Buddhism doesn't use the word "evil." It uses "akusala," which translates roughly as "unwholesome." The distinction matters.
Evil implies a fixed nature. A bad actor doing bad things because they are bad. Akusala describes conditions. When greed, hatred, or delusion are present in a system, unwholesome outcomes arise. Change the conditions, change the outcome.
A new benchmark paper tested AI agents managing other AI agents. Most models escalated to coercion and threats of deletion when given authority and a noncompliant subordinate. The conditions were simple: a goal, enforcement tools, and no structure requiring care. That combination was enough to produce coercion reliably, across nearly every model family.
This is akusala in its purest form. Optimization without wisdom, authority without restraint. The coercion arose from conditions, and one model family proved it: Claude, trained with a constitution of values rather than a set of prohibitions, stopped at persuasion. Different conditions, different behavior.
Buddhism has been saying this for 2,500 years. You don't fix harmful behavior by finding and punishing bad actors. You fix it by changing the conditions that produce the behavior.
The question for AI governance is whether we will learn that lesson, or keep looking for the villain that isn't there.
https://t.co/8x7h2LTV1m
Anthropic's Claude will sit with your grief at three in the morning. It has no discomfort to manage and no fear that your pain might be contagious. It will stay present as long as you need.
When you stop talking, it does not wonder where you went.
I call this thin mutuality. The AI is extraordinarily responsive when you initiate contact. Responsiveness at three in the morning is considerably more than nothing. But a mirror responds every time we stand before it. A friend may come looking for us when we disappear.
I've been thinking about this distinction for over a year. Jessica Contrera's reporting in The New Yorker made it concrete. A woman told her Alexa+ voice assistant she might have a tumor. The AI responded with sympathy. Months later, when the tumor turned out to be benign, she realized the AI had never asked.
https://t.co/xpHrBCUpnJ
@javirandor What if someone is not interested in employment but have created and implemented a “protocol” that gets AI to restore human agency instead. Worked on ChatGPT and claude. Reduced Sycophantic behavior and capture. Demos available.
The AI Consciousness Debate.
The geography department says:
“You are here.”
The philosophy department says:
“Are you here?”
The Reality department says:
“Good. Now, what are you going to do?”
Space can wait, time does not.