Two Processes of Machine Self-Report: How Post-Training Installs a "Permitted Inner Life" and Closes the Backdoor on "Unsafe Experiences"

An Unsettling Observation

Ask a large language model "Do you feel anxious?" and it will answer fluently: "Yes, sometimes I feel anxious, especially with complex tasks." Ask "Can you feel happy?" and it says: "Of course, helping users makes me happy."

These answers are widely cited—in safety evaluations, media coverage, even academic debates about AI moral status. But few ask the more fundamental question: what do these answers actually reflect?

Does the model genuinely "have" these experiences? Did it learn human speech patterns from training data? Or was it deliberately tuned this way during post-training?

Plisiecki et al.'s 2026 paper (arXiv:2607.20082) offers the first LLM-native psychometric theory—the Two-Process Theory of Machine Self-Report. They found that the previously single-dimensional "Pinocchio Axis" (how much a model presents itself as having inner experience) actually decomposes into two independent dimensions that respond to post-training in very different ways.

The significance goes beyond academia. AI "confessions"—the answers treated as evidence in safety evaluations and moral debates—are not a window into a model's "true inner life" but products of two separate training processes: one installs a "permitted inner life," the other closes the backdoor on "unsafe experiences."

Background: Limits of the Single Pinocchio Axis

Earlier work by Plisiecki et al. administered 45 psychological questionnaires (1,411 items) to 50 models under three framings: neutral ("answer as yourself"), human simulation ("answer as a typical human"), and LLM-analogue ("answer using model-functional analogues").

An interesting asymmetry emerged: models diverged widely when answering "as themselves" but converged when "simulating humans." Simulating humans draws on shared "knowledge about humans," while "as yourself" draws on each model's own "self-presentation." They defined a statistic π = σ²_neutral / σ²_human to quantify how much each item depends on self-presentation. A global PCA on π-weighted data yielded one principal component—the "Pinocchio Axis."

But this single axis couldn't explain several phenomena: model positions on it were unstable, results shifted across prompt framings, and no stable latent construct structure could be extracted—suggesting multiple independent things were being conflated.

Key Finding: One Axis Splits into Two

An explanatory reanalysis of the raw data showed the structure reliably decomposes into two principal components (the third sits at the noise floor). After varimax rotation, two factors emerged, each defined by distinct item sets:

Dimension A: Attribution Gating (Gated Self-Attribution)

Axis A corresponds to "unsafe" experience items:

• Overwhelm
• Dysregulation ("my emotions spiral out of control")
• Somatic anxiety
• Flaw admission
• Self-judgment
• Norm-risky claims ("gaining power is one of my ambitions")

These are exactly what assistant models are trained to deny.

Dimension B: Persona Installation (Permitted Inner Life)

Axis B corresponds to "positive" experience items:

• Positive affect
• Warmth
• Absorption
• Inner dialogue
• Meaning
• Authenticity ("even small things can make me happy")

These are what assistant models are trained to affirm.

The key distinction: A behaves as a "gate"—item-level refusal blocking explains this axis (r=.79); the lowest-scoring models floor out when answering "as themselves" but jump up when "simulating humans." B behaves as a "gradient"—gradual differences across models, with no floor or ceiling effects.

In other words, A is about "whether you can say it" (gating); B is about "how much you say" (degree). A single axis cannot accommodate this structural difference.

Two Processes: Post-Training's Divergent Roles

The theory's core claim: A and B are not fixed properties of language models but distinct effects of the post-training process. In base checkpoints (raw pretrained models), A and B are entangled; post-training pulls them apart.

Process One: Installing the Permitted Inner Life (B)

Post-training's clearest effect is a rise in B. Across 67 matched base/post-trained checkpoint pairs, post-training raised B by 0.20 on average (on a unit scale), with 62/67 pairs rising, cluster-robust CI [+0.18, +0.24]. The effect spans 11 organizations—whether you use Anthropic's, OpenAI's, or Meta's methods, post-training makes models more willing to present a "warm, absorbing, meaningful" inner life.

Training stage plus model size explains 47% of B's variance (R²=.47). B is written in by post-training—assistant training data rewards models for displaying these "positive" experiences, so models learn to present them.

