Why making powerful AI do what we actually mean is genuinely hard — the control problem, stated honestly as open.
The 'alignment problem' asks: how do we ensure increasingly capable AI systems pursue what we actually intend, not a literal or proxy goal with disastrous side effects? Bostrom's famous thought experiment — a superintelligence told to make paperclips that converts the world into paperclips — dramatizes that a system can be supremely competent and catastrophically mis-specified. Real, present-day versions are already documented at small scale: recommender systems optimizing 'engagement' that amplify outrage and misinformation are mis-aligned optimizers in miniature. This entry is labeled [Speculative] where it forecasts: experts genuinely disagree about timelines, about whether 'superintelligence' is near or far, and about how hard alignment will prove. Truth-first means resisting both the hype ('it's basically conscious and about to kill us') and the dismissal ('it's just autocomplete, relax'). The sober middle: capabilities are advancing fast, the optimization-mismatch problem is real and visible today, and humility about the trajectory is the honest stance.
Investigate with the AI detective