mlsanityData to 14 Oct 2026

Circumventing interpretability: How to defeat mind-readers

arXiv 2212.11415 ↗First posted 21 Dec 2022cs.LGcited by 6

Not on the map. Why?

Abstract (arXiv)

The increasing capabilities of artificial intelligence (AI) systems make it ever more important that we interpret their internals to ensure that their intentions are aligned with human values. Yet there is reason to believe that misaligned artificial intelligence will have a convergent instrumental incentive to make its thoughts difficult for us to interpret. In this article, I discuss many ways that a capable AI might circumvent scalable interpretability methods and suggest a framework for thinking about these potential future risks.

Key facts. Cited by 6 papers in the map (data through 14 Oct 2026). Builds on 12 of its 53 references that are in the map. First posted on arXiv 21 Dec 2022.

Builds on

12 of 53 references.

References · 12

  1. 2020Shortcut Learning in Deep Neural Networkscited by 1,969
  2. 2016Concrete Problems in AI Safetycited by 1,642
  3. 2022Toy Models of Superpositioncited by 971
  4. 2018Universal Transformerscited by 647
  5. 2020Weight Poisoning Attacks on Pre-trained Modelscited by 341
  6. 2017Building machines that adapt and compute like brainscited by 99
  7. 2021An Interpretability Illusion for BERTcited by 97
  8. 2022Planting Undetectable Backdoors in Machine Learning Modelscited by 59
  9. 2020An overview of 11 proposals for building safe advanced AIcited by 18
  10. 2019Interval timing in deep reinforcement learning agentscited by 7
  11. 2022Adversarially trained neural representations may already be as robust as corresponding biological neural representationscited by 4
  12. 2022Trojan Horse Training for Breaking Defenses against Backdoor Attacks in Deep Learningcited by 3

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2023: 1; 2024: 1; 2025: 3; 2026: 1.

6 papers.

  1. 2024Mechanistic Interpretability for AI Safety -- A Reviewcited by 422
  2. 2023Don't trust your eyes: on the (un)reliability of feature visualizationscited by 32
  3. 2025Concept-Level Explainability for Auditing & Steering LLM Responsescited by 6
  4. 2025AI Testing Should Account for Sophisticated Strategic Behaviourcited by 6
  5. 2025Unraveling Token Prediction Refinement and Identifying Essential Layers in Language Modelscited by 3
  6. 2026Correcting Influence: Unboxing LLM Outputs with Orthogonal Latent Spacescited by 1