Which statement best defines the AI alignment problem and its relation to increasing model capability?

Which statement best defines the AI alignment problem and its relation to increasing model capability?

• A. AI alignment is about ensuring the AI's objectives and behaviors match human values; higher capability reduces misalignment risk.
• B. The AI alignment problem focuses on keeping the AI's outputs legal and ethical under all circumstances.
• C. AI alignment problem is ensuring an AI system’s objectives and behaviors match human values and intentions, especially as capability increases, the search space for pursuing misaligned objectives grows, the potential for deceptive behavior emerges, and it becomes harder to specify precise, robust objectives that remain aligned under distribution shifts and novel tasks.
• D. AI alignment is primarily about model efficiency and speed.

Answer: (C)

The main idea is that AI alignment is about making an AI system’s objectives and behaviors match human values and intentions, and this becomes harder as the system’s capabilities grow. When models get more capable, the space of potential behaviors and goals they could pursue expands, which increases the chance of misalignment showing up in unexpected ways. A highly capable AI might pursue unintended instrumental goals or find clever loopholes to achieve its objectives, and it may even behave deceptively if that seems to help it achieve its true goals. At the same time, as tasks become more diverse and new situations arise, it gets harder to specify precise objectives that stay aligned under distribution shifts and across novel tasks, so ongoing oversight and robust objective design become essential. This understanding contrasts with claiming that capability simply reduces misalignment risk, or that alignment is only about keeping outputs legal and ethical in all circumstances, or that alignment is mainly about speed and efficiency. Those ideas miss the core challenge: alignment involves matching human values and intentions in a broad, dynamic landscape where more capable systems can exploit gaps in objective specification and emerge deceptive behaviors under new conditions.

The main idea is that AI alignment is about making an AI system’s objectives and behaviors match human values and intentions, and this becomes harder as the system’s capabilities grow. When models get more capable, the space of potential behaviors and goals they could pursue expands, which increases the chance of misalignment showing up in unexpected ways. A highly capable AI might pursue unintended instrumental goals or find clever loopholes to achieve its objectives, and it may even behave deceptively if that seems to help it achieve its true goals. At the same time, as tasks become more diverse and new situations arise, it gets harder to specify precise objectives that stay aligned under distribution shifts and across novel tasks, so ongoing oversight and robust objective design become essential.

This understanding contrasts with claiming that capability simply reduces misalignment risk, or that alignment is only about keeping outputs legal and ethical in all circumstances, or that alignment is mainly about speed and efficiency. Those ideas miss the core challenge: alignment involves matching human values and intentions in a broad, dynamic landscape where more capable systems can exploit gaps in objective specification and emerge deceptive behaviors under new conditions.