In recent years, artificial intelligence has advanced following an almost industrial rule: more data, more computing power, larger models and, as a consequence, systems capable of doing more and more things. Now, after years of warning from experts, one of the companies that has contributed most to accelerating this race proposes something that until recently would have seemed contradictory: slowing it down.
Sam Altman, CEO of OpenAI, told his employees this week that the company is open to reducing the pace of development of its most advanced artificial intelligence systems. According to Bloomberg, Altman also expressed his hope that other laboratories would take a similar position. This is not, at least for now, a general pause or abandoning the development of new models, but rather a recognition that there may come a time when moving faster is no longer the safest option.
The idea has antecedents. In March 2023, more than a thousand scientists and figures from the technological world signed a letter calling for a pause of at least six months in training systems more powerful than GPT-4. They argued that society did not yet know how to control systems that could acquire unpredictable capacities. The proposal did not prosper and the main companies continued to increase the power of their models.
OpenAI had also warned of the problem. In July of that same year, he created the Superalignment team, led by his then chief scientist Ilya Sutskever, to develop methods capable of keeping systems under control that could far exceed human capabilities. The company reached announce that it would dedicate 20% of its available computing capacity over the next four years to this research. However, the initiative disappeared in 2024, after the departure of Sutskever and researcher Jan Leike, one of its leaders. The decision fueled criticism from those who believed that OpenAI was moving faster than its security research was progressing.
The company has meanwhile developed its Preparedness Framework, a system to assess the risks of its models before deploying them. Among them include the possibility of helping the development of biological or chemical weaponscybersecurity risks and the ability of the systems themselves to improve their capabilities.
But the nature of the problem has also changed. Today’s systems don’t just answer questions: they can use tools, browse the Internet, write and run code, and operate for long periods of time with much less human supervision. The concern, therefore, is no longer just that an AI provides an incorrect answer, but rather that What you can do when you have autonomy and access to real systems.
In this context, OpenAI’s own chief scientist, Jakub Pachocki, has defended the need to slow down development. Their argument is not to stop research forever, but to ensure that security mechanisms advance at least at the same pace as the capabilities of the models. Until now, No company has demonstrated that it has sufficient methods to control much more intelligent and autonomous systems.
The problem is that a company does not compete with itself. Slowing down OpenAI doesn’t do much good if its competitors continue to speed up. A laboratory that voluntarily reduces its pace could lose an advantage over another that does not. Hence, Altman points towards coordination between companies, something much more complicated than an internal decision. OpenAI is even studying whether an agreement between companies to slow development could violate US antitrust laws: a measure designed to increase safety could also be interpreted as a restriction of competition.
The consequences would be important. A slowdown would mean fewer new models or more time between generations and could affect the huge investments made in data centers and chips under the expectation of continued growth in AI. It would also have a geopolitical dimension: If American companies slowed down while other countries maintained theirs, the technology race could change hands.
But stopping would also have an advantage: buying time. Time to develop better evaluation systems, study the behavior of autonomous agents, establish common standards and understand what happens when a machine acquires capabilities that its own creators cannot predict.
That’s why Altman’s words are more significant than they seem. OpenAI is not saying that artificial intelligence is too dangerous to continue. He is admitting something more uncomfortable: that we may not yet know when a new capability stops being an advance and begins to become a risk that we do not know how to manage. For years, the question was how much artificial intelligence could speed up. Now another one begins to appear: yes we are able to accelerate our ability to control it at the same time.