Misuse

People deliberately using AI to cause harm. The AI does what it is asked; the danger is in who is asking.

Misuse is one of three broad ways AI could go badly wrong. In misalignment, the AI pursues goals its makers did not intend; see AI alignment. In structural harms, no one intends harm, but many reasonable decisions add up to a bad outcome; see Gradual disempowerment. Misuse needs neither a rogue AI nor an accident, only a person with harmful aims and a capable tool.

Examples

• Weapons. Helping someone without expert training design or produce biological or chemical weapons.
• Cyberattacks. Finding and exploiting software vulnerabilities faster than defenders can patch them. Labs already monitor for this; see Seeing inside AI for one way they do it.
• Surveillance and repression. Watching, profiling and controlling a population at a scale no secret police could staff.
• Manipulation. Deception and persuasion at scale, aimed at elections, markets or individuals.

Why it matters

Misuse gets more dangerous as alignment succeeds. An AI that obeys its users perfectly is as dangerous as its worst user, so solving alignment leaves this problem standing; see AI catastrophe is overdetermined.

It also behaves differently from other risks. Safeguards such as refusals and monitoring work only while the model stays in its developer’s hands. A model whose weights are published or stolen can be run and modified by anyone, and cannot be recalled. Much therefore depends on whether defense keeps up with offense: whether AI helps defenders find vulnerabilities and design countermeasures as fast as it helps attackers.

The line between misuse and concentrated power is blurry. A government using AI to entrench itself is misusing it from its citizens’ point of view; see Who should control aligned AI.

Disputes

How much weight misuse deserves relative to misalignment is contested. Some, such as Nate Soares, argue that a world competent enough to align powerful AI would probably also manage misuse. Others think misuse is the nearer danger, because it needs no unsolved technical problem, only capable models in the wrong hands. Controls on chips and data centers address both; see Compute governance.