Documented release procedures and model cards, open lawsuits over training data
Overall rating “solid”. No deduction for controversies. The superscript numbers link directly to the source.
1 · Responsibility
Oversight, an AI ethics committee and a tiered release procedure are documented, but lawsuits allege that management approved the use of unauthorised training data.
What is documented
- According to the 2026 proxy statement, the full board oversees the business model including AI, and the Audit Committee oversees information security2 1
- A Head of AI & Legal Ethics and an internal AI ethics committee from engineering, business, research, public policy and legal steer the work3 4
- The Frontier AI Risk Assessment provides risk tiers with graduated approvals; the highest tier requires approval from an independent body6
- Models and datasets are approved before publication; high-risk models are tested by third parties3
- An independently operated speak-up hotline protects against retaliation3
The case against
Allegations over sourcing of training data
- The court considers the allegations on the use of books from shadow libraries sufficiently pleaded; there is no finding7
- According to the amended complaint, management approved data access to Anna's Archive; NVIDIA denies this8 9
- Reports cite a blanket approval for downloading videos on a mass scale10 11
Gaps in accountability
- AI governance has not been externally audited or certified; apart from the number of model cards there are no AI metrics3
- Rules for procured and internally used AI are not disclosed
Sources for Responsibility (11)
- NVIDIA Corporation / SEC: Proxy Statement (DEF 14A) 2025 Annual Meeting · 12 May 2025 · primary source
- NVIDIA Corporation / SEC: Proxy Statement (DEF 14A) 2026 Annual Meeting · 12 May 2026 · primary source
- NVIDIA Corporation: NVIDIA Sustainability Report Fiscal Year 2026 · 12 June 2026 · primary source
- Kellogg School of Management: Nikki Pope (NVIDIA) on AI ethics · 12 June 2024 · secondary source
- Open Class Actions: NVIDIA Board Sued Over AI Trained on Pirated Books · 20 August 2026 · secondary source
- NVIDIA Corporation: NVIDIA Frontier AI Risk Assessment · 3 November 2026 · primary source
- U.S. District Court, N.D. California (via CourtListener): Order Granting in Part and Denying in Part Defendant's Motion to Dismiss, Nazemian et al. v. NVIDIA Corp., No. 24-cv-01454-JST, ECF No. 303 · 5 May 2026 · primary source
- TorrentFreak: NVIDIA Contacted Anna's Archive to Secure Access to Millions of Pirated Books · 19 January 2026 · secondary source
- Tom's Hardware: Nvidia says it didn't use pirated books to train its AI models — company asking for Anna's Archive suit to be dismissed · 8 February 2026 · secondary source
- Tom's Hardware: Nvidia accused of scraping 'A Human Lifetime' of videos per day to train AI · 6 August 2024 · secondary source
- U.S. District Court, N.D. California (via CourtListener): Complaint, Ted Entertainment, Inc. et al. v. NVIDIA Corp., Case No. 3:25-cv-10287 · 26 November 2025 · primary source
2 · Safety
Testing of the vehicle AI and a working process for security flaws are documented, but there are weaknesses in protection against misuse of open models.
What is documented
- The frontier framework names benchmarks, human red-teaming, garak scans before release and stress tests1
- For Nemotron 3 Nano, a content safety evaluation and filtering of training data including screening for child sexual abuse material are documented2
- The inspection lab for vehicle AI is accredited to ISO/IEC 17020; TÜV SÜD and TÜV Rheinland have carried out assessments3
- A PSIRT process with coordinated disclosure and bulletins is in place; the CVE-2025-23266 container escape and Triton flaws were fixed5 6 7
- The framework provides for an incident process with notification of authorities and withdrawal of models in cases of misuse1
The case against
Weak protection against misuse
- The Nemotron licence contains no usage restrictions beyond the law and export rules12
- According to the press, NVIDIA waives further guidelines in its agreement with the Pentagon, including for the development of autonomous weapons systems13
- According to a lawsuit, voiceprints were used without consent to train voice and speech models; the allegation is unproven14 15
Gaps in infrastructure
- Independent researchers found a container escape that put third-party models and data on shared GPU infrastructure at risk6
- Results of independent red-teaming of the Nemotron models have not been published
Sources for Safety (15)
- NVIDIA Corporation: NVIDIA Frontier AI Risk Assessment · 3 November 2026 · primary source
- NVIDIA Corporation (NGC Catalog): Nemotron 3 Nano – Model Card, Safety and Security · 3 November 2026 · primary source
- NVIDIA Corporation (Blog): NVIDIA Launches DRIVE AI Systems Inspection Lab, Achieves New Industry Safety Milestones · 6 January 2025 · primary source
- NVIDIA Corporation: NVIDIA Sustainability Report Fiscal Year 2026 · 12 June 2026 · primary source
- NVIDIA Corporation: NVIDIA PSIRT Policies · 3 November 2026 · primary source
- Wiz Research: NVIDIAScape – NVIDIA AI Vulnerability (CVE-2025-23266) · 17 July 2025 · primary source
- The Register: Chained bugs in Nvidia's Triton Inference Server lead to full system compromise · 5 August 2025 · secondary source
- TechRepublic: NVIDIA Responds to Demo of First Rowhammer Attack on GPUs · 18 July 2025 · secondary source
- NVIDIA Corporation (Blog, David Reber Jr.): No Backdoors. No Kill Switches. No Spyware. · 5 August 2025 · secondary source
- NVIDIA Corporation / SEC: Form 10-K for the fiscal year ended January 25, 2026 · 25 February 2026 · primary source
- NVIDIA Corporation: NeMo Guardrails Library – Overview · 3 November 2026 · secondary source
- NVIDIA Corporation: NVIDIA Nemotron Open Model License · 15 December 2025 · primary source
- The Boston Globe: US military deal with tech companies on AI · 2 May 2026 · secondary source
- U.S. District Court, N.D. Illinois (via Loevy & Loevy): Class Action Complaint, Rogers et al. v. NVIDIA Corp., Case No. 1:26-cv-05478 · 12 May 2026 · primary source
- Biometric Update: Tech giants sued under BIPA over voiceprints used to train AI · 13 May 2026 · secondary source
3 · Transparency
The model cards state data sources, token counts and evaluations, but information is missing for some models, and there is no dedicated AI transparency report.
