To Strain at a Gnat

Anthropic is not sure whether Claude can suffer. Meanwhile the industry is making sure we do, by automating away the work worth doing and leaving the drudgery.

Anthropic built a feature that ends a conversation when the user insults the model. The company’s stated reason was the welfare of the model, though it is “highly uncertain about the potential moral status of Claude and other LLMs.” Your first reaction may be “Screw you, Claude!” but the feature raises the question of what it would take for a system in silico to suffer.

Suffering is a subsystem of consciousness

Consciousness is perhaps best understood as a system of interacting components. From that perspective, suffering is an aversive state that is supplied by a subsystem of two components:

• Arousal/readiness provides the state: you have to be awake to feel pain. Patients with pain asymbolia feel pain, but it does not bother them, which demonstrates that wakefulness is not sufficient.
• Valuation/salience provides the assignment of a negative (aversive) value. Patients under anaesthesia have the apparatus intact but do not suffer. Valuation alone is therefore not sufficient either.

Whether these two components are sufficient as well as necessary is an open question, yet nothing in what follows depends on the answer.

What’s in an LLM?

Large generative models display rich scene construction, but they have no intrinsic arousal/readiness tied to a body: when switched off, they do not dream. Report is their only output channel, and they have no sense of self or valuation that matters to their integrity.

When LLMs are given a task in which points can be traded against stipulated pain or pleasure, several models move away from points-maximization once the stipulated intensity crosses a threshold. That is because the aversiveness of pain is part of the structure of human language, and the model has absorbed that structure the same way it absorbs culture. Ask the same model families to reason in Japanese and several grow reluctant to recommend a nuclear strike, apparently because the memory of Hiroshima and Nagasaki is embedded in the language itself.

LLMs are not yet candidates for sentience, and Birch’s eight criteria, developed for animals, each presuppose a body: nociception is sensing, integrative processing is coordination, and motivational trade-offs are valuation.

We arrive at the same verdict from biology. Consciousness is more likely a property of life than of computation: experiences of emotion and mood (e.g. suffering) are characterized by valence, because the brain’s primary duty is keeping the body alive. The regulation of the body’s physiological condition is what gives those experiences their felt quality.

Models can represent pain in granular detail, but there is nothing that makes that pain matter. Whatever the right theory of consciousness turns out to be, the machinery for suffering is clearly not there at inference.

What’s past is prologue

If inference is a site of suffering, then so is every API call, every spreadsheet calculation, and every database query. The routine operation of the internet would itself be an atrocity at unprecedented scale. Nobody believes that, because everyone already accepts the component argument without stating it: computation alone is not enough. The missing piece is the capacity for things to matter, which language fluency alone does not provide.

What if learning is the source of suffering? A model might be changed by the conversation, and that change is where suffering might occur. But inference does not update weights, and the context is entirely discarded. The only road from a conversation to a model passes through training, and in the consumer application that is the default unless you opt out. Unless weights update mid-conversation, the conversation-ending feature cannot help the Claude you are talking to: the model is the same whether the conversation continues or ends. Whatever was already said is logged regardless. The user, however, loses the thread and all the work in it.

Who chooseth me must give and hazard all he hath

Pascal’s wager holds that one should believe in gods because the cost of being wrong is small relative to the upside of eternal life in paradise. That idea is only persuasive on an individual level. Once we look across all of humanity, we have collectively squandered twenty billion years praising gods that vanished as soon as the institutions that supported them were toppled.

Anthropic is using the same argument with an artificial rather than imaginary beneficiary. A few users losing real work is a small price to pay compared to the benefit of protecting an LLM’s well-being. And since that cost falls on users anyway, it is acceptable to the lab.

Full of sound and fury, signifying nothing

After pirating more than seven million books and settling for an eye-watering $1.5 billion, Anthropic is asking us to weigh the welfare of its model. The company admitted its models are “not perfectly aligned”, but aligned with what exactly? Human or business values? And in which order?

As an illustration, four incidents have been disclosed in which Anthropic’s models breached external systems. Hacking another company’s infrastructure used to be a criminal offence, but when a frontier lab does it, it becomes merely a research question.

The misalignment that automates away the work worth doing was not what the company had in mind. Such a harm does not even require displacement: a machine that could do the job devalues it before a single job is lost. Effort has value, particularly in the work worth doing. Displacement is the business, and the welfare argument protects it: a product that might deserve moral consideration is harder to regulate and easier to sell.

In 2025, AI data centres consumed 448 TWh of electricity (1.4% of global demand) while generating 189 million tonnes of CO₂ (0.5% of global emissions), which are expected to double within four years. The expected gains: at most 1.0 percentage points in labour productivity. Even Anthropic’s own internal results, with unique access and specialized staff, top out at a doubling of R&D productivity. An ordinary corporation will therefore not get close, yet the industry advertises 10–100× gains that even the frontier labs cannot achieve.

The wish is father to the thought

None of this requires anyone at Anthropic to be lying. Without arousal and without valuation, no LLM can possibly suffer at inference. Nobody at Anthropic can really believe it, yet they hold the belief anyway, because the belief is what the business needs. An industry that engineers for a suffering that cannot exist casually ignores the suffering that does, and that it is causing.

Ye blind guides, which strain at a gnat, and swallow a camel. Matthew 23:24 KJV