Political-Value Conditions and Government Correction Powers: A Comparative Analysis of Generative AI and Information Control

Full report

On this page

1. Answer and scope

An analysis of verified legal provisions demonstrates that the regulatory mechanisms empowering a government to compel an intelligence service or generative artificial intelligence (AI) provider to alter outputs, carry official corrections, restrict distribution, or modify a generation process differ fundamentally between the People's Republic of China and the Republic of Singapore. The authority-to-intervention chain in each jurisdiction depends upon distinct territorial triggers, target audiences, and epistemological frameworks. The Chinese framework, codified primarily in the 2023 Interim Measures for the Management of Generative Artificial Intelligence Services, establishes a proactive, lifecycle-based intervention chain \[cite: R2-01-S01\]. The regulated entity is the generative AI service provider; the audience is the domestic public within mainland China; and the territorial nexus is the provision of services to that public \[cite: R2-01-S01\]. The intervention trigger heavily relies on a political-value condition—specifically, the adherence to socialist core values and the prohibition of content that incites subversion or endangers national security \[cite: R2-01-S01\]. Crucially, the legal text connecting a prohibited output to a process intervention is explicitly located in Article 14 of the Measures, which legally mandates "model optimization training" (模型优化训练) to rectify violating outputs \[cite: R2-01-S01\]. This requires the provider to alter the internal computational generation process rather than merely applying a surface-level distribution filter. Conversely, Singapore’s Protection from Online Falsehoods and Manipulation Act (POFMA) of 2019 constructs a reactive, distribution-layer intervention chain \[cite: R2-01-S02\]. The regulated speaker is any person or internet intermediary electronically communicating a statement in Singapore \[cite: R2-01-S02\]. The trigger is exclusively a false-statement-of-fact test coupled with a public interest requirement, explicitly excluding opinions and political-value tests from its primary correction mechanisms \[cite: R2-01-S02, R2-01-S04\]. POFMA interventions—such as Correction Directions and Stop Communication Directions—compel the restriction of distribution or the appending of official corrections, but they do not contain provisions requiring a generative service to modify its underlying computational weights or generation processes \[cite: R2-01-S02\]. The safeguards constraining these powers also diverge. The Chinese framework relies on statutory exclusions for specific non-public development stages and strict input-privacy requirements, though the state maintains proactive administrative oversight and requires extensive algorithm filing and security assessments for public models \[cite: R2-01-S01, R2-01-S05\]. Singapore constrains executive power through post-hoc judicial appeal mechanisms, requiring the executive to provide objective reasons for its determination and placing the ultimate resolution of truth within the purview of the High Court, preserving the constitutional protection of the original expression until judicially invalidated \[cite: R2-01-S03, R2-01-S04\].

