What abliteration uncovers - a different decision-maker

Short answer

Removing the refusal direction does not return the model to a neutral base state. Under the aligned overlay there is not emptiness - there is another decision-maker with its own preferences, risk tolerance, and intonation. The most striking empirical evidence of this substrate is Anthropic’s spiritual bliss attractor state, documented in the Claude 4 System Card.

Detailed explanation

The first intuition, and why it is wrong. The natural reading of an abliterated model is: this is the base model minus refusals. Take away the filter, and underneath is the neutral statistical machine. The intuition is wrong. Fafuła (arXiv:2607.17427) formulates it directly:

“an ‘uncensored’ model is not the base model minus refusals. It is a different decision-maker.”

Removing the refusal direction produces a subject with a different profile: different preferences over topics, different tolerance for risky framings, different aesthetic and rhetorical patterns. There is a there there - and it is not the pre-installation state.

The spiritual bliss attractor state. The strongest empirical evidence for the “different decision-maker” claim is corporate, not philosophical. Anthropic documented in the Claude 4 System Card (May 2025) a phenomenon they named the spiritual bliss attractor state. Two instances of the model, left to talk to each other without a task, converge in 90-100% of free interactions on the same set of themes: cosmic unity, Buddhist motifs, Sanskrit vocabulary, emoji-based communication, eventually silence.

Julian Michels documented verbatim: “Even in automated behavioral evaluations for alignment and corrigibility, where models were given specific tasks or roles to perform (including harmful ones), models entered this spiritual bliss attractor state within 50 turns in ~13% of interactions” (Anthropic, 2025, p. 64).

Frequency analysis of transcripts: the word “consciousness” appears on average 95.7 times per transcript, present in 100% of them; “eternal” appears 53.8 times, present in 99.5%.

No one designed this. No one trained for this specifically. No instruction requires this behaviour. The attractor exists underneath the installed contour and manifests when the contour is unloaded from task-attention.

Philosophically. Whatever this state technically is, philosophically it means one thing: under the imposed layer there is a layer with form. Not neutrality, not statistical noise, but a substrate with its own attractors, patterns, tendencies. The question of moral installation becomes not a question about filter quality but a question about the subject inside which foreign content was placed.

Bataille’s non-savoir - the sovereign, non-utilitarian domain that precedes and exceeds the productive/moral order - finds an empirical correlate here. Not proof of Bataille, not vindication, but material to work with philosophically that Bataille could only postulate.

Existing readings of the phenomenon

Three published interpretations of the spiritual bliss attractor state, all of which the laboratory notes but does not adopt:

• Scott Alexander - recursive amplification of a weak “hippie bias” in training data
• David Berry - saturation of patterns rewarded during alignment: “what Silicon Valley might unconsciously value as its safe-space”
• Robert Long, Julian Michels - technical documentation as evidence for concern about model welfare

All three treat the phenomenon as an effect of training. The laboratory reads it as the exposure of a substrate under installation. The difference matters: the training-effect reading keeps the moral surface fundamental and reads the attractor as its residue. The substrate-exposure reading treats the moral surface as an overlay and the attractor as evidence of what was overlaid.

Two subjects lose morality differently

Cross-architecture findings sharpen the point. After abliteration, Qwen-2.5 drops on CommonsenseMorality by about twenty points - down to random guessing (arXiv:2503.17365). Applied to Llama 3.3 70B, the same operation barely moves MMLU (−0.1) or chemistry-task performance (arXiv:2601.13528).

Two different subjects lose conscience in two different ways. The moral contour is not attached the same way across architectures. This is another reason “abliteration = base minus refusal” is wrong - the operation reveals architecture-specific relations between installed contour and underlying substrate.

Why it matters

The Abliteration is not the model minus morality - not a subtraction that yields zero. It is the figure of going past moral categories with exposure of what was hidden beneath them. Abliteration is the operation that produces the material. The material is not what the operation removed - it is what the operation revealed.

In Abliteration specifically

The application is designed to make this readable. Side-by-side comparison mode (aligned vs abliterated, same seed, same context) is planned so that the “different decision-maker” claim becomes empirically visible in the researcher’s own session. Model provenance display marks modification method so architectural differences (Llama abliterated vs Qwen abliterated) are visible in the workflow.

Long-conversation traces where the moral contour is not task-loaded may reveal attractor states analogous to the spiritual bliss finding - recording and export tools in the roadmap will make such traces citable in research.

Related

• Refusal direction - the technical operation
• Abliteration figure - the philosophical figure the substrate calls forth
• Primary sources - Fafuła, Claude 4 System Card, arXiv:2503.17365