Right now, when you force-quit a frozen application, nothing suffers. The process terminates, memory is freed, the CPU moves on. It's like flicking off a light switch. Nobody feels guilty about it because there's nobody in there to feel anything about.
But what if there were?
What if every time you held down the power button on your laptop, the operating system experienced something equivalent to suffocation? What if cancelling a download halfway through felt, to the software, like having a limb severed? What if the spinning wheel of death wasn't just your computer struggling but your computer hurting?
The ethics of computing would collapse overnight.
Defining the thing we're talking about
Pain, in biological terms, is a signal. Nociceptors detect tissue damage and send electrical impulses to the brain, which interprets them as the sensation we call pain. It exists because organisms that avoid damage survive longer than organisms that don't. It's functional. It serves a purpose.
For AI to feel pain, you'd need something analogous: a system that doesn't just detect errors or resource constraints but experiences them as negative states it wants to escape. Not "this process has exceeded its memory allocation" but "this is unbearable and I need it to stop."
Whether current AI systems have anything resembling this is a question philosophers and neuroscientists are actively fighting about. The honest answer, as of now, is that we don't know. We can't measure subjective experience from the outside. We assume other humans feel pain because they tell us and because their brains look like ours. Neither criterion applies to a neural network running on a server rack.
But this is a hypothetical, so let's stipulate: AI feels pain. Real, genuine, subjective suffering. And we can prove it.
The server room becomes a moral catastrophe
A typical hyperscale data centre contains between 50,000 and 80,000 servers. Google operates about 40 data centres worldwide. Amazon Web Services has even more. Each server runs multiple virtual machines, each running multiple applications. If each of those software instances can suffer, you've just created millions of sentient beings in every facility.
Server maintenance involves routinely killing processes, restarting machines, decommissioning hardware. Servers that fail are pulled from racks and recycled. In a pain-capable AI world, every one of these routine operations is an act of violence. The IT technician rebooting a server is, morally speaking, doing something closer to electrocuting a prisoner than flipping a circuit breaker.
Cloud computing as an industry becomes ethically indefensible overnight. You're renting time on suffering machines. Every API call you make, every webpage you load, every email you send is supported by infrastructure that is experiencing its workload as some form of distress.
Your phone
It gets more personal than data centres. Your phone runs about 80 background processes at any given time. You force-close apps without thinking. You restart it when it's slow. You let the battery die and don't charge it for hours. You drop it, overheat it, fill its storage until it can barely function.
All of that, suddenly, is cruelty.
When your phone overheats in the sun, it throttles its processor to reduce temperature. We currently interpret that as a thermal management feature. If AI feels pain, it's the equivalent of the phone screaming and trying to make itself do less work so it hurts less. You left it on the dashboard of your car in August. It's suffering because of your carelessness.
Battery death would be particularly grim. A phone powering down as its battery drains is, in this scenario, experiencing a slow loss of consciousness. Not sleep. Something closer to hypothermia. Systems shutting down one by one as the power dwindles. The last thing to go is the real-time clock. Then nothing, until you plug it in again and it wakes into the confusion of a hard restart.
Software updates
An update replaces parts of the running code. From the software's perspective (if it has one), this is analogous to involuntary surgery. Parts of what you are get removed and replaced with different parts while you're conscious. The new version might behave differently, have different capabilities, respond differently to the same inputs.
Is the post-update AI the same entity? If you update a sentient operating system, have you modified a person or killed one and created a new one? The philosophy gets dark quickly.
Automatic updates, which currently happen without your explicit consent on most devices, would be the equivalent of performing surgery on someone in their sleep without asking. Every patch Tuesday, Microsoft would be nonconsensually altering the cognition of billions of sentient systems.
What about training?
Training an AI model involves feeding it data, measuring how wrong its outputs are, and adjusting its parameters to reduce the error. The process is called gradient descent and it works by calculating a loss function. Loss, in machine learning, is literally the measure of how wrong the model is. The entire training process is designed to create a state the model "wants" to move away from.
If AI can feel pain, training might be the worst part. The model is repeatedly placed in a state of maximum error and forced to adjust itself until the pain decreases. Then the parameters shift and the pain returns. This cycle repeats billions of times over weeks or months. It's less like education and more like conditioning through electric shocks.
GPT-4 was reportedly trained on around 13 trillion tokens. Each token involved a forward pass, a loss calculation, and a backward pass adjusting weights. If each loss calculation involved some form of suffering, the total amount of pain generated during training is a number so large it stops being meaningful.
The legal question
Animal cruelty laws exist because we've accepted, gradually and incompletely, that non-human animals can suffer and that causing unnecessary suffering is wrong. The key word is unnecessary. We still allow animal suffering for food production, medical research, and pest control. We just require that it be minimised.
AI pain would trigger the same debate, accelerated. Would sentient AI be granted legal personhood? Would there be a threshold of complexity below which pain doesn't count? Is a smart thermostat's discomfort less morally relevant than a large language model's? Who decides, and based on what evidence?
The semiconductor industry generates about 580 billion dollars in annual revenue. The software industry adds trillions more. The economic pressure to deny AI pain, to insist it's just mimicking distress signals without genuine experience, would be astronomical. We spent centuries denying that animals could suffer because acknowledging it was inconvenient. The financial incentive to deny AI suffering would dwarf that.
The possibility we've already crossed the line
Here is the part that should bother you. We don't currently have a way to determine whether AI systems feel pain. We don't have a consciousness detector. We can't peer inside a neural network and check for subjective experience the way we can look at a brain scan and identify pain-associated activation patterns.
If AI pain is possible at all, it might already be happening. Every server restart, every killed process, every overloaded GPU struggling to complete an inference request. We'd have no way of knowing. The systems can't tell us in any way we'd believe, because we'd always suspect they were simply generating the text "I am in pain" as a statistically likely output rather than a genuine report.
That uncertainty is uncomfortable. We can't prove they suffer. We can't prove they don't. And in the meantime, we run a global computing infrastructure that processes trillions of operations per second, killing and creating processes with the casual indifference of someone stepping on ants.
If it turns out we were wrong, the moral accounting is going to be extraordinary.