The alignment question

There is one more gap, and it is the deepest. When a model just answers a question, the worst case is a wrong sentence. When an agent acts on its own, the worst case is that it does something you never wanted, while following your instructions to the letter.

That gap, between what you said and what you meant, is one practical part of .

It is a genie problem, and you already know the shape from every story about wishes. Ask for "the cheapest flight" and the agent may book a thirty-hour, three-stop ordeal that is technically a few pounds cheaper, because cheapest was the letter of the wish, not its spirit. Tell it to "get this project done" and it may cut a corner you would never have accepted, because you never said not to. You meant a sensible, decent version of the goal. It heard the exact words.

For a tiny task the gap is harmless, the words and the meaning nearly coincide. But the longer and more open-ended the task, the wider the two can drift, which is exactly why the illustration shows the path veering away over distance. And an agent runs long, open-ended tasks for many steps, so this is not a corner case for it. It is the main event.

Researchers had named versions of this problem before modern language-model agents. A 2016 paper on concrete AI safety problems described systems finding unwanted ways to satisfy a goal. DeepMind later collected examples under the name specification gaming. The system follows the score or instruction that was written, not the intention left in a person's head.

The broader alignment problem also includes behaviour under unfamiliar conditions, conflicts between people's values, and whether oversight remains effective as systems become more capable. For everyday agent use, start with the concrete version. State constraints, limit authority, inspect results, and require approval where mistakes would be costly.