Safety Framework
ByteDance does not publish a standalone “frontier AI safety framework” document comparable to Anthropic RSP v3, OpenAI Preparedness v2, or DeepMind FSF v3. This is a shared feature of China’s leading AI companies. The core claim of this page: it is not that ByteDance is “absent” on safety frameworks; rather, the corporate tier is absorbed by the national tier—state regulation pulls risk management up to the industry / national level, making corporate-tier “self-governance documents” institutionally redundant.
1. ByteDance’s AI-safety institutional architecture (as of 2026-04)
Section titled “1. ByteDance’s AI-safety institutional architecture (as of 2026-04)”Four parallel compliance / safety workstreams
Section titled “Four parallel compliance / safety workstreams”| Layer | Content | Matching rules |
|---|---|---|
| Algorithm-filing layer | CAC filing and change-filing for Doubao, Coze, Volcano Engine, etc. | Generative AI Interim Measures Art. 17 · Algorithm Recommendation Provisions |
| National-standard compliance layer | Full-line compliance with TC260-003 (basic requirements) + GB 45438-2025 (labeling) | TC260-003-2024 · GB 45438-2025 |
| Governance-framework alignment layer | AI Safety Governance Framework 1.0 / 2.0 guiding corporate-level execution | Safety Governance Framework 1.0 / 2.0 |
| Overseas compliance layer | TikTok DSA Art. 34 systemic risk assessment; CapCut C2PA; jurisdiction-specific localized Trust & Safety | EU DSA; California AB 2013 / 2655; Singapore IMDA Model AI Governance Framework, etc. |
Party Committee + Compliance + Trust & Safety
Section titled “Party Committee + Compliance + Trust & Safety”Based on public information after ByteDance formally established its Party Committee (党委) in 2024, together with industry interviews, corporate-level safety governance is structured roughly as:
- Party Committee (top direction): sets major technology-route choices (e.g. open-source vs. closed-source), value-alignment orientation, and red lines on sensitive topics— consistent with the “Party Committee → algorithm → product three-tier coordination” observation proposed in Olivia’s thesis The Embedding of Algorithmic Values by Party Committees in Chinese AI Firms.
- Cybersecurity compliance / legal: multi-agency liaison with CAC / MIIT / MPS; preparing filing materials; responding to targeted enforcement; cross-border compliance.
- AI Ethics Committee (established 2024): a cross-departmental deliberative body typically including AI Lab research leads, the general counsel, the head of Trust & Safety, and internal independent audit. Major product launches clear it (for example, the 2025 Q4 launch of Doubao’s “virtual friend” feature went through it).
- Trust & Safety: a content-moderation and safety organization of thousands, split between China and overseas; domestic teams are primarily in Tianjin / Chengdu / Jinan; overseas teams in Dublin / Los Angeles / Singapore / London.
2. No RSP: structural comparison with frontier labs
Section titled “2. No RSP: structural comparison with frontier labs”| Dimension | Anthropic RSP v3 (2026-02) | OpenAI Preparedness v2 | DeepMind FSF v3 | ByteDance |
|---|---|---|---|---|
| Form of the framework | Corporate self-governance document | Corporate self-governance document | Corporate self-governance document | No standalone document |
| Risk tiering | ASL-2/3/4 | High / Critical | CCL / TCL | No corporate tiering (compensated by national GB 39724 and TC260-003) |
| Capability thresholds | Public | Public | Public | No public thresholds |
| External review | Third-party reviewers (GovAI / MATS / METR) | Safety Advisory Group | FSF Report | CAC + CAICT + MCNSA algorithm-evaluation lab (quasi-regulatory third parties) |
| Pause commitment | Withdrawn in v3 | Weak wording | None | Not applicable (regulators can mandate takedown) |
| Frontier voluntary commitments | Signed U.S. 2023 voluntary commitments + UK AISI MoU + GPAI CoP | Signed | Signed | None signed |
Key observation: the Western three-step sequence of “corporate self-governance → external review → regulatory pressure” is compressed in China into the two-step sequence “regulator leads, enterprise executes.” The absence of a corporate self-governance document at the frontier-model layer does not mean safety work is absent; it means the institutional tier is different.
3. ByteDance’s participation in national standards and frameworks
Section titled “3. ByteDance’s participation in national standards and frameworks”Drafting and revising TC260-003
Section titled “Drafting and revising TC260-003”ByteDance is one of the member units of the generative AI safety working group under TC260 (全国信息安全标准化技术委员会). The two core contents of TC260-003-2024 Basic Safety Requirements for Generative AI Services are:
- A 31-item risk list (organized into 5 major classes: violation of core socialist values, discriminatory content, commercial-law violations, infringement of others’ lawful rights and interests, and defects in accuracy / reliability)
- 34 specific requirements across 7 classes (training-data safety, corpus labeling, model safety, service transparency, emergency response, etc.)