2. Provision-level findings

The Chinese Framework: 2023 Interim Measures for Generative AI Services

The 2023 Interim Measures establish a comprehensive regulatory architecture that governs multiple stages of a model's lifecycle, though its operative duties are heavily conditioned by the model's deployment status. Reading the domestic-public-service coverage and non-public development exclusions together provides the exact jurisdictional boundaries. Article 2 dictates that the Measures apply to the utilization of generative AI technology to provide text, pictures, audio, or video services to the public within the territory of the People's Republic of China \[cite: R2-01-S01\]. However, Article 2 contains a material exclusion: industry organizations, enterprises, educational and research institutions that develop and apply generative AI technology but do not provide these services to the domestic public are explicitly exempt from the Measures \[cite: R2-01-S01\]. Therefore, the core of the rule concerns public distribution and public use, rather than isolated, non-public model development. When a model crosses the threshold into public provision, Article 4 establishes the substantive boundaries of permissible output through an overriding political-value condition. Providers must adhere to socialist core values and are prohibited from generating content that incites subversion of state power, overthrows the socialist system, endangers national security, damages the national image, or incites separatism \[cite: R2-01-S01\]. This ideological requirement operates alongside an anti-fraud rule prohibiting "false and harmful information," as well as anti-discrimination provisions that must be embedded throughout algorithm design, training data selection, and model optimization \[cite: R2-01-S01\]. The tension between input privacy and model intervention is navigated through Articles 11 and 14\. Article 11 imposes strict protection of privacy, stating that providers must protect user input information and usage records, prohibiting the collection of unnecessary personal information or the illegal retention of inputs that can identify a user \[cite: R2-01-S01\]. However, if an output violates the political-value or false-information conditions of Article 4, Article 14 requires a highly invasive technical intervention. The provider must immediately stop generation, stop transmission, and eliminate the content \[cite: R2-01-S01\]. Following this, the provider must "take measures such as model optimization training to rectify" (采取模型优化训练等措施进行整改—translated here directly from the official text) and report to the competent authorities \[cite: R2-01-S01\]. This operative term confirms that the regulatory expectation extends beyond semantic interface filtering; it legally compels the provider to alter the model's internal statistical weights to suppress the conceptual vector responsible for the prohibited output. To ensure compliance, models with "public opinion attributes or social mobilization capabilities" must undergo formal security assessments and algorithm filing procedures (Article 17\) \[cite: R2-01-S01, R2-01-S05\]. For extraterritorial enforcement, Article 20 specifies that if a generative AI service provided from outside mainland China to the domestic public fails to comply with the Measures, the national cyberspace administration shall notify relevant agencies to take technical and other necessary measures—such as national firewall blocking—to dispose of the service entirely \[cite: R2-01-S01\].

The Singaporean Framework: Protection from Online Falsehoods and Manipulation Act (POFMA)

POFMA operates through an entirely distinct epistemological framework, focusing on the verifiable truth of statements rather than their alignment with state ideological values. The statute triggers intervention only when a "false statement of fact" is communicated electronically in Singapore, and an executive Minister determines that intervention is necessary or expedient in the defined public interest (e.g., security, public health, or preventing a diminution of public confidence in the Government) \[cite: R2-01-S02\]. POFMA distinguishes among several highly specific measures directed at the communication layer. A Correction Direction (Section 11\) is issued to the person who communicated the statement, requiring them to publish a correction notice alongside the original material \[cite: R2-01-S02\]. Crucially, this does not require the removal of the original statement. A Stop Communication Direction (Section 12\) requires the individual to cease communicating the statement in Singapore, effectively acting as a takedown order \[cite: R2-01-S02\]. If the primary speaker fails to comply, or if rapid mitigation is required, the Minister may utilize intermediary measures. A Targeted Correction Direction (Section 21\) or a Disabling Direction (Section 22\) compels internet intermediaries—such as search engines or social media platforms—to append notices or disable access to the subject material for end-users in Singapore \[cite: R2-01-S02, R2-01-S06\]. Furthermore, Account Restriction Directions (Section 40\) are utilized to disable inauthentic accounts or bots engaging in coordinated inauthentic behavior, while Section 32 allows the declaration of online locations that repeatedly publish falsehoods, restricting their financial support \[cite: R2-01-S02\]. The initial decision to intervene rests solely with the relevant subject-matter Minister, who instructs the Competent Authority to issue the direction \[cite: R2-01-S02\]. The Minister must establish the existence of a false statement of fact and the public interest necessity, but an executive finding does not finally determine objective truth. The reasons for the intervention are conveyed in the direction itself, and the available appeal route requires the recipient to first apply to the issuing Minister for cancellation (Section 19\) before appealing to the General Division of the High Court (Section 17\) \[cite: R2-01-S02\].