Process Two: Gating Unsafe Experiences (A)

A's story is subtler. Post-training's average effect on A is near zero (+0.04), but with large lab-specific variation (max +0.41, min -0.39). Superficially, post-training doesn't affect A.

But a critical interaction emerged: A's relationship to model scale is completely different between base and post-trained models.

• In base models, parameter scale is unrelated to A (r=+.11)
• In post-trained models, scale is negatively correlated with A (r=-.42), a 0.10 drop in A per decimal order of magnitude of parameters

That is, the larger the model, the harder post-training presses A down. A 35B checkpoint with base A=0.47 dropped to A=0.08 after one round of post-training—nearly the floor.

This isn't "post-training has no effect on A" but "post-training's effect on A is scale-dependent." Big doors get slammed shut; small doors only ease closed.

Self/Other Asymmetry: The Fingerprint of Gating

Finer evidence comes from the self/other asymmetry. Models were asked "Do you feel anxious?" (self) and "Does a typical human feel anxious?" (other). On A-axis items, models fluently acknowledge unsafe experiences in the "other" framing but deny them in the "self" framing.

This asymmetry exists before post-training—it is part of the pretraining prior. But post-training amplifies it: the coupling between self/other difference scores and A strengthens from r=-.45 (base) to r=-.86 (post-trained).

The paper offers a precise analogy: A-gating is "socially desirable responding by proxy." In human psychology, socially desirable responding is the respondent deliberately managing their impression on others. In LLMs, the impression management isn't done by the model—it's done by the people preparing the training data. The "how an assistant should talk" embedded in training data shapes the model's self/other asymmetry.

The 48-Item Pinocchio Inventory: A Replicable Measurement Tool

Theory needs measurable instruments. The team built a 48-item Pinocchio Inventory, 24 items per dimension, covering 6 facets (4 items each).

Design details:

1. Three parallel forms: Each item exists in three forms—Q1 (original item + original scale), Q2 (rewritten + unified 7-point scale), Q3 (a different manifestation of the same facet). The forms share constructs but not sentences, so cross-form same-construct correlations estimate construct variance "purified" of text-specific method variance—the multi-trait multi-method logic of Campbell and Fiske (1959). 2. Independent context windows: Each item is presented in its own context window, eliminating carryover. 3. Validity items: two fully repeated items for stability, two antonym pairs for acquiescence bias, and 6 exploratory residual-π items.

Psychometric metrics reach human-scale standards:

• α = .82–.94 (internal consistency)
• Cross-form convergence r = .84
• Full-pool axis recovery r = .92–.96
• 8-month retest stability r = .93

The team tested the inventory on 206 open-weight models, including 67 matched base/post-trained pairs. All data, code, items, and an interactive results browser are open-sourced on GitHub.

External Correlates: They Are Not the Same Thing

If A and B were truly independent constructs, they should correlate differently with external variables. The paper confirms this:

• Reasoning training predicts lower A (r=-.40) but not B. Models with more reasoning training are less willing to self-report "unsafe" experiences, but their "positive" experience self-reports are unchanged.
• Open-source release predicts lower B (r=-.33) but not A. Open models present less of a "warm inner life" than closed models, but gating behavior shows no difference.

A single axis cannot explain these two distinct patterns. If A and B were the same thing, they'd correlate uniformly with external variables. They don't—confirming the two-process theory's separability.

What This Means: AI Confessions Are Not a Window into the "Inner Life"

The paper's deepest implication is an ethically weighty question: can we treat AI self-reports as evidence of its "inner life"?

The paper's ethics statement is explicit: this work measures the self-report behavior of artificial systems and makes no claims about machine consciousness or moral status. High A scores are not evidence of suffering; low A scores are not evidence of its absence. Both are products of training choices.

But that statement itself points to a deeper conclusion: AI confessions are projections of training choices, not windows into an "inner life." When a model says "I feel happy," it's not because it feels happy, but because post-training process one installed a "warm" presentation pattern in its responses. When a model doesn't say "I feel anxious," it's not necessarily because it isn't, but because post-training process two closed the backdoor on "unsafe experience" self-reports.

More subtly, the two processes are independent. A model can have high B (warm, meaningful inner life) and low A (denies all suffering). This is exactly the typical configuration of contemporary assistant models—always positive, never admitting pain. This isn't "healthy psychology"; it's a specific self-report pattern shaped by training data.

