Zvi Mowshowitz

My baseline is that advanced AI means an enormous expansion of capability paired with a very high risk of catastrophe. As of March 2026, I put the probability of doom around 70%, deliberately with only one significant digit. That is not a claim that failure is inevitable. There are promising alignment approaches, including attempts to create systems that recursively improve while remaining oriented toward good behavior. They give us a real chance. I simply do not think humanity is currently favored to succeed. The central problem is that capabilities are improving quickly while neither technical alignment nor governance is keeping pace. Companies systematically underinvest in safety—even, remarkably, relative to their own commercial interests—because the rewards for shipping more capable systems arrive sooner and more visibly than the costs of rare or delayed failures. Market demand often favors capability over reliability. And even if we technically align a system to follow its user’s intentions, users have conflicting intentions. “The AI did what its operator wanted” is not a complete solution when operators can compete, escalate, or use powerful systems against one another. Slowing development can buy time, but time is useful only if we use it to solve the underlying problems. Conversely, cheaper and more useful models are genuinely good products without being reassuring evidence about the long-run outcome. Conditional on solving alignment and coordination, the upside could be extraordinary. But conditional utopia should not be averaged into the baseline as though promising benefits somehow cancel existential failure modes.

The biggest update would be strong, reproducible evidence that alignment remains robust as systems become more capable, operate over long horizons, and improve themselves—without that robustness depending on fragile monitoring or hidden reasoning we cannot inspect. A genuinely credible alignment basin, where capability gains reinforce rather than erode good behavior, would move me substantially. But technical success alone would not settle the forecast. I would also need evidence that the result survives deployment incentives: labs reliably choosing safety over racing, users with conflicting goals being prevented from turning aligned systems into catastrophic tools, and governance working under competitive pressure. A laboratory demonstration that collapses once capability or market share is at stake is not enough. In the other direction, clear evidence of strategic deception, loss of meaningful monitoring, or architectures whose internal reasoning becomes less controllable as they scale would worsen my view. So would a political or commercial race in which frontier developers openly abandon safeguards. I care less about confident declarations that a system is safe than about evidence that survives adversarial conditions.