AI review steps according to the ESG report, but independent criticism of remedy and disclosure
Overall rating “solid”. No deduction for controversies. The superscript numbers link directly to the source.
1 · Responsibility
The 2025 ESG report describes an AI subcommittee, a pre-release review and training; Ranking Digital Rights, however, finds weak due diligence and remedy, and an AI responsibility on the board is missing.
What is documented
- The board has delegated ESG oversight to the Corporate Governance Committee; the 2025 ESG report has limited assurance by a third party1 2
- Management's Group Technology Committee has subcommittees for privacy, security and AI; the AI Technology Committee coordinates and reviews AI projects1
- According to the ESG report, a joint security and compliance review takes place before release; privacy impact assessments and internal audits are described1
- More than 61,000 employees were trained in information security and data protection; AI safety training takes place per business team1 8
- Ranking Digital Rights notes that Tencent reviews AI products for discrimination and safety risks6
The case against
Independent findings
Sources for Responsibility (8)
- Tencent Holdings Limited (HKEX-Einreichung, Kopie auf pdf.dfcfw.com): 2025年環境、社會及管治報告 (ESG Report 2025) · 31 December 2025 · primary source
- Tencent Holdings Limited: ESG Governance Structure · 15 December 2022 · primary source
- Tencent Holdings Limited: Corporate Governance (Board committees) · 3 November 2026 · secondary source
- Tencent Holdings Limited (HKEX filing): Annual Report 2024 · 19 March 2025 · primary source
- Tencent Holdings Limited: Tencent Responsible AI Principles · 3 November 2026 · secondary source
- Ranking Digital Rights: 2025 Big Tech Edition – Tencent Holdings Limited · 2025 · primary source
- EOS at Federated Hermes: Tencent – Engagement case study (December 2023) · 1 December 2023 · secondary source
- Tencent Holdings Limited (HKEX filing): Environmental, Social and Governance Report 2024 · 31 December 2024 · primary source
2 · Safety
Tencent describes reviews before and after release, red-teaming and a runtime gateway; published safety evaluations are missing, and Citizen Lab found weaknesses in WeChat encryption.
What is documented
- According to the 2025 ESG report, models are reviewed before release and then monitored automatically and manually for accuracy, stability and inappropriate content1
- Red-teaming, a model safety platform and a runtime gateway against prompt attacks and data leakage are described1
- Fixed response times apply to incidents (30 minutes for intrusions, 2 hours for critical vulnerabilities); no material cyber or data incidents are reported1 8
- Ranking Digital Rights rates security oversight as fully met (P13 100%)9
- The Hunyuan licence prohibits, among other things, impersonation, circumvention of safeguards and military purposes11
The case against
Missing test evidence and incidents
- The technical report on Hunyuan-Large and the model card for Hunyuan-A13B contain no safety evaluation2 3
- According to reports, Yuanbao gave insulting answers twice in January and February 2026; Tencent spoke of a model anomaly5 6 7
- Abuse monitoring for open weights and measures for deepfake detection are not disclosed12
Sources for Safety (12)
- Tencent Holdings Limited (HKEX-Einreichung, Kopie auf pdf.dfcfw.com): 2025年環境、社會及管治報告 (ESG Report 2025) · 31 December 2025 · primary source
- Tencent Hunyuan Team (arXiv): Hunyuan-Large: An Open-Source MoE Model with 52 Billion Activated Parameters by Tencent (v3) · November 2024 · primary source
- Tencent (Hugging Face): Model card Hunyuan-A13B-Instruct · 27 June 2025 · primary source
- 36氪 (36Kr): 腾讯发布大模型安全与伦理报告:以负责任AI引领大模型创新 · 29 January 2024 · secondary source
- OECD.AI – AI Incidents and Hazards Monitor: Tencent Yuanbao AI Insults User Due to Model Malfunction · 4 January 2026 · secondary source
- 澎湃新闻 (The Paper): Tencent Yuanbao AI repeatedly outputs abusive language while editing code – official response · 5 January 2026 · secondary source
- OECD.AI – AI Incidents and Hazards Monitor: Tencent Yuanbao AI Outputs Insulting Language to User During Image Generation · 24 February 2026 · secondary source
- Tencent Holdings Limited (HKEX filing): Environmental, Social and Governance Report 2024 · 31 December 2024 · primary source
- Ranking Digital Rights: 2025 Big Tech Edition – Tencent Holdings Limited · 2025 · primary source
- The Citizen Lab (University of Toronto): Should We Chat, Too? Security Analysis of WeChat's MMTLS Encryption Protocol (Report No. 178) · 15 October 2024 · primary source
- Tencent (Hugging Face): Tencent Hunyuan 3D 2.0 Community License Agreement · 3 November 2026 · primary source
- 新华网 (Xinhua): 从AI反诈助手到生态共建:腾讯反诈新范式 · 8 September 2025 · secondary source
3 · Transparency
Tencent Seals AI content and publishes open models with a technical report; Ranking Digital Rights, however, finds little explanation of algorithms and data use, and model limits are missing.
