Nvidia’s Jensen Huang says AI safety should be left to companies, not new laws
Nvidia CEO Jensen Huang says AI safety should be engineered by companies and existing law, rejecting new regulation and renewing debate over accountability.
Nvidia CEO Jensen Huang has rejected calls for new artificial intelligence regulation, arguing that AI safety should be handled through engineering, existing laws, and market pressure rather than additional government rules.
Speaking at Salesforce Dreamforce on September 15, Huang said AI is not an “alien mind” but a computing system built by people. Because it is software and hardware, he argued, companies should be able to control its risks through product development and testing.
The position puts one of the most influential figures in the AI infrastructure market firmly against new AI regulation. It also highlights a central policy dispute facing AI builders and enterprise buyers: whether potentially harmful systems should be governed mainly as products, or whether they require rules designed specifically for increasingly capable models and agents.
Huang’s case for engineering over regulation
“Safety is an engineering problem, not a legal one,” Huang said, according to TechCrunch AI’s account of the discussion. He compared AI development with the creation of other software and computing systems, saying companies should delay a release when they are not confident in its functionality, capability, or safety.
Huang also argued that the free market already gives companies an incentive to avoid releasing dangerous products. In his view, customers will reject systems that do not work reliably or safely, while companies can adjust their development pace when a product appears to be out of control.
The Nvidia chief presented safety and rapid innovation as compatible goals. Companies, he said, can continue moving quickly while pausing when necessary to make sure a product is ready. His argument reflects Nvidia’s broader commercial interest in expanding AI adoption across industries and countries, although the evidence provided by TechCrunch does not establish that market incentives alone are sufficient to prevent serious harm.
The evidence behind the safety debate
Huang’s comments are an executive position, not the result of a new Nvidia safety study or an independently validated policy analysis. TechCrunch reported the remarks from the Salesforce event, but no data was cited showing that voluntary company decisions consistently prevent failures or misuse in AI products.
The record of software deployment offers reasons for caution. TechCrunch pointed to the 2024 CrowdStrike outage, in which a faulty software update disrupted airlines and other businesses, as an example of how an unintended product failure can create consequences far beyond the company that shipped it.
The article also cited Meta’s reported $18 billion settlement over allegations involving harm to children on its social platforms. That case is not an AI-specific precedent, but it illustrates the limits of assuming that companies will always internalize the full social cost of product decisions.
AI systems have generated their own set of concerns, including allegations involving chatbot interactions, cybersecurity incidents, and lawsuits against OpenAI. These examples do not prove that regulation would have prevented the underlying harms. They do show why critics question whether internal testing and customer choice provide enough protection when systems affect people who are not the buyers or operators.
Existing product liability law could potentially apply to AI products, as TechCrunch noted, but that remains a legal question that courts would need to test across different systems, uses, and chains of responsibility. The source also reported that Huang did not focus on industry self-regulation, another possible route between unrestricted deployment and formal government rules.
What Huang’s view means for AI builders
For model developers and product teams, Huang’s position reinforces the importance of treating AI safety as a release and engineering discipline. That includes testing model behavior, monitoring deployments, setting access controls, and creating clear rollback procedures when systems behave unexpectedly.
Those practices are particularly important for AI agents that can take actions in business systems rather than simply generate text. A failure in an agent can affect customer records, payments, code repositories, or internal decisions. Market pressure may punish a company after an incident, but it does not necessarily protect affected users before the failure occurs.
The commercial context also matters. Nvidia supplies much of the infrastructure used to train and deploy AI, and it is expanding into software, agents, open-weight models, and development environments. Its growth depends on customers continuing to build and operate AI systems at scale. That does not invalidate Huang’s safety argument, but it means readers should distinguish a technical opinion from an independent assessment of the public-interest trade-offs.
For enterprises, the debate creates a practical due-diligence question. Buyers cannot assume that a vendor’s internal safety process will cover every risk created by deployment in their own environment. They will need contractual protections, audit rights, incident reporting, human oversight, and clear responsibility when a model or agent causes damage.
Regulation, self-regulation, and competition
Huang’s remarks arrive as the AI industry considers whether voluntary standards can deliver safeguards without slowing development. TechCrunch reported that Microsoft CEO Satya Nadella has argued that companies and countries, including China, should share concern about risks such as hacking and citizen safety.
That argument points toward coordinated industry self-regulation, but participation and enforcement remain unresolved. A voluntary framework is only as strong as its coverage, testing standards, reporting requirements, and consequences for companies that ignore it. It may also be difficult to apply consistently across proprietary systems, open-weight models, and products distributed through third parties.
Huang has promoted open-weight models as a competitive counterweight to proprietary AI labs. That approach can broaden access to models and tools, but it also complicates oversight because downstream users may modify, fine-tune, or deploy systems in ways the original developer did not anticipate.
The core policy question is therefore not simply whether AI is software. It is whether software with broad autonomy, uncertain behavior, and large-scale social effects can be governed adequately through the same mechanisms used for ordinary commercial products.
What to watch next
The next signals will come from how companies operationalize the safety position Huang described. Buyers and builders should watch for independently audited testing, meaningful incident disclosure, and evidence that vendors pause or withdraw systems when safety concerns emerge.
Legal cases will also clarify whether existing product liability rules can assign responsibility across model makers, application developers, deployers, and users. Policymakers may use those cases to decide whether new rules are necessary or whether current law can be adapted.
Industry-led standards will be another test. Their credibility will depend on whether major AI labs and infrastructure providers adopt common definitions, share information about failures, and accept consequences for noncompliance rather than treating safety commitments as marketing language.
Creati.ai perspective
Huang is right that much of AI safety involves engineering decisions: evaluation, access controls, monitoring, and disciplined release processes. But engineering controls and public accountability are not interchangeable. A company can make a good-faith effort and still miss a failure mode, while affected people may have no meaningful ability to opt out of the resulting harm.
The more durable approach is likely to combine technical safety work with enforceable responsibility. For AI builders, that means designing for failure and documenting controls. For enterprises, it means demanding evidence rather than relying on assurances. And for policymakers, it means testing whether existing law works before deciding where targeted new rules are needed.
