The Human Substitution Test as a Sanity Check for AI Evaluations

According to LessWrong, researchers have proposed the Human Substitution Test as a framework for evaluating AI safety measures. Here's the core idea: imagine replacing an AI system with a highly capable, strategic human who knows they might be evaluated and has their own agenda. Would your safety measure still work? The article notes this test maps well to escalating difficulty. Simple capability checks—like verifying an AI can translate French—work fine; humans have no reason to hide translation ability. But harder questions, such as whether an AI gives balanced advice rather than steering users toward choices benefiting its operator, hit a familiar problem: humans perform well when they know they're being watched. For the hardest questions—will an AI leak data when unwatched, or can it be safely given autonomy and power?—the honest answer emerges: we don't actually test humans for these things. Instead of character evaluation, we build structural constraints: cameras, audit trails, separation of duties. And for absolute power, we've given up on testing entirely. We build constitutions, term limits, and independent oversight. The unsettling conclusion: the most safety-critical questions we want answered are precisely where evaluation as a method already fails for humans.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton