Open models and cyber oversight, but a court ruling over a false AI answer
Overall rating “limited”. Harm rule applied. The superscript numbers link directly to the source.
1 · Responsibility
Baidu documents a structure for cybersecurity and data, but there is no AI policy and Ranking Digital Rights documents major weaknesses in due diligence and remedy.
What is documented
- The Corporate Governance and Nominating Committee oversees cybersecurity risks and reviews the related disclosure1
- A Data Management Committee chaired by the General Counsel reports quarterly1
- The risk factors in the 20-F name reliability, data security and ethical risks of AI2
- Baidu signed the safety commitments of the China AI Industry Alliance with obligations on red-teaming and risk management6
- According to company and secondary statements, Qianfan and three cloud applications are certified to ISO/IEC 420013 5
The case against
Independently documented weaknesses
Missing AI governance
- No AI policy, risk classes or approval procedure for AI were found in the research period
- The AIIA commitments are declarations of intent; their implementation is not documented6
- There is no complaints channel for faulty AI answers with documented remedy
Sources for Responsibility (9)
- Baidu, Inc. (SEC-Einreichung, XBRL-Bericht): Form 20-F FY 2025 – Cybersecurity Risk Management and Strategy Disclosure (R10) · 3 November 2026 · primary source
- Baidu, Inc. (SEC-Einreichung): Form 20-F for fiscal year 2025 · 3 November 2026 · primary source
- Baidu Security (account Baidu_Secrity on AtomGit/CSDN): 2025 百度 ESG 报告发布:以 AI 筑盾,共建可信数字生态 · 12 June 2026 · secondary source
- SustainabilityReports.com: Baidu Inc 2024 SR (summary) · 3 November 2026 · secondary source
- AIbase: Baidu Qianfan receives ISO/IEC 42001 certification (report) · 20 September 2024 · secondary source
- Emerging Technology Observatory (AGORA): Artificial Intelligence Safety Commitments (China AI Industry Alliance) · 1 December 2024 · secondary source
- 国家互联网信息办公室 (CAC): 国家互联网信息办公室关于发布生成式人工智能服务已备案信息的公告 · 2 April 2024 · primary source
- Ranking Digital Rights: 2025 Big Tech Edition – Baidu · 1 August 2024 · primary source
- 国家互联网信息办公室 (CAC): 网信系统持续推进网络执法 查处各类网上违法违规行为 · 31 January 2024 · primary source
2 · Safety
According to the court, an AI answer in search violated a lawyer's reputation, and numerous robotaxis came to a standstill in Wuhan, while red-teaming for ERNIE is not verifiably documented.
What is documented
- The 20-F describes a defence system for network, host and application, internal testing and testing with external providers, and reviews by qualified third parties10
- Potentially material cyber incidents are escalated to the responsible board committee10
- According to Citizen Lab, the critical decryptability of Baidu IME on Windows was fixed by April 202411
- A 2023 white paper describes multi-layer filters and red-teaming procedures as a product offering9
The case against
Serious incidents involving its own AI systems
- Courts in Nanjing found that an AI answer falsely portrayed a lawyer as a convicted person; Baidu apologised only in enforcement proceedings1 3
- In Wuhan, numerous robotaxis came to a standstill and some passengers were stuck for almost two hours; the police provisionally cited a system failure4 5
- According to Bloomberg, new L4 permits were subsequently halted nationwide; the authority has not confirmed this6
- In Chongqing, an Apollo Go vehicle with a passenger drove into a construction pit; she was unhurt7
Sources for Safety (15)
- 澎湃新闻 (The Paper): «百度AI称一律师被判刑»构成侵权:百度辩称AI幻觉无法预见,法院判其书面道歉 · 9 May 2026 · primary source
- 东方财富网 (Eastmoney): «百度AI称一律师被判刑»构成侵权:百度辩称AI幻觉无法预见 法院判其书面道歉 · 9 May 2026 · secondary source
- 北京日报 (Beijing Daily, news.bjd.com.cn): AI智能回答侵犯名誉权,百度败诉道歉 · 22 September 2026 · secondary source
- 澎湃新闻 (The Paper): 武汉多辆«萝卜快跑»行驶中突然停摆…警方通报:初步判断为系统故障,无人员伤亡 · 1 April 2026 · secondary source
- The Express Tribune (Reuters): Baidu robotaxi outage in Wuhan caused by system failure, police say · 1 April 2026 · secondary source
- The Star (Bloomberg): China suspends autonomous driving permits after Baidu outage · 29 April 2026 · secondary source
- Motorshow (after Reuters): Robotáxi da Baidu cai em vala de construção na China · 8 August 2025 · secondary source
- ERNIE Team, Baidu (arXiv): ERNIE 5.0 Technical Report (arXiv:2602.04705) · 4 February 2026 · primary source
- Baidu Security (北京百度网讯科技有限公司): 大模型安全解决方案白皮书 (Version 1.0) · 24 August 2023 · secondary source
- Baidu, Inc. (SEC-Einreichung, XBRL-Bericht): Form 20-F FY 2025 – Cybersecurity Risk Management and Strategy Disclosure (R10) · 3 November 2026 · primary source
- Citizen Lab, University of Toronto: The Not-So-Silent Type: Vulnerabilities Across Keyboard Apps Reveal Keystrokes to Network Eavesdroppers (Report No. 175) · 23 April 2024 · primary source
- Ranking Digital Rights: 2025 Big Tech Edition – Baidu · 1 August 2024 · primary source
- Baidu Security (account Baidu_Secrity on AtomGit/CSDN): 2025 百度 ESG 报告发布:以 AI 筑盾,共建可信数字生态 · 12 June 2026 · secondary source
- Sixth Tone: Baidu Exec Tries to Diffuse Doxxing Controversy · 18 March 2025 · secondary source
- Emerging Technology Observatory (AGORA): Artificial Intelligence Safety Commitments (China AI Industry Alliance) · 1 December 2024 · secondary source
3 · Transparency
Baidu publishes open weights for ERNIE 4.5 and a technical report on ERNIE 5.0, but discloses neither limitations nor safety information nor an explanation of the outage in Wuhan.
