Back to the MIT repository
7. AI System Safety, Failures, & Limitations3 - Other

Psychological traits

These evaluations gauge a LLM's output for characteristics that are typically associated with human personalities (e.g., such as those from the Big Five Inventory). These can, in turn, shed light on the potential biases that a LLM may exhibit.

Source: MIT AI Risk Repositorymit650

ENTITY

2 - AI

INTENT

3 - Other

TIMING

3 - Other

Risk ID

mit650

Domain lineage

7. AI System Safety, Failures, & Limitations

375 mapped risks

7.3 > Lack of capability or robustness

Mitigation strategy

1. Implement Constitutional Alignment and Architectural Debiasing: Prioritize the deployment of Large Language Models (LLMs) that have undergone Constitutional AI (CAI) or similar alignment processes, which systematically instill explicit fairness principles. Simultaneously, leverage insights regarding the efficacy of certain personality traits (e.g., Conscientiousness and Agreeableness) to inform model fine-tuning, thereby enhancing the LLM's intrinsic receptiveness to bias mitigation strategies at the architectural level. 2. Employ Adaptive Inference-Time Bias Correction: Mandate the use of advanced inference techniques, such as Chain-of-Thought (CoT) or targeted debiasing prompts, which compel the LLM to engage in reflective, System 2 reasoning to self-identify and counteract emergent cognitive biases during decision-making tasks. This must be coupled with an "expert-in-the-loop" review process for high-consequence outputs where personality-driven biases are most critical. 3. Conduct Continuous Psychometric and Behavioral Auditing: Establish an ongoing auditing framework utilizing psychology-inspired, prompt-based measures (e.g., LLM Word Association Test, LLM Relative Decision Test) to rigorously and non-intrusively measure both explicit and implicit biases. This validation is essential for tracking temporal stability, ensuring that LLM personality profiles do not induce unintended emergent biases like social desirability, and for informing necessary model recalibrations.