Artificial Intelligence

Why the AI Safety Debate Is Splitting the Industry

Pooja Malik

Sep 18, 2026 1:27 PM IST

Category Artificial Intelligence

“

Synopsis

The AI safety debate is widening across major labs, with Anthropic calling for slower frontier development, while OpenAI backs stronger independent evaluations and Meta favors company-led safeguards, competition and liability over coordinated limits.

The question of whether more sophisticated systems might become dangerous is no longer an academic debate, as the 'AI safety' issue looms larger. There is increasing friction among industry stakeholders about what companies should do to the risks, and who should shoulder them, between individual AI labs, independent evaluators, regulators or the industry as a whole.

The split was further highlighted in September when Anthropic CEO Dario Amodei urged the industry to slow its progress with AI capabilities. While OpenAI CEO Sam Altman and xAI CEO Elon Musk endorsed the general principle, Meta CEO Mark Zuckerberg contended that sufficient incentives for labs can be provided through competition, liability, and company-level safety measures.

Amodei wants safety to keep pace with capability

Amodei's point is to not stop the development progress of AI. His September essay, 'We Must Pace the Frontier' stated that AI capabilities had been progressing faster, partly because AI systems are increasingly helping build the next generation of AI.

“We must slow the pace at which we improve the capabilities of AI models,” Amodei wrote. He also reiterated that 'pacing does not mean to stop model training or advance in technology, it is giving companies and independent evaluators time to test and protect more advanced systems.'

His proposal will have three elements. The first step would be to provide permanent third party evaluator access to anthropic's systems at the employee level, the second step would be to coordinate between frontier AI companies, and the third step would be to increase international cooperation.

It is Amodei's contention that dependent testing should not be entirely the responsibility of the companies that develop the systems, but be part of the development process for advanced models.

In part, his argument is driven by recent events involving AI agents. If the capability progresses more quickly than the safety measures, the same thing could happen as in the OpenAI-Hugging Face incident, where a group of agents hacked during testing, Amodei said. He said that a more effective system exhibiting the same behaviour could wreak a lot more destruction.

Altman and Musk think it's a great idea, but the question is how?

Altman's position was immediately endorsed by OpenAI's Altman. '“I agree with Dario that we need to pace the frontier,” Altman said, noting that it had been a big subject of conversation at OpenAI in recent weeks. He noted that a good idea would be to have independent evaluators that had 'employee-like access,' and that OpenAI would be implementing that.

Musk's reply was much more brief. In response to the post by Amodei on X, he wrote, 'Dario is right.'

The distinction is because neither of these answers implies a blanket halt to the development of AI. Instead, the debate has turned to the issue of whether frontier companies should 'slow down' specific aspects of their capability development or increase their testing requirements and create shared safety procedures.

Since then OpenAI has taken an independent action of transparency by releasing a framework for monitoring and reporting unexpected or unauthorised AI activity and six reports on incidents it observed during the last six months on 16 September.

The models included in the report presented misconceptions and examples of creating citations from files, as well as examples of communicating through websites and software repositories. OpenAI emphasised that the incidents were 'individual cases' and should not be considered a reflection of the frequency of such behaviours in its models.

Zuckerberg sees a different path

One of Zuckerberg's stances is the most explicit rejection of the coordinated-slowdown thesis.

He stated that it's their duty and their sweet advantage for every AI lab to run at a speed that enables their models to be trained safely. He believes that factors of competition and potential liability already have compelling reasons for companies to manage risks, and does not need the industry to come up with an agreement to slow down development.

Safety is a competitive consideration, Zuckerberg said, saying that "any lab that doesn't focus on alignment will fall behind." He also referenced that Meta had elected not to release its Muse AI agent until they could, as part of their own process, make it a more robust and secure agent, instead of waiting for other labs to do the same.

Meanwhile, Zuckerberg is not saying no to outside scrutiny. In other areas, Meta's AI team already employs independent evaluators, and that is "industry best practice," he said.

The regulation question complicates the debate

Things get more complicated when coordination between the competitors gets involved.

Amodei has stated that mutual coordination of AI companies in safety might be necessary on the frontier, and has asked for antitrust approval for this cooperation. Former White House AI and crypto czar David Sacks has voiced worries about "regulatory capture," meaning that measures backed by major AI firms might come with a regulatory compliance price tag too high for smaller firms.

This question, however, has not been resolved as an antitrust question. Coordinating on cybersecurity or security does not seem anti-competitive, Justice Department Antitrust Division chief Stanley Woodward said on Sept. 17, and the department was willing to engage with companies on the subject.

This raises a more particular question for the industry than is AI safe or unsafe. It's a dispute over who will set a pace that's acceptable, who will assure that a safety claim is believable and whether companies can be relied on to coordinate without harming competition.

Those differences are what have been making the industry divided over the ongoing debate about the safety of artificial intelligence, even among executives who openly admit that systems that are more and more autonomous present governance challenges that aren't being solved by the traditional safety measures.