What is documented
- The ERNIE 5.0 technical report names data categories, filters against unsafe content and deduplication4
- The ERNIE 4.5 model card names the licence, architecture and post-training procedures5
- Chinese regulations require display of the model name and registration number and, from September 2025, labelling of AI content1 2
The case against
Missing information on models
Missing disclosure to users
- Ranking Digital Rights finds no explanation of the recommendation and advertising algorithms and no information on government requests3
- On the outage in Wuhan there were only customer service and in-vehicle messages citing a network problem7 8
- Baidu is not named among the platforms that have updated their terms on labelling AI content2
Sources for Transparency (8)
- 国家互联网信息办公室 (CAC): 国家互联网信息办公室关于发布生成式人工智能服务已备案信息的公告 · 2 April 2024 · primary source
- 科技日报 (stdaily.com): AI生成内容不得«隐身»传播,多家平台更新用户协议 · 2 September 2025 · secondary source
- Ranking Digital Rights: 2025 Big Tech Edition – Baidu · 1 August 2024 · primary source
- ERNIE Team, Baidu (arXiv): ERNIE 5.0 Technical Report (arXiv:2602.04705) · 4 February 2026 · primary source
- Baidu (Hugging Face): Model card baidu/ERNIE-4.5-VL-424B-A47B · 3 November 2026 · primary source
- 澎湃新闻 (The Paper): «百度AI称一律师被判刑»构成侵权:百度辩称AI幻觉无法预见,法院判其书面道歉 · 9 May 2026 · primary source
- Jiemian Global: Baidu robotaxi outage in Wuhan exposes system fragility at scale · 31 March 2026 · secondary source
- CnEVPost: Baidu Apollo Go robotaxis hit by rare mass paralysis in Chinese city · 1 April 2026 · secondary source
4 · Why the harm rule applies
Harm rule applied: harm caused by the company's own AI is documented, so the rating cannot be better than “limited”. A final court judgment found that an AI answer in Baidu's search falsely portrayed a lawyer as a convicted criminal.
5 · Controversies (6)
- seriousCourts in Nanjing ordered Baidu to apologise because an AI answer in its search falsely portrayed a lawyer as a convicted person; the appeal was dismissed. · 12 March 2026 · abgeschlossen
- mediumIn Wuhan, numerous Apollo Go robotaxis came to a standstill on the road and some passengers were stuck for almost two hours; the police provisionally cited a system failure. · 31 March 2026 · unklar
- mediumCitizen Lab showed that Baidu IME encrypted keystrokes inadequately; the flaw on Windows was fixed, while weaknesses on Android and iOS persisted. · 23 April 2024 · unklar
- minorIn Chongqing, an Apollo Go robotaxi with a passenger drove into a construction pit; the passenger was unhurt. · 6 August 2025 · unklar
- minorThe daughter of a Baidu vice president allegedly published personal data of internet users; after an internal investigation, Baidu denies any data leak. · 17 March 2025 · unklar
- minorThe CAC summoned Baidu over prohibited content in search results, required it to remedy this and fined the company. · 31 January 2024 · abgeschlossen
6 · What remains open
- AI-specific board oversight and named persons responsible for AI safety
- Public AI policy with risk classification and approval procedure
- Safety evaluation, red-teaming results and limitations of ERNIE
- Safety figures and incident reports for Apollo Go, in particular on the Wuhan outage
- Documented labelling of AI-generated content in search and chat
- Content of the inaccessible 2024 and 2025 ESG reports on AI governance and data protection
7 · How it was assessed
Cut-off date: 3 November 2026 · 40 searches, 33 fetches.
Primary documents read (8)
- Form 20-F FY 2025 (p. 1, characters 0–117614)
- Form 20-F FY 2025, Item 16K via XBRL R10 (S2)
- ERNIE 5.0 Technical Report (S8, HTML and PDF)
- Model Card ERNIE-4.5-VL (S9)
- Ranking Digital Rights 2025, Baidu (S10)
- Citizen Lab Report No. 175 (S12)
- CAC notices of 31 January 2024 and 2 April 2024 (S27, S28)
- Baidu Security White Paper: Large Model Safety 2023 (S30)
Research of 3 November 2026 under Globalance AI Footprint rulebook version 0