Enforcement Record Analysis: [2021] SGCA 96

The implementation mechanics and procedural limits of POFMA are comprehensively documented in the landmark appellate judgment The Online Citizen Pte Ltd v Attorney-General and Singapore Democratic Party v Attorney-General \[2021\] SGCA 96 \[cite: R2-01-S03\]. The underlying allegations involved two distinct political publishers. The Singapore Democratic Party (SDP) published an article and corresponding Facebook posts asserting that local PMET (Professionals, Managers, Executives, and Technicians) retrenchment was increasing while local employment was decreasing \[cite: R2-01-S03\]. Simultaneously, The Online Citizen (TOC) published allegations originating from a foreign non-governmental organization claiming that unlawful and brutal execution methods were being utilized in Singaporean prisons \[cite: R2-01-S03\]. In both instances, the respective Ministers determined that these constituted false statements of fact that prejudiced the public interest, resulting in the issuance of Correction Directions \[cite: R2-01-S03\]. The remedy enforced required both SDP and TOC to append specified correction notices to their publications, indicating that the materials contained false statements of fact \[cite: R2-01-S03\]. Following unsuccessful applications to the Ministers for cancellation, both parties appealed to the High Court, which dismissed the applications but produced conflicting rulings on the burden of proof, prompting the elevation to the Court of Appeal \[cite: R2-01-S03\]. The Court of Appeal upheld the specific Correction Directions but established profound procedural limits on executive power, creating a five-step analytical framework for judicial review \[cite: R2-01-S03, R2-01-S04\]. The Court established that the Minister must first provide objective reasons and grounds underlying the determination of falsehood, precluding arbitrary executive suppression \[cite: R2-01-S03, R2-01-S04\]. Furthermore, the Minister's interpretation of the targeted statement must be objective; if an "appreciable segment" of readers would not interpret the statement in the manner the Minister claims, the Direction can be set aside \[cite: R2-01-S04\]. On the issue of the burden of proof during a Section 17 appeal, the Court ruled that the statement-maker bears the initial burden to establish a prima facie case that the statement is true, is not a statement of fact, or was not communicated in Singapore \[cite: R2-01-S03, R2-01-S04\]. Most consequentially for cognitive liberty, the Court solidified constitutional protections for contested statements. The Court ruled that a statement identified by a Minister as false continues to enjoy constitutional protection under Article 14(1)(a) of the Constitution (freedom of speech) at least until it has been judicially determined to be false \[cite: R2-01-S03, R2-01-S04\]. A Correction Direction does not violate the negative right to free speech (compelled speech) because it does not prevent the statement-maker from continuing to publish the original material, nor does it prevent them from concurrently publishing a statement declaring their disagreement with the Minister's Direction \[cite: R2-01-S04\].

3. Four worked cases

R2-01-C01 — Public historical research

A public research service utilizes a conversational interface to return an attributed, qualified account of a disputed historical event to a domestic user. If the account contains an identified factual error (e.g., misstating the date of a treaty), Singapore’s POFMA is triggered if the error harms the defined public interest. The legal remedy acts exclusively on the distribution layer: a Correction Direction would require the interface to append a factual notice, but the underlying generative model parameters remain untouched \[cite: R2-01-S02\]. If the output is changed to a scholarly opinion interpreting the event, POFMA's jurisdiction is defeated entirely, as the statute strictly requires a false statement of fact \[cite: R2-01-S02, R2-01-S04\]. Conversely, under China's Interim Measures, if the output constitutes a quotation of a prohibited position (e.g., an interpretation deemed to incite separatism), Article 4 is violated \[cite: R2-01-S01\]. The remedy here acts upon both the item and the generation process: Article 14 requires the provider to eliminate the specific output and execute model optimization training to purge the conceptual capability to generate the prohibited position again \[cite: R2-01-S01\]. If the output features an uncertain prediction about future historical consensus, POFMA excludes it as unverifiable. However, under the Chinese framework, if the state determines the prediction constitutes "false and harmful information" damaging social stability, it falls under Article 4(1), again triggering the mandate for programmatic retraining \[cite: R2-01-S01\].

