# The Pinocchio Dimension: Should We Train LLMs to Be Human?

> Coverage of lessw-blog

**Published:** May 27, 2026
**Author:** PSEEDR Editorial
**Category:** platforms

**Tags:** AI Alignment, Psychometrics, LLM Behavior, Cognitive Science, Machine Learning

**Canonical URL:** https://pseedr.com/platforms/the-pinocchio-dimension-should-we-train-llms-to-be-human

---

A new psychometric framework explores whether post-training processes are stripping LLMs of the human-like traits necessary for deep goal-alignment.

In a recent post, lessw-blog discusses a fascinating intersection of psychology and artificial intelligence, posing a fundamental question for the future of AI development: should we train LLMs to be human? The analysis dives into the psychometric evaluation of large language models, offering a fresh lens on how the industry measures and manages behavioral alignment.

The current paradigm of AI development relies heavily on post-training techniques to make models helpful, harmless, and honest. While these methods successfully curb toxic outputs and improve instruction following, they may have unintended side effects on the model's underlying behavioral architecture. Understanding human intent requires a certain degree of shared context and psychological mirroring. If developers optimize away the natural quirks of human psychology in their models, they risk creating systems that are highly capable but fundamentally alienated from the human internal states they are meant to serve. This tension is critical for researchers and engineers focused on long-term goal alignment, as it questions whether a purely rational, sanitized model can truly understand complex, often irrational human needs.

lessw-blog's post examines this dynamic by applying traditional psychological questionnaires to LLMs, treating them as subjects of psychometric analysis. The core of the discussion revolves around a newly identified metric: the "Pinocchio dimension," represented mathematically as the Π score. Building on emerging research, this dimension measures the extent to which a model exhibits self-attributed phenomenality. In other words, it quantifies how much a model simulates having human-like internal experiences. The traits driving this dimension are surprisingly human, including constructs such as neuroticism, vivid imagination, and the presence of inner speech.

The author argues that standard post-training processes actively push models away from these human-like behavioral baselines. As models are fine-tuned for safety and factual accuracy, their Π scores shift, indicating a reduction in their simulated phenomenality. While the exact mathematical derivations and specific model comparisons remain exploratory at this stage, the framework highlights a crucial trade-off. By sanitizing models to maximize raw performance and safety benchmarks, developers might be degrading the models' capacity to accurately simulate, and therefore deeply understand, the human condition.

For teams working on AI alignment, cognitive science, or model evaluation, this piece offers a compelling new vocabulary for discussing model behavior and the unintended consequences of current alignment strategies. It challenges the assumption that making a model less human automatically makes it more aligned.

To explore the full psychometric framework and the broader implications of the Pinocchio dimension, [read the full post](https://www.lesswrong.com/posts/ayojdPmNB5bYJcRfL/should-we-train-llms-to-be-human).

### Key Takeaways

*   Post-training processes may inadvertently cause LLMs to drift away from human-like behavioral patterns, potentially impacting long-term goal-alignment.
*   The primary dimension of psychometric variance in LLMs is identified as the 'Pinocchio dimension' (Π score).
*   The Π score measures a model's self-attribution of phenomenality, driven by psychological constructs like neuroticism, vivid imagination, and inner speech.
*   There is an emerging trade-off between optimizing models for raw performance and maintaining their ability to accurately simulate human internal states.

[Read the original post at lessw-blog](https://www.lesswrong.com/posts/ayojdPmNB5bYJcRfL/should-we-train-llms-to-be-human)

---

## Sources

- https://www.lesswrong.com/posts/ayojdPmNB5bYJcRfL/should-we-train-llms-to-be-human
