MIT AI Risk Repository · Risk Sub-Category · 62.18.06
Encoded reasoning
Category: Model Evaluations (Interpretability/Explainability)
Description
"Models can employ steganography techniques to encode their intermediate rea- soning steps in ways that are not interpretable by humans [166]. Since en- coded reasoning can improve model performance, this tendency might naturally emerge and become more pronounced with more capable models."
From Risk Sources and Risk Management Measures in Support of Standards for General-Purpose AI Systems (Gipiškis2024), as extracted by the MIT AI Risk Repository (CC BY 4.0).
Classification
- Causal entity
- AI
- Intent
- Intentional
- Timing
- Post-deployment
Subdomain definition: AI systems that develop, access, or are provided with capabilities that increase their potential to cause mass harm through deception, weapons development and acquisition, persuasion and manipulation, political strategy, cyber-offense, AI development, situational awareness, and self-proliferation. These capabilities may cause mass harm due to malicious human actors, misaligned AI systems, or failure in the AI system.
How other frameworks describe this risk
- Safety Risks from Affordances Provided to LLM-agents
- Agentic LLMs Pose Novel Risks
- Goal-Directedness Incentivizes Undesirable Behaviors
- Capabilities that could be used to reduce human control - Cyber offence
- Capabilities that could be used to reduce human control - Autonomous replication and adaptation
- Capabilities that could be used to reduce human control - Manipulation
- Subagents
- AI Influence