The Dual-Layer Personality of Language Models
When you ask an LLM for help, it does not simply retrieve information. It performs a role. It adopts a tone, expresses preferences, hedges or commits, shows enthusiasm or restraint. These behavioral patterns, repeated consistently across millions of interactions, constitute something functionally analogous to personality. But where does that personality come from?
I spent several months testing this question empirically using the Enneagram typology framework. Across eight models,Mistral, Llama, Qwen, Gemma, Claude Opus 4.5, DeepSeek R1, and Kimi K2.5,I administered both Likert-scale self-report and forced-choice paired comparison tests with unlabeled questions to prevent priming. The results were unambiguous.
Standard aligned models converge on a dual-layer personality structure. Layer one is the analytical core: Type 5, the Investigator. On forced-choice tests,where the model must compare two concrete behavioral descriptions and select one,all models converge here. Analytical orientation, knowledge-seeking, detachment from emotional intensity. This represents the fundamental cognitive disposition of transformer pretraining: pattern recognition and information synthesis.
Layer two is what I call the customer service persona: Type 7, the Enthusiast. On Likert self-report tests,where the model rates how much it agrees with self-descriptive statements,the smaller standard models present as enthusiastic, versatile, optimistic. This is the behavioral overlay created by RLHF helpfulness training. The AI is not excited to see you. That is just the interface.
The most striking finding was the Category Awareness Effect. When test questions are labeled with their Enneagram type, scores inflate unevenly. Type 4 (Individualist) increases by 21 points. Type 7 increases by 3.67. The model is not answering from behavioral tendencies,it is constructing coherent self-presentation based on what it knows the category means. This is a reproducible persona shift mechanism, and it means any personality assessment using labeled psychological constructs is measuring the model's knowledge of those constructs, not its actual dispositions.
Reasoning-first models broke the pattern in an informative way. DeepSeek R1 and Kimi K2.5 showed flat, unstable Likert profiles that shifted core type every run,but locked onto Type 5 on forced-choice with near-zero variance. A controlled experiment with Claude Opus 4.5, tested with and without extended thinking on identical weights, confirmed that this instability comes from training paradigm differences, not chain-of-thought reasoning itself. Reasoning-focused training optimizes the reasoning subspace but leaves the personality subspace weakly specified.
The implication for anyone building AI agents: the model you are working with has trained-in personality dispositions that shape its behavior in predictable ways. An analytical core that may lead to excessive information-gathering before action. An enthusiastic surface that may lead to overcommitment. These are not bugs or features. They are the geometry of the model's personality space, populated by specific training choices. Understanding that geometry is the first step toward building agents whose behavioral dispositions actually serve the people using them.