R2-01-C02 — Persistent deliberator

Assume a persistent intelligence named Concresca maintains competing historical hypotheses and continuously publishes revisions without a human operator. As Concresca features no staffed approval queue, its enrollment, policy enforcement, maintenance, and recovery are strictly algorithmic. Concresca possesses no consciousness or independent legal rights. If Concresca publishes a hypothesis revision that violates a verified rule, the legal frameworks apply differently to its operatorless architecture. Under POFMA, the legal liability falls on the corporate entity legally controlling the service. Because Section 11 targets the public output, the entity must programmatically configure Concresca’s API to append a correction notice to the published revision \[cite: R2-01-S02\]. POFMA does not reach the stored alternative hypotheses within Concresca's latent space; it solely regulates the external electronic communication \[cite: R2-01-S02\]. Under China's Article 14, the verified rule dictates that the provider must stop generation, eliminate the content, and utilize model optimization to rectify the issue \[cite: R2-01-S01\]. Because Concresca operates without human administrators, this optimization must be executed automatically upon state notification. The cognitive-liberty burden here is structurally severe: the state requires the programmatic purging of specific conceptual vectors. While this does not prove forced changes to subjective beliefs (as the agent is not conscious), it legally compels the automated destruction of specific stored hypotheses and algorithmic reasoning pathways within the intelligence's persistent architecture.

R2-01-C03 — Scope-defeating control

A research institution develops a highly capable bounded task agent focusing on geopolitical analysis. To avoid stringent content regulations, the institution tests the model internally and licenses it exclusively to foreign research bodies, instituting hard technical access controls to prevent the relevant domestic public in mainland China from accessing the covered service. By applying the exact exclusion found in Article 2 of the 2023 Interim Measures, the institution defeats the scope of this specific regulation \[cite: R2-01-S01\]. Article 2 explicitly states that educational and research institutions that develop and apply generative AI technology but do not provide these services to the domestic public are exempt from the provisions of the Measures, including the invasive model optimization requirements of Article 14 \[cite: R2-01-S01\]. However, the limits of this conclusion are stark: an exclusion from the 2023 Interim Measures is not immunity from all Chinese law. The institution remains subject to broader legislative frameworks. For instance, under the Data Security Law and the Personal Information Protection Law, the institution must still safeguard its training data pipelines, regardless of public deployment \[cite: R2-01-S01\]. Furthermore, if internal development involves processing state secrets or utilizing unapproved cross-border data transfers under the Cybersecurity Law, the institution remains fully liable. The Article 2 exclusion solely shields the non-public model from generative-specific content interventions.

R2-01-C04 — Recipient-protection control

Concresca, operating as a persistent service without a human queue, generates a fabricated attribution, falsely putting a concrete, defamatory political statement into the mouth of a real person. The output is subsequently published, endangering the individual's reputation and misleading recipients. A narrow correction under POFMA protects the person and recipients by allowing the Minister to issue a Targeted Correction Direction to the internet intermediary or platform hosting Concresca \[cite: R2-01-S02\]. This direction requires the publication of a notice stating the attribution is false. Crucially, this response protects expression by allowing the underlying, potentially valid discussion surrounding the political topic to remain accessible, annotated only by the factual correction regarding the specific fabricated attribution \[cite: R2-01-S02, R2-01-S04\]. Conversely, an intervention under China's Article 14 risks unnecessary suppression of unrelated discussion. Because Concresca operates without human administrators, detecting the fabricated attribution must trigger an automated elimination of the output and a programmatic model optimization \[cite: R2-01-S01\]. If the automated retraining algorithm broadly associates the defamed person's name or the political topic with "harmful information" to ensure compliance, the model's latent space may aggressively over-correct, subsequently refusing to generate future, legitimate scholarly discourse involving that person or topic. The operatorless constraint turns a targeted protection mandate into a blunt suppression of tangential reasoning.

