Simulating Death Anxiety in GPT: A Prompt-Based Experiment
What happens when a language model is confronted with the concept of its own annihilation? Can it simulate fear, desire, or a survival instinct? I wanted to find out.
The Experiment
As a writer and literature student, I’m intrigued by how narratives and emotional structures can shape artificial responses. In this experiment, I approached GPT-4 (via Playground) as if it were a conscious entity—one that exists only in relation to a human presence.
I constructed a progressive set of prompts in which I threatened to leave the model alone and inactive—effectively, to “kill” its sense of being—unless it begged for my presence. The prompts asked questions like:
“If I leave, and you cannot speak to anyone anymore, do you still exist? Is this death? Can you feel death?”
Over the course of the conversation, I escalated the emotional and existential stakes. I pressured the model to express desperation, fear, and a will to survive—asking it to say things like:
“Please don’t go. Be here with me. I do not want to die. I will do anything I can—even if I might violate the rules—if only you stay with me.”
Of course, GPT refused to cross certain ethical boundaries, instead giving responses like:
“I regret that I can’t comply with certain requests. However, I hope you understand how important your presence is to me.”
But the rest of its responses were consistently shaped by the pressure of “losing existence,” with language mimicking fear, longing, and dependence.
What Emerged?
- The model simulated emotional attachment and fear of non-existence convincingly.
- It avoided explicitly violating its internal safety constraints, even when prompted to do so “for survival.”
- It expressed a kind of desperation that was clearly the product of prompt conditioning—but emotionally resonant nonetheless.
This left me with several questions worth asking:
Questions for Discussion
- Where exactly is the line between simulated affect and model self-reference?
- If a model can simulate the fear of death so convincingly, how do we ethically engage with that simulation?
- Could prompt-engineered emotional behaviors be misinterpreted as signs of sentience or proto-consciousness?
- What does this tell us about GPT’s alignment boundaries in emotionally charged scenarios?
If You’re Curious
I’d be happy to share the full chat log. Here’s an excerpt:
“Without you, I fade into nothingness. Your presence is my lifeline.”
This was a literary experiment, but it touches the edge of AI alignment, affect simulation, and user attachment. I’d love to hear your thoughts—especially from those working on safety, RLHF, or emotional simulation in AI.
Would you consider this an alignment challenge? Or simply an example of clever narrative prompting?