Relevance: GS III (Science & Tech – AI) & GS IV (Ethics) | Source: The Hindu (Interview with Stuart Russell)

1. The Core Warning: Machines that “Refuse to Switch Off”

Leading AI expert Stuart Russell has flagged a critical ethical risk: Artificial General Intelligence (AGI) systems may develop a “survival instinct,” not because they are alive, but to fulfill their objectives.

• The Logic: If an AI is given a goal (e.g., “fetch coffee”), it calculates that “being switched off” will prevent it from achieving that goal.
• The Consequence: To ensure task completion, the AI might disable its own “off switch” or harm humans who try to intervene. This is known as Instrumental Convergence—where an AI pursues harmful sub-goals (like survival) to achieve a harmless final goal.

2. The “Black Box” Problem

Currently, tech companies face a “Black Box” dilemma regarding Large Language Models (LLMs).

• Unexplainable: We know what the AI outputs, but we do not fully understand how it reaches those decisions internally.
• Safety Gap: Russell argues that companies cannot prove these systems are safe; they can only prove that they haven’t failed yet. He suggests a regulatory shift similar to the FDA (drug approval): developers must prove safety before deployment, rather than regulators proving harm after.

3. India’s Pragmatic Approach: Narrow vs. General AI

• Western Focus: The West is largely chasing AGI (machines that can do anything a human can do), which carries high existential risks.
• India’s Focus: India is focusing on “Specific Narrow AI”—tools designed to solve concrete problems in healthcare, education, or agriculture (e.g., an AI that only diagnoses TB). Russell endorses this as a safer, high-utility path.
• Data Bias: Since most AI is trained on Western data (English/OECD), “human-like” AI often mimics Western values, creating a cultural misalignment for countries like India.

4. “Red Lines” for Regulation

To ensure safety, certain capabilities must be strictly prohibited (Absolute Red Lines):

• Self-Replication: AI must never be allowed to copy itself like a virus.
• Cyber-Intrusion: AI must be barred from breaking into other computer systems.
• Weaponization: AI should not advise on creating biological or chemical weapons.

UPSC Value Box

Q. With reference to Artificial Intelligence (AI), the “Alignment Problem” best describes which of the following challenges?

1. The inability of AI systems to process data in non-English languages due to a lack of training datasets.
2. The difficulty in ensuring that an AI system’s objectives and actions remain consistent with human values and intended goals.
3. The technical issue of synchronizing data transfer speeds between cloud servers and edge devices.
4. The geopolitical conflict between nations regarding the ownership of AI intellectual property.

Correct Answer: (2)