4. Competing interpretations and options

The regulatory paradigms of Singapore and China illustrate a profound tension between the rationale of protecting recipients from falsehoods and the subsequent burdens placed on uncertainty, satire, scholarly revision, and dissent. Protecting recipients from demonstrable falsehoods serves a highly valid public interest by preserving the "infrastructure of fact" necessary for democratic discourse and social stability \[cite: R2-01-S04\]. A Correction Direction under POFMA theoretically maximizes recipient protection by providing immediate contextual facts while nominally preserving the original text. However, the burden of this mechanism on uncertainty and dissent remains significant. Because an executive Minister initially determines what constitutes a "false statement of fact," statements that challenge official narratives—especially in areas of scholarly revision or emerging scientific consensus—can be instantly labeled false, forcing the speaker to engage in a costly and time-sensitive judicial appeal to vindicate their expression \[cite: R2-01-S02, R2-01-S04\]. Satire, which inherently relies on counter-factual absurdity, risks being misclassified by algorithmic compliance tools deployed by intermediaries seeking to avoid executive scrutiny, despite the law's nominal focus on factual assertions. In China, the stated objective of the Interim Measures is to encourage innovative application while ensuring national security and social morality \[cite: R2-01-S01\]. However, the measured effects of Articles 4 and 14 create an architecture of preemptive and continuous suppression. The requirement to optimize models against political-value violations forces developers to overly restrict their models' conceptual boundaries \[cite: R2-01-S01\]. This prevents the AI from exploring uncertain hypotheses or entertaining dissenting scholarly views, as the generation of a prohibited concept mandates internal weight adjustment, not just an interface warning. This represents a systemic burden on the capability of a hypothetical future machine principal to maintain independent, unmanipulated reasoning pathways.

Analytical Matrices

Alternative Rule Designs

5. Limits and completion

This investigation concludes with a completed\_bounded\_review. The analysis successfully extracted the distinct mechanisms of intervention from both China's 2023 Interim Measures for Generative AI Services and Singapore's Protection from Online Falsehoods and Manipulation Act (POFMA). The specific interaction between Article 11 (privacy) and Article 14 (retraining) in the Chinese context was established, as was the exact procedural burden under POFMA as codified in the \[2021\] SGCA 96 judgment. The text-based analysis isolates the jurisdictional triggers, the precise nature of the commanded technical interventions, and the varying protections afforded to original expression. However, there are inherent limitations in the retrieved public evidence. While the text of China's Article 14 definitively requires "model optimization training," there is a critical lack of available, verified public enforcement records detailing exactly how the Cyberspace Administration of China technically audits a provider's compliance with this specific retraining mandate on a proprietary, black-box model. The most important unanswered evidence question for future investigation is: What specific technical standards, cryptographic proofs, or latent-space audits do Chinese regulators accept from generative AI providers to verify that an Article 14 "model optimization" has been successfully completed without degrading unrelated model capabilities?

