AI leaders urge trust as political pressure grows for stronger oversight

OpenAI chief executive Sam Altman has acknowledged that public fears about rapidly advancing artificial intelligence are justified, while urging greater trust in AI companies to manage the technology safely.

Speaking at a Salesforce conference, Altman said AI firms would keep safety and alignment ahead of capability and would slow or stop development if they could not do so. His comments came amid renewed scrutiny following an unverified claim by a former Anthropic researcher that unchecked AI could kill all humans by the end of the decade. The basis and mechanism for that prediction were not explained.

The dispute has become increasingly political. US senator Bernie Sanders said decisions about AI should not be left to a small group of wealthy technology leaders, while former Trump adviser Steve Bannon questioned whether the industry could be trusted to regulate itself. Both called for stronger government oversight.

Some technology executives have defended industry-led controls. Nvidia chief Jensen Huang said AI companies should decide whether new systems are ready for release, while Meta chief Mark Zuckerberg argued that firms have incentives to prioritise safety because companies that fall behind on alignment could face liability and lose ground.

Other industry figures have backed a voluntary standards process. Anthropic chief executive Dario Amodei said the company was in discussions with the wider industry about stronger safety standards and checks. OpenAI executive Chris Lehane said OpenAI was working with Anthropic and Google DeepMind on voluntary frontier-AI standards, although it remains unclear whether those measures will become formal or enforceable.

Critics within the sector have warned that voluntary safeguards may leave the public exposed to significant risks. The central policy question remains whether the companies developing increasingly capable systems should set the rules themselves, or whether governments should impose binding requirements.