In an early October 2026 appearance on Politico's Decoded podcast, Sam Altman laid out OpenAI's risk tolerance in unusually direct terms: society should accept scams, hacks, and other limited harms as the price of broad AI access, but catastrophic risk is a different category entirely that shouldn't be tolerated. He also pushed back on treating AI models with quasi-religious significance, calling that framing itself a safety concern.
Taken together, the comments sketch a more specific worldview than OpenAI has typically articulated publicly, one that accepts real costs from AI's current deployment in exchange for continued iteration and access, while drawing a harder line only at the far end of the risk spectrum, a position that leaves considerable room for disagreement about exactly where the line between “limited harm” and “catastrophic” actually falls.