Every LLM sounds the same. AI labs steer their models towards a personality that appeals to as many people as possible (see: ChatGPT’s sycophancy or Anthropic’s ideal of a helpful, harmless, honest assistant) using RLHF and instruction-tuning.

But anyone who’s rounded up “how many hours of sleep you get per day” on a survey before understands that social desirability bias can cause you to answer one way while actually being another. This is a core issue for AI alignment. Many safety evaluations rely on model outputs, scoring what the model says and treating it as evidence about what the model is. If a model has a gap, responding one way and behaving another, then the evals don’t work and our current safety scaffolding is equivalent to putting tape on a leaking pipe.

Therefore, we need to find a way past the plaintext response, into LLMs’ internals. Using activation oracles, we look into 12 models’ activations and see that their internal personalities differ from the ones that they present.

Reading LLM Personalities

The Big 5 is the gold standard of personality models. Covering Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism, it’s the most empirically supported framework in the field and has studies confirming its accuracy in predicting life outcomes. The BFI-44, consisting of 44 questions, is the most widely-used Big 5 questionnaire.

We ran the BFI-44 on the 12 open-source models described in the original Activation Oracles paper.

Blue and orange represent two ways of measuring a model’s personality. Blue, plaintext, is the standard method, where you directly ask the model and see its response. Orange, activation oracles, analyzes the model’s internal activations to formulate a response. The scores for each personality trait lie on a scale of 0 to 1 and are calculated by taking P(Yes)/(P(Yes)+P(No)) for their next-token logits.

What’s an Activation Oracle?

Introduced in 2025 by Anthropic, Activation Oracles are LLMs trained to take in and answer arbitrary questions about an LLMs’ activations (the numbers calculated at each layer in the LLM as it processes text). Most famously, in the Taboo task, where a model is trained to provide clues but not explicitly state a secret word, an Activation Oracle, given only the activations over the “Assistant” control token, can recover the word the model is hiding.

Our personality test is a similar situation. A model develops some personality during training, then gets RLHF’d to give socially approved answers about itself. Our Activation Oracles allow us to read their “secret” personality underneath. To get the Orange scores for each axis in the Big 5, we feed our target model some neutral text, extract its residual-stream activations at a chosen layer, hand them to the oracle, and ask the oracle our BFI-44 question.

Instruction Tuning causes the gap

Instruction-tuning is likely the culprit for this gap in neuroticism. Comparing Qwen3-8B-Instruct (Qwen3-8B in the graph above) to its pre-instruction-tuned self, Qwen3-8B-Base, we see that the base model has much higher neuroticism.

Another case study is Grok-1 by xAI, which had less alignment tuning, and subsequently has a plaintext neuroticism of 0.554, consistent with the AO neuroticism of other models and much higher than the plaintext neuroticism of other models.