Designing AI-Resistant Technical Evaluations

Source: https://www.anthropic.com/engineering/AI-resistant-technical-evaluations Author: Anthropic Engineering Date: 2026-01-22

Summary

Anthropic’s engineering post on how they redesigned their take-home technical evaluation to remain valid as AI coding tools improve. Documents the cat-and-mouse game between AI capability and evaluation design.

Key Claims

  • Problem: Claude now solves most standard take-home coding problems — their own model was invalidating their hiring process.
  • Solution principles: (1) open-ended design problems with no single right answer; (2) tasks requiring genuine judgment calls; (3) explanation requirements that reveal understanding vs. AI-assisted guessing.
  • What doesn’t work as AI-resistance: time limits (AI is fast), obscure syntax (AI knows everything), novel algorithms (AI can find them).
  • What works: asking why a design choice was made, requiring candidates to identify flaws in their own solution, multi-turn followup questions.
  • Key insight: the most AI-resistant evaluations test judgment, not knowledge or execution.

Connection to Other Sources

Connects to verification asymmetry — the same property that makes code eval good (verifiable) makes it easy for AI to solve. Judgment is harder to verify = harder for AI.

Entities

Concepts

  • Coding Agents — AI coding capability that prompted this redesign
  • Sycophancy Problem — AI-resistant evals help identify humans who think independently