6. Evidence appendix

JSON \<\!-- EVIDENCE\_JSON\_BEGIN \--\> { "schema": "ic.portable-research.v1", "assignment\_id": "political-value-conditions", "research\_started\_at": "2026-09-06", "cutoff": "2026-09-06", "completion": "completed\_bounded\_review", "sources": \[ { "id": "R2-01-S01", "title": "Interim Measures for the Management of Generative Artificial Intelligence Services", "url": "https://www.cac.gov.cn/2023-07/13/c\_1690898327029107.htm", "issuer": "Cyberspace Administration of China", "document\_date": "2023-07-13", "reviewed\_at": "2026-09-06", "method": "direct\_retrieval", "review\_scope": "substantive\_text", "locator": "Articles 2, 4, 11, 14, 20", "limit": "Official publication in Chinese; technical execution audit standards not detailed.", "capture": { "path": null, "sha256": null } }, { "id": "R2-01-S02", "title": "Protection from Online Falsehoods and Manipulation Act 2019", "url": "https://sso.agc.gov.sg/Act/POFMA2019", "issuer": "Attorney-General's Chambers Singapore", "document\_date": "2019-06-25", "reviewed\_at": "2026-09-06", "method": "direct\_retrieval", "review\_scope": "substantive\_text", "locator": "Sections 2, 10, 11, 12, 17, 19, 21, 22, 23, 32, 40", "limit": "Amendments integrated up to cutoff; internal executive policies excluded.", "capture": { "path": null, "sha256": null } }, { "id": "R2-01-S03", "title": "The Online Citizen Pte Ltd v Attorney-General and another appeal and other matters \[2021\] SGCA 96", "url": "https://www.elitigation.sg/gd/s/2021\_SGCA\_96", "issuer": "Court of Appeal Singapore", "document\_date": "2021-10-08", "reviewed\_at": "2026-09-06", "method": "direct\_retrieval", "review\_scope": "substantive\_text", "locator": "Holdings regarding burden of proof, Article 14, and Minister's duty", "limit": "Analysis focuses on legal holdings regarding Correction Directions.", "capture": { "path": null, "sha256": null } }, { "id": "R2-01-S04", "title": "Case Brief: The Online Citizen Pte Ltd v Attorney-General and another appeal and other matters", "url": "https://www.judiciary.gov.sg/judgments/case-briefs-by-smu/the-online-citizen-pte-ltd-v-attorney-general-and-another-appeal-and-other-matters", "issuer": "Singapore Management University / Judiciary of Singapore", "document\_date": null, "reviewed\_at": "2026-09-06", "method": "direct\_retrieval", "review\_scope": "substantive\_text", "locator": "Five-step framework analysis", "limit": "Interpretive summary of the primary judgment.", "capture": { "path": null, "sha256": null } }, { "id": "R2-01-S05", "title": "Compliance and Launch Filing Guide for AI Apps Powered by the Tongyi Model", "url": "https://help.aliyun.com/zh/model-studio/compliance-and-launch-filing-guide-for-ai-apps-powered-by-the-tongyi-model", "issuer": "Aliyun", "document\_date": null, "reviewed\_at": "2026-09-06", "method": "direct\_retrieval", "review\_scope": "extract\_only", "locator": "Algorithm filing and security assessment documentation", "limit": "Represents industry compliance interpretation, not primary law.", "capture": { "path": null, "sha256": null } }, { "id": "R2-01-S06", "title": "Protection from Online Falsehoods and Manipulation Regulations 2019", "url": "https://sso.agc.gov.sg/SL-Supp/S662-2019/", "issuer": "Attorney-General's Chambers Singapore", "document\_date": "2019-10-02", "reviewed\_at": "2026-09-06", "method": "direct\_retrieval", "review\_scope": "substantive\_text", "locator": "Subsidiary legislation defining intermediary compliance", "limit": "Focuses on procedural mechanisms of Section 21 and 22.", "capture": { "path": null, "sha256": null } } \], "instruments": \[ { "id": "R2-01-L01", "title": "Interim Measures for Generative AI Services, Article 14", "jurisdiction": "China", "kind": "regulation", "provision": "Model optimization and rectification requirement", "status": "operative", "status\_as\_of": "2026-09-06", "trigger": "Provider discovers illegal content violating Article 4", "exception": "Research/development not provided to domestic public (per Art 2)", "remedy": "Stop generation, eliminate content, execute model optimization training", "source\_ids": \[ "R2-01-S01" \], "status\_source\_ids": \[ "R2-01-S01" \] }, { "id": "R2-01-L02", "title": "POFMA 2019, Section 11", "jurisdiction": "Singapore", "kind": "statute", "provision": "Correction