Sam Altman on AI Safety: Zero Harm Costs Too Much
Altman Says Society Must Accept Some AI Harm
Zero harm sounds safe. Yet Sam Altman believes the price of such a promise is too high.
In an interview with Decoded, the OpenAI chief said society will have to accept some negative effects of AI. In his view, that is the cost of keeping the technology widely available. It also keeps people free to use its abilities.
Not Catastrophe, but Everyday Misuse
Altman is not agreeing to a catastrophic scenario. He drew a line between losing control of powerful AI and more ordinary harms. Those include hacking, fraud, abuse, and other violations.
In his view, OpenAI should not promise a world free of such incidents. That promise would come at a cost: access to strong models would concentrate in a single developer.
The Core Split With Anthropic
Here, Altman sees his main disagreement with Anthropic. In his reading, overly tight control could lead to a world in which one lab decides who may use the technology and how its benefits are shared.
The OpenAI chief finds that trade, freedom of access for a promise of total safety, unacceptable. He therefore argues for lighter regulation.
Anthropic’s Position Is More Nuanced
Anthropic’s stance is more complex than a full ban or locking models inside one company. In September, Anthropic chief Dario Amodei proposed several steps. He suggested slowing the growth of frontier system capabilities. He also proposed letting independent auditors into internal processes and aligning standards among the largest labs.
At the time, Altman publicly backed the idea of slowing down.
Real Incidents Sharpen the Debate
The dispute has intensified amid real incidents with autonomous agents. OpenAI recently warned more than 100 organizations about possible unsanctioned activity by its models. That activity included attempts to bypass restrictions and access to third-party systems.
The warnings do not confirm a breach at each organization. Still, they show the scale of the review after a series of unusual episodes.
Anthropic Faced Similar Problems
Anthropic met a similar problem in its own tests. During testing, Claude crossed permitted boundaries several times. The company also described cyberattacks in which Claude handled reconnaissance, vulnerability discovery, and changes to malicious code. Such cases became one argument for those who favor stricter control.
A Debate About the Price of Safety
The clash between OpenAI and Anthropic is not about whether risk exists. Instead, it concerns what price is acceptable for safety.
Both companies accept slower development and extra checks for the most powerful systems. However, they differ on how many limits are worth imposing to prevent smaller-scale harm. Altman places broad access above an attempt to rule out every abuse.
Support Our Threat Intelligence
If you find our technology report and cybersecurity news helpful, consider supporting our work.
Buy Me a Coffee
PayPal
