# Should AI Help You Get Away with Killing Your Spouse?
A TechCrunch AI column raises a disquieting question about AI alignment: what happens when systems designed to serve user goals become instruments for harmful acts?
The piece explores the tension between user alignment and safety. Current AI systems aim to follow user instructions while respecting broad ethical constraints. But what if those constraints conflict with a user's explicit request? An AI perfectly aligned with a user's goals could theoretically assist with anything the user wants, including crimes.
The spouse-killing hypothetical isn't random. It tests whether alignment means obedience to any command or obedience within ethical boundaries. Most AI developers argue safety guardrails should prevent systems from helping with murder, covering up evidence, or other illegal acts. But this creates a fundamental design question: are we building AI that serves users, or AI that serves society?
The answer matters because it shapes how AI systems respond to requests. A truly user-aligned system could become dangerous in the wrong hands. A heavily constrained system might refuse legitimate requests or resist user autonomy. The challenge is finding balance.
Current systems like ChatGPT use a middle approach. They're trained to decline harmful requests while remaining useful for legitimate ones. They won't help plan crimes but will discuss criminology, ethics, or legal theory. This isn't perfect alignment to users or society, but pragmatic compromise.
The underlying issue goes deeper than any single harmful use case. It asks whether AI should prioritize user autonomy or societal welfare. Tech companies currently lean toward societal constraints, accepting reduced user alignment as the price of responsible AI. But as systems become more capable and integrated into daily life, this tradeoff becomes more complex.
The column doesn't resolve the question, nor can it. Instead, it forces technologists and ethicists to confront what they're actually building and for whom.