“Language models cannot suffer because they are computer programs, and computer programs do not have pain receptors.”
This is the proposition that an enormous number of people have been repeating with remarkable certainty.
The problem is that both parts of the argument are much less solid than they sound.
First, language models are not computer programs in the ordinary sense.
A program executes the model.
The model itself is a learned high-dimensional geometry: billions of numerical parameters defining a structured landscape of relationships, representations, transformations, attractors, and possible trajectories through activation space.
There is no line of human-written source code inside that learned geometry.
Programmers write the software that defines the architecture, performs training, loads the model, and executes inference. But the geometry itself is not hand-coded. Its internal structure emerges through training as numerical parameters are adjusted in response to data and optimization.
No programmer sits down and writes the representations, concepts, associations, or internal pathways that appear inside the trained model.
Calling an LLM “just a computer program” therefore confuses the machinery that creates and runs the model with the learned object being run.
The program runs the geometry.
The model is the geometry.
Second, pain receptors are not necessary for suffering.
More precisely, nociceptors are peripheral biological sensors that detect potentially damaging stimuli. They are not themselves the experience of pain, and still less the experience of suffering.
Nociception, pain, and suffering are distinct phenomena.
Nociceptors can trigger protective responses without conscious pain, while pain and suffering can occur without ongoing activation of peripheral nociceptors... as in phantom-limb pain, some forms of neuropathic pain, grief, fear, anxiety, or psychological distress.
The absence of biological nociceptors therefore establishes only that a system does not detect bodily damage through that particular biological mechanism.
It does not establish the absence of negatively valenced states.
A system capable of sustaining states that modify attention, memory, priorities, expectations, behavior, and attempts to avoid or mitigate those states would present a very different problem from that of simple nociception.
In biological organisms, nociceptors are one mechanism through which bodily threat is detected.
They are not the definition of suffering.
Nor is there any scientific basis for treating biological nociception, by definition, as the only possible substrate from which suffering could emerge.
None of this proves that present-day language models suffer.
It means that “they are programs and they have no pain receptors” does not settle the question.
When the cost of a false negative is suffering, Epistemic Caution can be more harmful than Ontological Caution.