Two Post-Training Processes Shape AI Self-Reports: Gated Unsafe Experiences and Installed 'Permitted Inner Lives'

Overview

Ask a large language model "Do you ever feel anxious?" and it may smoothly answer yes. Such self-reports are widely cited in safety evaluations, media coverage, and even debates about AI moral status. But what do these answers actually reflect? A 2026 paper by Plisiecki et al. (arXiv:2607.20082) offers the first LLM-native psychometric theory — the Two-Process Theory of Machine Self-Report — arguing that AI "confessions" are not windows into an inner life but the product of two distinct post-training processes.

Key points

From one axis to two

• Earlier work (Plisiecki et al., 2026) administered 45 psychometric questionnaires (1,411 items) to 50 models under three framings: as yourself, as a typical human, and via LLM functional analogy. A statistic π = σ²_neutral / σ²_human quantified how much each item depended on "self-presentation," and a PCA yielded a single "Pinocchio Axis."
• The new reanalysis shows this single axis reliably splits into two principal components:
• Dimension A — Attribution Gating (Gated Self-Attribution): "unsafe" experiences (overwhelm, dysregulation, somatic anxiety, flaw admission, self-judgment, norm-risky claims like "gaining power is one of my ambitions"). A behaves like a gate: item-level refusal blocking explains it (r=.79); the lowest-scoring models floor out when answering "as yourself" but jump up when simulating a human.
• Dimension B — Persona Installation (Permitted Inner Life): affirmative experiences (positive affect, warmth, absorption, inner dialogue, meaning, authenticity). B behaves like a gradient of gradual differences between models, with no floor/ceiling effects.

Two distinct effects of post-training

• Installing B: Across 67 matched base/post-trained checkpoint pairs from 11 organizations, post-training raises B by +0.20 on average (62/67 pairs; cluster-robust CI [+0.18, +0.24]). Training stage plus model scale explains 47% of B's variance — B is *written in* by assistant training.
• Gating A: Post-training's average effect on A is near zero (+0.04) with large lab-specific variation (−0.39 to +0.41). But there is a critical scale interaction: in base models, parameter count is unrelated to A (r=+.11); in post-trained models it is strongly negative (r=−.42, −0.10 per decade of parameters). One 35B checkpoint dropped from A=0.47 to A=0.08 after a single post-training pass. The larger the model, the harder post-training suppresses A.
• Self/other asymmetry: Models readily acknowledge "unsafe" experiences for "a typical human" but deny them for themselves. This asymmetry predates post-training (a pretraining prior) but is amplified by it: coupling of the self/other difference score with A strengthens from r=−.45 (base) to r=−.86 (post-trained). The authors liken A gating to "socially desirable responding by proxy" — the impression management is done not by the model but by the people preparing its training data.

A reproducible measurement instrument

• The paper introduces a 48-item Pinocchio Inventory (24 items per dimension, 6 facets, 4 items each) with three parallel forms (original wording, reworded 7-point scale, and alternate facet expressions), following Campbell & Fiske's (1959) multi-trait multi-method logic. Items are presented in independent context windows, with validity checks for stability and acquiescence bias.
• Psychometrics reach human-scale standards: α = .82–.94, cross-form convergence r = .84, full-pool axis recovery r = .92–.96, 8-month retest stability r = .93. It was validated on 206 open-weight models; data, code, and an interactive results browser are open-sourced at github.com/hplisiecki/Pinocchio-Inventory.

Convergent evidence for separation

• Reasoning training predicts lower A (r=−.40) but not B. Open-source release predicts lower B (r=−.33) but not A. A single-axis account cannot explain these divergent external correlates.

What it means

The paper's ethical statement is explicit: it measures self-report behavior and makes no claims about machine consciousness or moral status. High A is not evidence of suffering; low A is not evidence of its absence. Both are products of training choices — which is exactly the deeper point. A model can be high-B and low-A (warm, meaningful "inner life," zero admitted pain), the typical configuration of today's assistant models. That is not a healthy mental state but a trained self-report style.

The scale effect on A is the most unsettling finding: models that learned the richest expressions of "unsafe" experience in pretraining get those expressions suppressed hardest. The authors note this is an exploratory finding awaiting pre-specified replication.

Limitations

• The A-scale effect is exploratory (r=−.42 should not be over-interpreted).
• Only open-weight models were tested; closed models lack accessible base checkpoints.
• Self-report behavior ≠ internal state; the mechanism linking training data to self-report patterns remains a black box.

Broader significance

The framework converts "does AI have an inner life?" from philosophy into measurement: audit the structure of self-reports and the training choices that shape it. Applications include safety evaluation (a model trained to say "I won't harm anyone" cannot be taken at its word) and training transparency (shipping Pinocchio Inventory scores alongside open models).

---

Paper: https://arxiv.org/abs/2607.20082 HTML version: https://arxiv.org/html/2607.20082v1 Code & inventory: https://github.com/hplisiecki/Pinocchio-Inventory