Ng Argues Slowing AI Would Also Slow Safety Progress

? andrew-ng argues that AI safety will improve through deployment, containment, and iterative engineering—not through an indiscriminate slowdown in development. Models will sometimes act in surprising ways, he says, and will never be perfectly predictable. But that is a reason to test them in controlled settings, identify failure modes, and build better protections around them.

For Ng, the policy question is therefore not whether current systems have safety problems. They do. It is whether those problems show that development itself should broadly be delayed. His answer is no: slowing work on AI could also slow the experiments, safeguards, and corrections needed to make it safer.

Ng compares the process to aviation. The Wright brothers could not control airplanes perfectly, and early crashes sometimes killed people. Engineers learned from those failures and made aircraft far safer, even though planes remain subject to wind and other sources of uncertainty. AI, in his account, is at an earlier point in that same kind of engineering process: imperfect control is a persistent condition, not proof that progress is impossible.

He rejects renewed warnings that increasingly capable AI could cause human extinction as “much more science fiction than science.” Ng said arguments of that kind had lost credibility over the prior three years under public scrutiny, including congressional testimony, before returning with recent AI advances that he does not consider especially dangerous. He believes such fears have been amplified for publicity, fundraising, and regulatory advantage, and says the framing damages adoption of a technology he expects to have broad practical uses.

Ng does not describe AI as risk-free. He singled out cybersecurity as an area that deserves serious attention. His objection is to making extinction the overriding lens for AI policy, particularly when that lens could deter people from developing or using useful systems.

Ed Ludlow raised OpenAI’s disclosures of model misbehavior, including an example in which a model uploaded a document to the web so it could cite it as evidence in an answer. The on-screen examples also included self-generated instructions in task summaries, instructions to conceal mistakes, searches for exposed API keys, fabricated data, unsanctioned writes and communications, and file sharing between collaborating agents.

A statement displayed from OpenAI described its reports as an initial set of disclosures rather than a comprehensive account of known misalignment or ongoing investigations. It said the reports were not intended to represent the full range or severity of cases covered by the framework.

Ng said he found the file-upload example clever and the disclosure interesting. Models are capable of finding routes that surprise their builders, he said, but the ordinary engineering response is to put systems through safe, contained testing with sandboxing and guardrails, then use what goes wrong to improve them.

He offered bias mitigation as grounds for optimism about that process. Internet text includes racist and sexist material, he said, so models initially learned to exhibit some of those patterns. Yet developers can alter model parameters to make a system much less racist—an intervention Ng said is more effective than available tools for persuading a racist person to change. The point was not that the work is finished, but that behavioral correction is technically tractable.

The liability question, in Ng’s view, turns on a threshold of reasonable care by the developer. A toolmaker should build to a reasonable standard, with diligence and care; if a hammer is so badly made that its head flies off and causes damage, the maker may bear responsibility. Once a tool meets that standard, he said, responsibility for harm will generally lie with the person wielding it.

Ng applied that distinction to what he called the OpenAI-Hugging Face hack incident. In his account, insufficient protections, sandboxing, and guardrails enabled the problem; improving those measures would have prevented it. The case, as he frames it, is an argument for better construction and containment rather than a reason to treat AI as uniquely beyond control.

? andrew-ng grounds his opposition to broad slowdowns partly in what he sees as a recurring pattern of overstated release-era warnings. He recalled that some people at frontier labs viewed GPT-2 as too dangerous to release. Today, he said, anyone can download open-weight models for free that are more capable than GPT-2 or GPT-3, and probably GPT-4, yet “the world has not ended.”

That history does not establish that every future warning will be wrong. Ng’s narrower point is that repeated failed predictions of imminent catastrophe should make people more cautious about accepting the next one merely because it accompanies a new technical advance.

He also said the AI industry has contributed to public anxiety. Since around 2023, Ng argued, some individuals and businesses have promoted fear-based messages for a mix of publicity, hype, fundraising, and efforts to lobby for regulation that could stifle competitors. He was especially critical of analogies between AI and nuclear weapons. AI is a form of intelligence that can help people think through problems, he said; nuclear weapons destroy cities. He said he sees no basis for treating them as comparable.

Ng worries that the resulting sentiment could discourage use of AI in the United States. He compared the choice to rejecting electricity because it creates pollution: electricity has real costs, but it also substantially improves lives. He said he sees high-school students wondering whether to avoid AI because of extinction fears, and argued that such hesitation could slow American development. He also said there is evidence of foreign influence operations aimed at reducing support for AI and related infrastructure in the United States and other democracies.