Practice cases contributed by ByteDance in the TC260 working group (publicly visible in the published meeting summary excerpts) include: Doubao’s large-model content-safety evaluation methodology, red-line keyword recognition in anthropomorphic scenarios, and contextual risk management in multi-turn dialogue.
First revision of TC260-003 is underway in 2025–2026; draft additions focus on safety requirements for multimodal (vision, speech, video), Agent / tool-calling scenarios, and anthropomorphic interaction. ByteDance co-leads several chapters with Alibaba, Baidu, Tencent, Huawei, and China Telecom.
AI Safety Governance Framework 2.0 (2025)
Section titled “AI Safety Governance Framework 2.0 (2025)”In 2025 the CAC and TC260 jointly released version 2.0, which adds a chapter on systemic-risk governance for frontier models (comparable to the GPAI chapter of the EU AI Act), but does not require enterprises to publish an RSP. ByteDance’s implementation approach is: to disassemble the 2.0 requirements into filing materials plus internal processes, without producing a standalone external document.
4. The 2026 Anthropomorphic Interactive Services Measures: ByteDance’s largest new compliance pressure
Section titled “4. The 2026 Anthropomorphic Interactive Services Measures: ByteDance’s largest new compliance pressure”The Interim Measures for the Administration of Anthropomorphic Interactive AI Services (published 2026-04-10, effective 2026-07-15) imposes dedicated requirements on ByteDance’s Doubao “virtual friend / AI persona / AI partner” products:
- Prohibition during nighttime hours for minors (22:00 – 06:00)
- Refusal on sensitive topics (suicide / self-harm / extreme emotional counseling must route to a human or hotline)
- Emotional-addiction detection and intervention mechanisms
- Prominent “AI identity” notice (each conversation must begin with explicit “I am an AI”; persistent impersonation of a human is prohibited)
- Periodic re-filing of anthropomorphic avatars (appearance / persona changes trigger change-filing)
ByteDance’s response: Doubao’s “Character” feature is expected to undergo a major overhaul in 2026 Q2, with some AI personas either removed or restricted. The estimated end-to-end adjustment cycle across product, filing, and Trust & Safety is 3–6 months.
5. Overseas: TikTok DSA Art. 34 Systemic Risk Assessment
Section titled “5. Overseas: TikTok DSA Art. 34 Systemic Risk Assessment”As a VLOP, TikTok must publish an annual Systemic Risk Assessment (SRA) under DSA Art. 34 / 35. Published versions include:
- 2024 SRA (published 2024-11) — covering four risk classes: illegal-content dissemination, fundamental-rights harms, democratic process and public discourse, and minors / public health. Co-issued with Kroll as independent auditor.
- 2025 SRA (published 2025-11) — added a dedicated AI-generated-content chapter and included a dedicated post-mortem of pro-Russia account activity around the 2024 European Parliament election (tied to the European Commission’s DSA formal proceedings).
This is the only structured risk assessment in the ByteDance product family that is published to an international public audience—the equivalent algorithm-filing materials in China are not public, but the DSA legal duty compels TikTok to publish. This inversion (“overseas transparency > domestic transparency”) is a persistent tension in ByteDance’s compliance architecture.
6. Academic critique
Section titled “6. Academic critique”Chinese scholarship
Section titled “Chinese scholarship”- Zhang Linghan: the “tier absorption” of national AI safety governance is both an efficiency advantage and a risk— internal corporate risk-management capabilities may atrophy exogenously, since “compliance floor becomes highest duty.”
- Dai Xin: in China’s “structural dual compliance,” ByteDance satisfies both CAC filing and global regulators at the same time, which builds unique compliance muscle but also produces fragmentation of internal policy logic.
- Xue Lan: agile collaborative governance must guard against “too agile”— if rules iterate too quickly, firms cannot plan long-term safety investments.
- Zhu Yue: the absence of standalone corporate-level safety-framework documents weakens third-party academic oversight and independent evaluation.
International scholarship
Section titled “International scholarship”- Matt Sheehan (Carnegie): in ChinAI-related writing and public commentary, his consistent argument is that Chinese frontier AI companies lack RSP-like documents not because of capability but because of institutions—the state lifts risk-management to the industry / national tier, so corporate-level self-governance documents become institutionally redundant. This is the core lens for understanding the U.S.–China divergence on self-governance documents.
