Nvidia's Jensen Huang Says AI Safety Is an Engineering Problem, Not a Legal One

Nvidia founder and CEO Jensen Huang used his appearance at Salesforce's Dreamforce conference on Tuesday, September 15, to argue against new AI legislation. AI safety is "an engineering problem, not a legal one," he said, according to TechCrunch.

Huang described AI as hardware and software built by humans, and therefore controllable by humans under existing laws. He said there is no need for a separate legal regime for model development, deployment or use. Regulation should focus on measurable risks rather than broad restrictions on research or products, a position he also laid out in early September, according to Yahoo Finance.

He placed responsibility on vendors. Companies should not release a product or service if they are not confident in its functionality, capability or safety, Huang said. They should pace themselves or pause until they get it right. He argued free market forces provide enough discipline because buyers will reject unsafe products.

Huang also rejected language that frames models as a new form of "alien mind." In his framing, AI systems are built artifacts that can be tested, patched and recalled, much like cars or phones receive fixes and recalls.

His comments landed in an active policy sequence. In early September, the United States pressed G20 members to take a hands-off approach to AI regulation and avoid creating new rules, according to Reuters. President Donald Trump said in September the United States already has guardrails in place to regulate and prosecute AI companies, according to Reuters. On September 14, TechCrunch reported Huang had told Trump "we're not going to let [an AI slowdown] happen," in a piece by Amanda Silberling.

Not all labs share that hands-off posture. In August 2026, OpenAI called for California to add more safeguards to its AI safety bill passed the prior year.

Nvidia's official language is narrower than a no-new-laws headline implies. Its AI trust center states AI should respect privacy and data protection regulations, operate securely and safely, and function transparently with accountability. In its Open Weights and American AI Leadership paper in July, Nvidia said openness may be one of the most important paths to AI safety and security, and relying solely on closed models is not inherently safe. The company joined leaders from the White House, Congress and the tech industry in September 2023 to discuss AI standards and best practices.

Huang has consistently framed AI as infrastructure rather than application. He said AI is no longer a single breakthrough or application, it is infrastructure every company will use and every nation will build. At Dell Technologies World, he called every company an "Intelligence Manufacturer." In March 2026, he said every company should have an OpenClaw strategy. In April 2025, he appeared to have struck a deal with the Trump administration to avoid export restrictions on Nvidia's H20 AI chips.

The broader context here matters for reading his position as a chip supplier. For practitioners, safety as engineering means familiar work: evaluation benchmarks (standard tests that score model behavior), adversarial testing (deliberate attempts to trick or break a model), data provenance controls (records of where training data came from), access management, inference monitoring (watching a model while it serves users), incident response and staged rollouts.

In my view, market pressure disciplines visible failures well and systemic failures poorly. Customers can punish a chatbot that hallucinates, meaning it invents facts, or an agent that deletes data. They are slower to detect biased scoring, quiet data leakage across retrieval pipelines, systems that pull outside data into answers, or correlated failures when many enterprises deploy the same foundation models and safety tools. Engineering rigor is necessary. Whether it substitutes for independent standards depends on how much failure a buyer can observe before purchase.

Looking at what this means for builders, the open-weights argument may prove more consequential than the no-new-laws argument. Open weights means model files are shared for inspection. If broad inspection, reproducible evaluation and competitive patching improve safety, then export policy, licensing terms and disclosure norms matter as much as any safety bill. Infrastructure often gets governed through procurement, insurance, tort and existing sector rules before new statutes arrive. That is where Huang's position will be tested in practice, and where careful engineering can expand what enterprises trust with production workloads.