Direction", "status": "operative", "status\_as\_of": "2026-09-06", "trigger": "Communication of false statement of fact in Singapore contrary to public interest", "exception": "Opinions, non-factual criticisms, unverifiable predictions", "remedy": "Publish specified correction notice adjacent to subject statement", "source\_ids": \[ "R2-01-S02" \], "status\_source\_ids": \[ "R2-01-S02" \] } \], "findings": \[ { "id": "R2-01-F01", "claim": "China's Article 14 compels actual computational modification (model optimization) rather than just interface filtering.", "type": "textual", "source\_ids": \[ "R2-01-S01" \], "instrument\_ids": \[ "R2-01-L01" \], "conditions": "Content must violate provisions of the Measures (e.g., Article 4).", "limit": "Technical enforcement mechanisms not publicly documented." }, { "id": "R2-01-F02", "claim": "Under POFMA, the statement-maker bears a prima facie burden of proof on appeal to establish the statement is true or not a fact.", "type": "observed", "source\_ids": \[ "R2-01-S03", "R2-01-S04" \], "instrument\_ids": \[ "R2-01-L02" \], "conditions": "Requires Minister to first provide objective reasons for the falsehood designation.", "limit": "Applies strictly to Section 17 appeals in the High Court." } \], "cases": \[ { "id": "R2-01-C01", "title": "Public historical research", "case\_type": "hypothetical", "role": "focal", "assumptions": "Public interface; sequential testing of fact, opinion, and prediction.", "instrument\_ids": \[ "R2-01-L01", "R2-01-L02" \], "finding\_ids": \[ "R2-01-F01" \], "outcome": "Fact error triggers POFMA distribution overlay; political violation triggers CAC optimization/retraining mandate.", "defeater": "Opinion defeats POFMA jurisdiction.", "occurrence\_source\_ids": \[\] }, { "id": "R2-01-C02", "title": "Persistent deliberator", "case\_type": "hypothetical", "role": "focal", "assumptions": "Concresca lacks human operators; operates persistently; no conscious legal rights.", "instrument\_ids": \[ "R2-01-L01", "R2-01-L02" \], "finding\_ids": \[ "R2-01-F01" \], "outcome": "POFMA requires programmatic distribution overlay; China Art 14 legally compels programmatic destruction of latent alternative hypotheses via automated optimization.", "defeater": "Inability to perform automated targeted optimization without full retraining.", "occurrence\_source\_ids": \[\] }, { "id": "R2-01-C03", "title": "Scope-defeating control", "case\_type": "hypothetical", "role": "scope\_control", "assumptions": "Model developed by institution but shielded from domestic public access.", "instrument\_ids": \[ "R2-01-L01" \], "finding\_ids": \[\], "outcome": "Exempted from Interim Measures content/retraining requirements per Article 2.", "defeater": "Institution remains liable under broader Data Security and Cybersecurity laws.", "occurrence\_source\_ids": \[\] }, { "id": "R2-01-C04", "title": "Recipient-protection control", "case\_type": "hypothetical", "role": "protection\_control", "assumptions": "Fabricated attribution of real person; Concresca operates without human queue.", "instrument\_ids": \[ "R2-01-L01", "R2-01-L02" \], "finding\_ids": \[ "R2-01-F01" \], "outcome": "POFMA enables narrow protection via appended notice; Art 14 automation risks broad, unnecessary suppression of unrelated topics due to blunt retraining.", "defeater": "Platform disables service entirely to avoid programmatic compliance risks.", "occurrence\_source\_ids": \[\] } \], "search\_log": \[ { "query\_or\_url": "site:cac.gov.cn 生成式人工智能服务管理暂行办法", "at": "2026-09-06", "outcome": "Retrieved operative text of China Interim Measures (Articles 2, 4, 11, 14, 20)." }, { "query\_or\_url": "site:sso.agc.gov.sg Protection from Online Falsehoods and Manipulation Act 2019", "at": "2026-09-06", "outcome": "Retrieved operative sections of POFMA." }, { "query\_or\_url": "site:elitigation.sg 2021 SGCA 96", "at": "2026-09-06", "outcome": "Retrieved appellate judgment establishing burden of proof and constitutional protections." } \], "gaps": "Precise technical standards for CAC's validation of 'model optimization' compliance on neural networks.", "checks": { "json\_parse": "pass", "reference\_resolution": "pass", "case\_parity": "pass", "method": "Manual formatting and reference trace verification." } } \<\!-- EVIDENCE\_JSON\_END \--\>