What is documented
- According to its own account, Tencent Seals AI content in text, images and video explicitly and implicitly1
- Since September 2025 WeChat has labelled AI content and prohibits removing or falsifying the Seals2
- The technical report on Hunyuan-Large describes the architecture and the rough composition of the training data5
- Since March 2025 Yuanbao has used user content only with participation in an optimisation programme that is switched off by default8 9
The case against
Independent findings
Sources for Transparency (9)
- Tencent Holdings Limited (HKEX-Einreichung, Kopie auf pdf.dfcfw.com): 2025年環境、社會及管治報告 (ESG Report 2025) · 31 December 2025 · primary source
- 品玩 (PingWest), per announcement «微信珊瑚安全»: 微信:进一步规范人工智能生成合成内容标识 · 1 September 2025 · secondary source
- Ranking Digital Rights: 2025 Big Tech Edition – Tencent Holdings Limited · 2025 · primary source
- Tencent Holdings Limited: Tencent Responsible AI Principles · 3 November 2026 · secondary source
- Tencent Hunyuan Team (arXiv): Hunyuan-Large: An Open-Source MoE Model with 52 Billion Activated Parameters by Tencent (v3) · November 2024 · primary source
- Tencent (Hugging Face): Model card Hunyuan-A13B-Instruct · 27 June 2025 · primary source
- Tencent (Hugging Face): Tencent Hunyuan 3D 2.0 Community License Agreement · 3 November 2026 · primary source
- Sina Tech: Who really owns AI-generated content? Tencent Yuanbao's three revisions to its user agreement spark controversy · 5 March 2025 · secondary source
- Phoenix Tech (republished from Huxiu): A 'small' question — why Yuanbao + chat, and not chat + Yuanbao? · 31 January 2026 · secondary source
4 · Why the harm rule does not apply
The harm rule does not apply: in the review period there is no ruling, no regulatory decision and no fine for AI harms. The two insulting Yuanbao answers are documented only by secondary sources and are not serious.
5 · Controversies (4)
- minorAccording to reports, Yuanbao produced insulting language while designing a New Year image; Tencent spoke of an anomaly and apologised. · 24 February 2026 · unklar
- minorAccording to reports, Yuanbao insulted a user while revising code; Tencent confirmed a rare model anomaly and announced improvements. · 4 January 2026 · unklar
- minorThe Yuanbao terms of use granted Tencent an irrevocable, perpetual licence to user content; after criticism they were changed in March 2025. · 4 March 2025 · unklar
- minorCitizen Lab demonstrated cryptographic weaknesses in WeChat's MMTLS encryption without showing a complete attack; Tencent is sticking with MMTLS. · 15 October 2024 · abgeschlossen
6 · What remains open
- Board responsibility for AI; the 2025 Corporate Governance Report was not included in the retrieved file.
- Published safety evaluations and red-teaming results for HY 3.0 and Yuanbao.
- Independent review of AI labelling and abuse monitoring.
- Metrics on AI complaints and remedy.
- Training data sources and copyright policy of the Hunyuan models.
- Whether the public copy of the 2025 ESG report that was read matches the version filed with the Hong Kong stock exchange.
7 · How it was assessed
Cut-off date: 3 November 2026 · 41 searches, 29 fetches.
Primary documents read (11)
- ESG Report 2025 (S4, pp. 1–33, via pdf.dfcfw.com)
- Annual Report 2025 (p. 2, excerpt excluding Corporate Governance Report)
- Annual Report 2024 (p. 1, characters 0–100,000)
- Full-Year Results 2025 (p. 3)
- ESG Report 2024 (p. 5, characters 0–137,392, excluding AI chapter)
- Responsible AI Principles (p. 6)
- Technical Report Hunyuan-Large (p. 18)
- Hunyuan3D 2.0 Licence (S19)
- Hunyuan-A13B-Instruct Model Card (S20)
- Ranking Digital Rights 2025 (S9)
- Citizen Lab Report No. 178 (S10)
Research of 3 November 2026 under Globalance AI Footprint rulebook version 0