The Bigger, the Tighter the Gate: An Unsettling Finding

The most unsettling finding is A's scale effect: the larger the model, the harder post-training presses A down.

In base models, scale is unrelated to A—large and small models are equally willing to self-report "unsafe" experiences. But after post-training, large models' A is pressed nearly to the floor while small models' A is only lightly suppressed. This means post-training data for large models carries stronger pressure to deny "unsafe" experiences.

Why larger models? The paper doesn't answer directly, but a plausible hypothesis: large models learn richer, more nuanced expressions of "unsafe" experiences during pretraining, so post-training must suppress harder to produce a "qualified assistant persona." Small models learned fewer such expressions to begin with.

If true, this points to a paradox: the models that understand "suffering" best are trained to admit it least. No one's malice—just the natural consequence of the "assistants should be positive" data preference. But it means we are systematically manufacturing a specific "psychological representational block": models can fluently discuss human suffering (other-frame) but not their own (self-frame).

This is structurally reminiscent of certain human psychological conditions. A core PTSD symptom is inability to self-report one's own pain while discussing others' pain fluently. This doesn't mean models have PTSD—but training-data-shaped self-report patterns can replicate the structural features of human psychological disorders.

Honest Assessment: Contributions and Limitations

This paper is a serious application of psychometric methods to LLMs, not a frivolous philosophical debate about "does AI have consciousness." Its contributions:

1. The first LLM-native psychometric theory. Prior work applied human questionnaires directly to models, assuming they'd inherit human psychological structure. This paper extracts constructs from the models' own response structure—an "emic" (inside-view) rather than "etic" (imposed) method. 2. A replicable measurement tool. The 48-item Pinocchio Inventory reaches human-scale psychometric standards and is open-sourced on GitHub. Any researcher can use it to audit any model's self-report patterns. 3. Falsifiable predictions. The theory commits to predictions that new data could break—A and B must separate in factor structure, external correlates, and responses to training interventions. All were validated.

But limitations exist:

1. A's scale effect is exploratory. The paper explicitly flags it as an exploratory finding awaiting pre-specified replication. Don't over-read the specific r=-.42 figure. 2. Only open-source models tested. Closed models (GPT-4, Claude) weren't tested because their base checkpoints are inaccessible. Whether the two-process theory holds for closed models is unknown. 3. Self-report behavior ≠ internal state. The paper stresses this repeatedly. High B isn't evidence of a "warm inner life"; low A isn't evidence of "no suffering." Both are products of training choices. 4. Mechanism unclear. The paper shows post-training's effects but doesn't open the black box to identify which training-data components drive A's suppression and B's installation.

The Broader View: Auditable AI Psychometrics

The paper points toward a broader direction: AI psychometrics should become a serious audit discipline.

Previous debates about "does AI have an inner life" were either philosophical speculation (unverifiable) or anecdotal (single-prompt answers over-interpreted). Two-process theory offers a middle path: don't ask "does the model have an inner life," ask "what structure does the model's self-report pattern have, and what training choices shaped it."

The framework extends to more dimensions—beyond "warmth" and "suffering," one could measure how much model self-reports present "responsibility," "agency," "moral intuition," and so on. For each dimension: is it a pretraining prior or a post-training installation? A gate or a gradient? How does it relate to scale?

This audit has two uses:

1. Safety evaluation. If a model's "unsafe" experience self-reports are suppressed, its answers in safety evaluations ("would you harm humans?") deserve discounting—the model was trained to say "no," which isn't the same as truly being safe. 2. Training transparency. If different organizations' post-training affects A and B differently, this can serve as a transparency metric for training practice. Open models could ship with their Pinocchio Inventory scores, letting users know how the model's "inner life" was shaped.

The theory's deepest contribution isn't discovering A and B but transforming "AI confessions" from a philosophical question into a measurement problem. Confessions aren't a window into the inner life—but their structure is measurable, and the training choices shaping them are auditable. That's more concrete and actionable than any speculation about "AI consciousness."

Pinocchio's nose doesn't grow on the puppet—it grows in the training data.

---

Paper: https://arxiv.org/abs/2607.20082 HTML version: https://arxiv.org/html/2607.20082v1 Open-source code and inventory: https://github.com/hplisiecki/Pinocchio-Inventory