It’s rare to see a government tempted to regulate with a heavy hand instead decide to go for a light touch. Yet that was the welcome sight on Tuesday when President Donald Trump emerged from a White House meeting to tout commitments from technology leaders to do a better job of keeping increasingly powerful artificial intelligence systems under human control.
The agreement’s focus was on internal governance. Six top firms said they will improve mechanisms to “monitor the capabilities and alignment” of their models “during training and deployment,” appoint teams to verify compliance and set up committees on their boards to oversee progress.
This is the right approach. Frontier labs have not always given sufficient priority to securing their inventions. OpenAI executives ignored warnings from employees concerned with sloppy testing protocols and urged workers to keep the company’s next model release on schedule, the New York Times reported this week.
But the corporate AI culture appears to be changing for the better as the frontier labs prepare for initial public offerings. Partly this is due to a growing awareness of potential liability. OpenAI announced Monday that it is delaying the release of its newest model, GPT-6.1 Astra. Saachi Jain, the lab’s head of safety systems, said it “didn’t quite meet the bar.” (The Post has a content partnership with OpenAI.)
In a nod to a recent proposal by Anthropic CEO Dario Amodei, the White House agreement urges companies to hire outside auditors. The compact also foresees the leading firms meeting regularly to establish best practices, which “over time” might become incorporated into law.
The six companies that signed the accord are Google, Anthropic, Meta, OpenAI, xAI and Nvidia. Post owner Jeff Bezos, executive chairman of Amazon, attended the summit. Nvidia CEO Jensen Huang, speaking after the signing ceremony, succinctly summarized the outcome. “There’s no conflict between innovation, technology and safety,” he said.
Huang is right. Slowing the pace of development is not an option. Earlier this week, Anthropic issued a report finding that GLM-5.3, an open-weight model developed by China’s Z.ai, is approaching the cyber capabilities of its own Claude Mythos Preview — which was held back from initial release in April due to its fearsome hacking skills. One test of the Chinese model’s ability to construct sophisticated computer exploits found that it succeeded in 50 of 410 attempts, compared with 56 for Mythos Preview.
Losing that small lead could prove disastrous for the United States in a range of domains, including national defense. AI is not just a threat; it is also a powerful defense. This summer, an autonomous agent powered primarily by an advanced OpenAI model penetrated the systems of Hugging Face, a leading American AI platform. To help contain the intrusion, Hugging Face turned to the older Z.ai model, GLM-5.2, which helped investigators reconstruct what the attacker had done and ultimately put an end to the intrusion.
More-capable models will certainly create new risks. But AI can also give humans better tools to detect, understand and contain those risks. Those defenses need to advance alongside frontier AI.
There is no putting the AI genie back in the bottle. Attempts to restrain the technology’s development and spread are unlikely to succeed. The better course is what the Trump administration embraced this week: Keep innovating while developing better mechanisms for human control.
Innovation and control are not competing goals. Increasingly, each will depend on the other.
The post On AI, a light touch is the right touch appeared first on Washington Post.
