What evidence would make the risk look more serious?

The debate should move when evidence moves.

Several developments would materially increase concern:

• AI agents that can operate reliably for days or weeks on complex goals without human correction.
• Repeated evidence that systems strategically conceal capabilities or manipulate evaluations.
• Reliable autonomous replication or unauthorized resource acquisition in realistic environments.
• Major improvements in offensive cyber capability that exceed skilled human baselines across full attack chains.
• Strong evidence that AI access substantially increases real-world biological threat capability beyond existing resources.
• Widespread deployment of frontier systems with broad permissions inside critical infrastructure and weak monitoring.
• AI systems making or executing high-consequence military decisions at speeds that eliminate meaningful human intervention.
• Safety evaluations consistently failing to predict dangerous behavior after deployment.

Notice how specific these signals are. None requires guessing whether a model has a soul. They are empirical questions about capability, access, behavior, and control.