regarding product design, compliance costs, and the survival of business models.

Increased Granularity of Regulation: Data Sources and Operating Mechanisms Must Be Disclosed

The granularity of AI regulation in our country is undergoing a leap. The "Interim Measures for the Management of Generative Artificial Intelligence Services" implemented in July 2023 only required "conducting training data processing activities in accordance with the law" at the data level, representing a baseline regulation. However, Article 56 of the current "Revised Draft" explicitly requires that providers of intelligent information services must disclose information about the basic principles of their relevant technologies, their purposes and significance, major operating mechanisms, and sources of training data.

This means that transparency has shifted from a soft advocacy to a hard entry threshold; AI products that do not disclose relevant information will face risks of being unable to be listed or being taken down.

Tracing 20 Trillion Tokens: The Compliance Ledger for Startup Teams

Mandatory disclosure of training data sources presents a fuzzy accounting problem for AI startup teams. Currently, mainstream large models in the country tend to emphasize "scale over source" in their data disclosures. For instance, the Tongyi Qianwen Qwen 2.5-Max reportedly has over 20 trillion tokens in its pre-training data, according to organizations like Zhiyuan Community; technical reports of models like Kimi also reference data covering web pages, code, etc. However, much of this information remains at the macro description level in technical reports, and there are very few cases disclosing specific data source lists and copyright filtering details aimed at end users. In early 2025, a test of seven mainstream large models by The Paper pointed out that privacy issues had become a common problem, with serious deficiencies in information disclosure transparency.

For small and medium entrepreneurs relying on open-source datasets or web crawlers, the implementation of the new regulations necessitates the establishment of a complete "data lineage" ledger. The data tracing and copyright filtering mechanisms covering hundreds of trillions of tokens require substantial legal and computational costs. If they cannot disclose or if their disclosures have flaws, products will face the risk of being taken down. This essentially raises the industry's entry barrier, eliminating teams lacking compliance infrastructure capabilities from the starting line.

Prohibition of Forced Use of AI: The Traffic Logic of Apps Forcing Features Becomes Ineffective

The draft explicitly prohibits the mandatory use of intelligent information services by users and strictly forbids using algorithms to interfere with online public opinion order. This provision directly targets the widespread issue of "forcing everything to have AI" in current internet products.

In April 2026, 18 large model manufacturers in the country jointly issued the "New Generation Artificial Intelligence Industry Function Specification Management Initiative," advocating for the "turn-offable AI principle," criticizing tactics like hidden switches and automatic restarts. This indirectly corroborates the widespread phenomenon of various apps defaulting to turning on AI helpers, setting multiple hidden switches, and silently collecting data to feed back into models.

The implementation of the new regulations will directly cut off large companies' ability to acquire AI interaction data through "indispensable pop-up windows." For AI entrepreneurs and developers, the logic of product design must fundamentally change. AI functionality can no longer be forced on users but must return to genuine user value. Product interfaces need to offer clear one-click off options and completely terminate background operations. The logic of acquiring traffic must shift from "forced exposure" to "value attraction," which raises higher requirements for the retention capability and core experience of AI products.

Significant Labeling and Prevention of Online Violence: Review Logic Must Be Added to Generation Interfaces

In terms of content governance, the draft requires significant labeling of generated synthetic content and stipulates that platforms must establish a database of characteristics of online violent information and provide users with protective functions such as blocking and prohibiting reproduction.

Combined with the "Measures for Labeling Artificial Intelligence Generated Synthetic Content" officially implemented in September 2025 and mandatory national standards, explicit and implicit labeling of AIGC content has become a hard requirement. This means that the AI generation chain must undergo deep technical modifications. For developers providing APIs, it is necessary to add content security review logic at the interface level and inject digital watermarks or metadata as implicit labels. Simultaneously, generated content must connect to the database of online violent information characteristics for pre-emptive interception; content governance is no longer a remedial action but part of product infrastructure.

As compliance requirements rise, the application scenarios for AI content detection tools are also expanding. They will transition from testing toys for geek players to auxiliary infrastructure for platform review and corporate compliance, helping to identify AI-generated content that is not labeled according to regulations, thus improving the content governance ecosystem.

From the assessment results, even leading large model products are still in a vague processing state when it comes to the detailed disclosure of data sources, lacking transparency. In terms of content labeling, although basic AIGC labeling has become widespread, technologies like implicit watermarks have not yet been comprehensively covered. Additionally, in terms of the ability to turn off AI functions, the industry generally faces the pain point of being unable to completely disable them; some integrated AI helpers remain in a default enabled state. Currently, the industry as a whole is still significantly distant from the new regulatory requirements, particularly in terms of the disclosure of data sources and the ability to disable AI functions where compliance gaps exist.

Compliance reviews need to be proactive at the product development stage, rather than being remedied by legal departments after launch. Entrepreneurs must incorporate data transparency and anti-manipulation capabilities into core evaluation indicators when choosing external AI tools or developing their models. Under the new compliance framework of transparency and anti-manipulation, only by accounting for compliance costs in the R&D foundation can AI products qualify for market entry.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。