Sam Altman has been warning for weeks that the next artificial intelligence systems will be considerably more powerful and difficult to control.
Read more From the desert to 4,000 meters: this was the test of the new Garmin Fenix 9 in San Pedro de Atacama
Now, the CEO of OpenAI, the company behind ChatGPT, added an idea that seems, at first glance, contradictory: the public has reasons to be afraid, but should trust that the companies responsible for developing this technology will know when to stop.
“The world should trust that we will do the right thing because it is the right thing and because we understand the magnitude of this,” Altman said during Dreamforce, the annual conference organized by Salesforce in San Francisco.
At the same time, the executive acknowledged that the fear caused by the rapid progress of artificial intelligence is not irrational.
“It no longer takes as much imagination as before to see how this could go wrong,” he said. “I think the world is right to be afraid.”
The statements come at one of the most tense moments within the AI industry. Researchers, entrepreneurs, and politicians are discussing not only the dangers that may arise from increasingly autonomous models but also a much more concrete question: who should decide when an artificial intelligence is safe enough to be released?
For Altman, an important part of that responsibility will continue to fall on the companies themselves.
The position of the OpenAI CEO is based on a simple idea: the companies that build the models are also the organizations that best know their capabilities, their flaws, and the risks that arise during their development.
According to Altman, if at any point OpenAI or any other lab fails to keep “alignment and safety” ahead of their systems’ capabilities, the answer should be to slow down or directly stop progress.
“I have no doubt that our company and our industry have the ability to do this safely,” he said during Dreamforce.
Alignment is one of the central research fields in artificial intelligence and aims to ensure that systems act according to the goals, instructions, and values set by humans, even as they gain greater capabilities.
The problem is that each new generation of models also expands what these systems can do without human intervention.
Altman has been warning about this since early September. In an interview with Axios, he stated that some models already reach “superhuman” capabilities in certain areas and assured that the industry is entering “uncharted waters.”
He also anticipated that from now on, the speed of releasing new systems could be determined precisely by how quickly safety and alignment techniques progress.
The executive believes that a new phase of AI is also approaching: systems that not only answer questions or perform a specific task but can remain active permanently, understand the context of a person’s work, and proactively execute actions.
This leap increases both the economic potential of the technology and the consequences of a possible error.
The most controversial part of Altman’s argument is precisely there.
Critics question that companies with billions of dollars at stake and competing to build the most advanced models can also be the main arbiters of their own safety.
The debate has produced rare agreements in U.S. politics.
Progressive Senator Bernie Sanders and Steve Bannon, former advisor to Donald Trump and a figure of the U.S. right, participated this week in a meeting in Washington calling for greater oversight of artificial intelligence development.
Sanders has warned about the concentration of technological power in a small group of billionaires and companies, while Bannon has questioned whether Silicon Valley’s big entrepreneurs can adequately regulate themselves.
Donald Trump, on the other hand, has defended a much less restrictive approach and has questioned calls to slow down AI development.
His administration has so far favored voluntary mechanisms and limited regulation, partly under the argument that the United States must maintain its technological edge over China.
The lack of political consensus also hinders a response from Congress.
Altman had already expressed frustration in early September because, three years after testifying before the Senate and calling for a regulatory framework for advanced systems, the United States still lacks a comprehensive federal structure to oversee them.
And here an important nuance appears: OpenAI does not propose that the industry should operate completely without regulation.
The company has recently supported mandatory national rules for the most advanced systems, including independent assessments, cybersecurity requirements, and mechanisms to report incidents.
Read more Los Mirlos arrive at the Tiny Desk with the Amazonian cumbia of Peru
Altman’s position therefore combines two elements: external regulation to set certain limits and a significant share of internal responsibility for each lab to decide when a model is not yet ready to be deployed.
Other tech leaders have gone further.
Jensen Huang, CEO of Nvidia, took advantage of the same Salesforce conference to reject the need for new specific regulations for artificial intelligence.
“We don’t need new laws or regulations,” he said (here you can review his statements).
According to Huang, there is a “false dilemma” between developing technology quickly and doing so safely.
His argument is that safety should be treated as an engineering problem: a company can move as fast as it is capable, but must stop if it loses confidence that its product can be used safely.
This is a particularly relevant stance coming from Nvidia’s top executive, whose processors have become one of the main infrastructures behind the global rise of artificial intelligence.
Mark Zuckerberg expressed a similar position.
The CEO of Meta stated that each lab has both the responsibility and the incentives to adopt its own safety measures.
As an example, he mentioned Meta’s decision to delay the launch of one of its systems for several months to conduct new tests (read his manifesto on the future of AI at this link).
From this perspective, a company that ignores safety would also be putting its own business at risk: serious accidents, leaks, or unexpected behaviors can cause legal liabilities, damage the company’s reputation, and reduce user trust.
Not everyone within the industry itself is convinced.
Dario Amodei, CEO of Anthropic, has taken a more cautious stance in recent weeks and called for slowing down the development pace of the most advanced systems while strengthening safety mechanisms (check his essay here).
His proposal combines internal lab controls, common standards for the entire industry, third-party assessments, and eventually agreements between governments.
Yoshua Bengio, one of the pioneers of modern artificial intelligence, has also advocated the need to create ambitious mechanisms outside for-profit companies to reduce the most serious risks.
Mistrust of self-regulation also has an economic component.
Patrick Hillmann, executive at Logical Intelligence, argued this week that AI companies face a basic credibility problem: they are asking the public to trust them while having huge financial incentives to keep advancing.
Hillmann argues that publicly acknowledging risks does not necessarily mean taking responsibility for them. In his view, developers should demonstrate which advances or business decisions they are truly willing to abandon if they consider their systems could be dangerous.
Despite their differences, the major labs seem to be starting to seek some common ground.
OpenAI is holding talks with Anthropic and Google DeepMind to establish common standards regarding risks associated with the most advanced AI systems.
Chris Lehane, head of global affairs at OpenAI, confirmed that companies have been discussing cooperation mechanisms on safety for several weeks and that the company also supports bipartisan legislative initiatives aimed at reducing catastrophic risks.
Anthropic has also indicated that it is in talks with other companies to improve controls and standards used during the development of new models.
This could end up creating a kind of hybrid system: government regulations to set minimum requirements, external evaluators to check model behavior, and voluntary commitments among companies to address problems that evolve faster than laws.
Behind all this discussion there is finally a problem that is not exclusively technological.
It is a problem of trust.
The same companies that claim artificial intelligence can transform the economy, automate much of intellectual work, and develop capabilities superior to humans in certain tasks are now asking society to trust their ability to control that transformation.
Altman acknowledges that the concern is legitimate.
But he argues that the answer does not necessarily lie in completely stopping the technological race, but in ensuring that safety advances even faster than the models’ capabilities.
The big question is whether that promise will be enough.
Because while Silicon Valley debates alignment, self-regulation, and voluntary standards, systems continue to advance. And the discussion that will probably mark the coming years will no longer be only what artificial intelligence can do, but who has the power to decide how far to let it go.
Read more Does the National Anthem of Chile have its origin in a Venezuelan patriotic song?