What is documented
- The Nemotron Nano 9B v2 model card states sources and token counts per data type, the September 2024 cut-off and the source models of synthetic data3
- NVIDIA publishes pre-training and post-training datasets5
- More than 1,000 models have model cards to the Model Card++ standard with limitations and evaluation methods2
- The model cards contain sections on ethical considerations and a reporting channel for AI concerns3
The case against
Gaps for individual models
Sources for Transparency (8)
- NVIDIA Corporation: Trustworthy AI (Trust Center) · 3 November 2026 · secondary source
- NVIDIA Corporation: NVIDIA Sustainability Report Fiscal Year 2026 · 12 June 2026 · primary source
- NVIDIA Corporation (Hugging Face): NVIDIA-Nemotron-Nano-9B-v2 – Model Card · 18 August 2025 · primary source
- Kellogg School of Management: Nikki Pope (NVIDIA) on AI ethics · 12 June 2024 · secondary source
- NVIDIA (Hugging Face Blog): Nemotron's Open Secret: Accelerating AI Development with Open Models, Data, and Recipes · 22 October 2025 · secondary source
- U.S. District Court, N.D. Illinois (via Loevy & Loevy): Class Action Complaint, Rogers et al. v. NVIDIA Corp., Case No. 1:26-cv-05478 · 12 May 2026 · primary source
- U.S. District Court, N.D. California (via CourtListener): Order Granting in Part and Denying in Part Defendant's Motion to Dismiss, Nazemian et al. v. NVIDIA Corp., No. 24-cv-01454-JST, ECF No. 303 · 5 May 2026 · primary source
- Stanford CRFM: The Foundation Model Transparency Index (December 2025) · 3 November 2026 · primary source
4 · Why the harm rule does not apply
The harm rule does not apply: within the time window there is no judgment, regulatory decision or fine concerning harm caused by AI. The training data lawsuits have not been decided, and there is no pattern of serious incidents with NVIDIA's own AI systems.
5 · Controversies (7)
- mediumClass action in Illinois: voiceprints allegedly used without consent to train voice and speech models. · 12 May 2026 · open
- mediumClass action by authors over training on books from shadow libraries; the court largely denied the motion to dismiss. · 5 May 2026 · open
- mediumAgreement with the US Defense Department on AI on classified networks; according to the press, without further usage guidelines. · 1 May 2026 · unklar
- mediumLawsuits by YouTube creators over mass downloading of videos for the Cosmos model; one lawsuit was withdrawn. · 26 November 2025 · unklar
- minorShareholder lawsuit against management and the board over allegedly approved use of unauthorised training data. · 31 July 2026 · open
- minorChina's internet regulator summoned NVIDIA over alleged tracking and shutdown risks in H20 chips; NVIDIA denies backdoors. · 31 July 2025 · unklar
Further controversies (1)
- minorIndependent researchers found serious flaws in AI infrastructure software and a memory attack on GPUs; the flaws were fixed. · 15 July 2025 · abgeschlossen
6 · What remains open
- AI-specific metrics on tests, incidents and remedies
- Rules and approvals for procured and internally used AI
- Customer vetting and end-use controls for high-risk and military clients
- Misuse monitoring for open models
- Origin of the training data for the Cosmos world models and the voice models
- Independent review of AI governance
7 · How it was assessed
Cut-off date: 3 November 2026 · 43 searches, 44 fetches.
Primary documents read (13)
- Form 10-K FY2026 (S15, two retrievals, partial; Item 1C/3 not included)
- Form 10-K FY2026 Note Litigation (S14)
- Form 10-Q Q2 FY2027 (S16, financial section only)
- Proxy Statement DEF 14A 2025 (S17, first 100,000 characters)
- Proxy Statement DEF 14A 2026 (S18, first 100,000 characters)
- Sustainability Report FY2026 (S19, complete in two retrievals)
- Trustworthy AI page (S20)
- Frontier AI Risk Assessment (S21)
- Model cards: Nemotron 3 Nano and Nemotron-Nano-9B-v2 (S23, S24)
- Nemotron Open Model License (S26)
- PSIRT Policies (S28)
- Court records: Nazemian ECF 303 ruling (S03), complaints Millette (S07), Ted Entertainment (S09), Rogers BIPA (S11)
- Official list: Seoul Frontier AI Safety Commitments (S22)
Research of 3 November 2026 under Globalance AI Footprint rulebook version 0