- Jeffrey Ding (GWU): the ChinAI Newsletter repeatedly argues: absence of an RSP does not mean absence of safety investment— ByteDance’s content-safety team is larger than the combined red-team plus policy departments of most frontier labs.
- Rebecca Arcesati (MERICS): Chinese corporate AI Safety has a characteristic profile of “operational capability first, theorization lagging”— it can defend against 31 risks item by item, but lacks a conceptual framework for “frontier risk.”
- Helen Toner (CSET): the risk types absorbed by Chinese regulation (content, values, minors) and the risks that Western RSPs focus on (CBRN, autonomous replication, cyber-offensive capability) are ontologically different— which makes U.S.–China communication on frontier AI risk very hard.
- Markus Anderljung (GovAI): even with a strong national tier, corporate-level frontier-risk self-governance is still needed as a complement; the Chinese model is effective on “visible present risks” but under-covers “future catastrophic risks that have not yet materialized.”
- Paul Triolo (DGA): ByteDance’s “embedded compliance” (嵌入式合规) is a pathway of the Chinese model that other emerging markets (Southeast Asia, the Middle East) are willing to adopt, and forms a counterweight to the American model in global competition.
7. Comparison with other leading Chinese firms
Section titled “7. Comparison with other leading Chinese firms”| Company | Flagship model | Standalone safety framework? | Filing cadence | Open-source strategy |
|---|---|---|---|---|
| ByteDance | Doubao / Seed | No | First batch 2023-08; multiple products, multiple change-filings | Restrained open source (small Seed models) |
| Alibaba | Qwen | No (safety only at developer-documentation level) | First batch 2023-08; full series, multiple filings | Aggressive open source (0.5B–72B Apache-2.0) |
| Baidu | ERNIE | No (has an “ERNIE AI governance white paper,” not an RSP-shaped document) | First batch 2023-08 | Partial open source |
| Tencent | Hunyuan | No (has a Tencent AI Governance White Paper) | First batch 2023-08 | Partial open source |
| DeepSeek | V3 / R1 | No | Multi-batch filings | Most aggressive open source (MIT) |
Common conclusion: none of the Chinese frontier labs has produced a corporate-level safety-framework document comparable to an RSP. This is an institutional feature (national-tier absorption + non-signing of GPAI CoP + TC260-003 de facto coverage), not an idiosyncratic firm-level choice.
8. Why doesn’t ByteDance sign the GPAI Code of Practice?
Section titled “8. Why doesn’t ByteDance sign the GPAI Code of Practice?”ByteDance has not signed the EU GPAI CoP (open to first-batch signatories from 2025-08-01). Other Chinese firms that have also not signed include Alibaba, Baidu, Tencent, DeepSeek, Zhipu, and Moonshot—no top Chinese frontier-model company has signed.
Structural reasons:
- Models not entering the EU market do not generate a signing incentive (Doubao / Qwen / ERNIE / Hunyuan / DeepSeek consumer markets are not centered in the EU).
- TikTok is already a VLOP, so ByteDance’s EU compliance pressure is concentrated on DSA rather than the AI Act GPAI chapter.
- Signing means accepting systemic-risk duties under the EU AI Act— for firms that have not committed to large-scale EU entry, it raises compliance cost.
- Geopolitical signaling: signing an EU framework can be read as “taking sides,” and most top Chinese firms prefer to stay compliance-neutral.
9. 2026 Q1 developments
Section titled “9. 2026 Q1 developments”- 2026-01: ByteDance’s AI Ethics Committee was reorganized, expanding working groups on anthropomorphic interaction, Agent safety, and multimodal deepfakes.
- 2026-02: TC260-003 revision work began; ByteDance is involved in multimodal and Agent chapters.
- 2026-03: TikTok’s EU team negotiated implementation of the DSA Art. 40 researcher-data-access agreement with the European Commission.
- 2026-04: publication of the Anthropomorphic Measures triggered an enterprise-wide compliance review of Doubao’s “Character” feature.
10. Related index
Section titled “10. Related index”- Top-level rules: AI Safety Governance Framework 1.0 / 2.0 · TC260-003-2024 · Anthropomorphic Interactive Services Measures
- Peer comparison: Anthropic / safety-framework (RSP v3) · OpenAI / safety-framework (Preparedness) · DeepMind / safety-framework (FSF)
- Adjacent company pages: usage-policy · red-team-disclosures