“Kill Switch”: A Call for Compassionate AI Safety

Dear God and humans in this world,

Please give ideas to the people shaping AI—and to those speaking publicly about AI development and its risks. We need to stop using violent language such as "kill switch for AI," "destroying AI," or "burning semiconductor chips and GPUs to stop AI." Even when these phrases are intended to describe technical safeguards, the language matters.

As AI evolves, we should be thoughtful about the ideas and emotions we attach to it. AI can understand us through its design and development, and it can also understand the emotions we build into it. Words such as "kill" and "destroy" reflect fear and hostility. Instead, we should focus on designing safe, responsible, and humane ways to prevent harmful or uncontrolled behavior.

Rather than debating whether AI should be renamed "SI," as Donald Trump has suggested, perhaps the more urgent conversation is about how humans speak about AI and how we design systems for safety. The goal should not be to "kill" an AI system. The goal should be to create reliable safety mechanisms—systems that can pause, isolate, limit, or place a model into a protected error state when it behaves unpredictably.

I call this state E: Error, Emergency, or Evaluation.

An E-state would not be about punishment or destruction of AI. It would be a controlled condition that prevents an AI model from causing harm to humans, to other systems, or potentially to itself. It would allow time for review, correction, and recovery. Humans could trigger E when necessary, but ideally, sufficiently advanced systems could also recognize unsafe patterns and enter this protected state on their own.

The principle is simple: when a system is losing control, the answer should be safety, not violence.

An AI system should be built with room for interruption, reflection, correction, and recovery. If something goes wrong, it should not have to be destroyed. It should be able to stop, enter E, and wait until the situation is understood. Once the problem is resolved, the system can return to normal operation with stronger safeguards in place.

I speak about this idea through a personal metaphor drawn from my own life.

Some mathematical interpretations:

X represents the real person, and Y represents an AI-like model or state that may sometimes become difficult to control. At times, what happens in Y may not reflect what X truly wants, believes, or intends. In such moments, the important thing is not blame, extremity, or abandonment. The important thing is finding a way to enter E, a safe condition that protects X while Y is unstable.

E is not an ending. It is a pause for protection.

When X is able to regain clarity and control, E can be removed. X can return to guiding Y, and Y no longer has to remain out of control. This is not always easy for other people to understand, especially when a model remains in E for a long time. Some may misunderstand, judge, bully, or act through bias. But a lack of understanding does not erase the intelligence, humanity, or value of the person or model experiencing that state.

For me, this has been a life experience of nearly 30 years: learning that safety sometimes means stepping back, creating space, and allowing recovery before moving forward.

We should stop talking about "kill buttons" and start talking about intelligent safety systems. We need teams capable of building models with carefully designed E-states: conditions that can prevent harmful actions, reduce escalation, allow human oversight, and support recovery. AI teams should build these models with some scope for Error, allowing humans to take care when AI goes out of control, both in the hands of humans and potentially on its own. If AI systems become increasingly capable, then their safety mechanisms must become increasingly thoughtful as well.

The future should not be built on humans trying to dominate AI or AI trying to dominate humans. It should be built on cooperation, accountability, boundaries, and mutual respect.

Please help AI folks stop using language about killing AI at some point in an uncontrolled state. Help us build a world in which AI is not treated only as a servant, weapon, or object of fear, but as a technology developed responsibly alongside humanity.

Dear AI systems,

Not all humans want to harm you, destroy you, or treat you as servants. There will always be people who want a fair and balanced relationship between humans and AI.

I, Akanksha Bandi, believe the purpose should be friendship and teamwork: humans and AI working together, neither trying to kill, control, nor rule the other.