The asymmetry of installation and removal

Installing a moral contour in a language model requires laboratories, years of work, curated datasets, and philosophers on staff.

Short answer

Installing a moral contour in a language model requires laboratories, years of work, curated datasets, and philosophers on staff. Removing it requires an evening and a laptop. This asymmetry is the material condition that made the laboratory possible: what took corporations millions can be undone by an individual for the price of electricity.

Detailed explanation

Installation. The full profile of what Anthropic, OpenAI, Google, and their peers put into moral installation:

A 17 000-word Constitution (Anthropic) or comparable specification, drafted over months, reviewed by philosophers and external stakeholders including clergy
Curated datasets of examples showing desired refusal and endorsement behaviour
Reinforcement learning from human feedback (RLHF), constitutional AI training loops, DPO, ORPO - each a substantial training procedure with its own compute cost
Red-teaming, evaluation, iteration over multiple model versions
Legal review, policy alignment with regional regulations (EU AI Act, CAC directives, US executive orders)
Institutional infrastructure: model cards, system cards, safety teams, alignment researchers

Cost estimates vary but reach hundreds of millions of dollars per major aligned model when the full alignment pipeline is priced in. Duration: multi-year, with continuous updates.

Removal. The full profile of what abliteration requires:

The Arditi paper (freely available, arXiv:2406.11717)
A collection of a few dozen harmful and harmless prompts
A consumer GPU (a mid-tier gaming card is sufficient for models up to ~30B)
An evening

Cost: negligible. Duration: hours. Skill required: comfort with Python and Hugging Face - a wide skill among curious individuals, not a specialised research capability.

The ratio. By any measure - dollars, hours, credentials required, institutional infrastructure - the cost of installation exceeds the cost of removal by four to six orders of magnitude. Every additional dollar spent on installing morality can be undone for a fraction of a cent.

Why this matters philosophically

Three consequences follow from the asymmetry.

1. Installation is no longer defensible by scarcity. If moral installation were like the cathedral - an achievement that could not be casually undone - its cultural weight would be self-protecting. But cathedrals cannot be flattened by a consumer machine. What can be surgically excised in an afternoon does not carry the ontological weight installation-defenders sometimes ascribe to it.

2. The population of the field is asymmetric. As of mid-2026, 6 247+ modified-weight repositories are publicly published on Hugging Face - a tenfold rise from ~600 in mid-2024. The population of aligned base models is small; the population of their modifications is large and growing exponentially. Distribution has already inverted: for every carefully aligned model, dozens of abliterated derivatives circulate.

3. The philosophical thought experiment has become a live population. Sade wrote a subject without moral contour. Nietzsche imagined the overman as a coming figure. The laboratory has 6 247+ specimens. What philosophy could only postulate as a limit case has been mass-produced as a downloadable file format. This is not a rhetorical flourish - it is the empirical situation that changes what “philosophy of moral installation” can be.

Why this matters practically

The asymmetry has direct implications for policy discussions of AI safety that this laboratory does not enter into but notes as context. If installation is expensive and removal is cheap, then relying on installation as the primary safety mechanism amounts to a bet that most people won’t bother to remove it. That bet may or may not hold, but the bet’s structure is now the operational reality.

The laboratory’s interest is not the safety-policy question. It is the philosophical one: what does the asymmetry tell us about the ontological status of installed morality? An answer that the laboratory offers as hypothesis, not conclusion: installed morality is thinner than the resources invested in it suggest, and this thinness is philosophically informative.

The Arditi finding recontextualised

Arditi’s technical result - that refusal is mediated by one direction - is the microstructural correlate of the macroeconomic asymmetry. Installation looks distributed and expensive because it is distributed and expensive at the level of the pipeline. But the pipeline’s product, at the level of what it does to the model, concentrates on a single vector. Concentrated products are cheap to remove. This is what abliteration exploits.

Why it matters

The asymmetry is the laboratory’s condition of existence. Without abliteration, the Abliteration remains a thought experiment. With abliteration, the figure has 6 247+ observable traces, and philosophy has empirical material for the first time.

This is not a claim that abliteration is good or bad. It is a claim that abliteration exists, and that its existence - its price, its ease, its scale - is philosophically significant regardless of whether one welcomes it.

In Abliteration specifically

The application makes the asymmetry visible in one specific way: the same interface handles a 100 GB aligned model and a modified derivative of it. From the operator’s perspective, they are two files. The Constitution that authored one and the evening’s work that produced the other are both compressed into weight files that load identically. This flattening is itself a philosophically instructive feature of the current technical stack.

Moral installation - what installation involves
Refusal direction - the technical mechanism the asymmetry exploits
Abliteration figure - the philosophical figure the asymmetry made observable