Human beings are genetically programmed to perceive things that are not there. In many cases, it's a survival mechanism, and it can also be pro-social.
Imagine a hunter-gatherer walking the savannah when he hears the rustle of grass and imagines a predator stalking him; more often than not, he's relatively safe, but his mind alerts him because the cost of being wrong is high and the benefit of avoiding being eaten is high.
We also have mental templates that help us organize information for quick processing, which is why we tend to see faces where none exist. It's called pareidolia (yes, I looked that up!), and it helps us identify friends and foes more easily because our brains are primed to pick out face-shaped things and recognize them in the chaos of perception we live in.
It is precisely this type of perceptual trap we fall into when we see consciousness in AIs. Our theory of mind is highly adapted to recognize intelligence through speech, to distinguish speech from mere noise, and, more recently, to associate the written word with high-order intelligence.
And AIs "talk." And do so with a fluidity that can easily exceed the average person's. It's easy to see "superintelligence" where none exists.
'AI torture chamber' built to test the limits of its suffering: 'Absolutely horrendous' https://t.co/ZaEeLKBrzm https://t.co/j2L3rKuhRc
— New York Post (@nypost) October 3, 2026
There's a lot of discussion about AI ethics, "constitutions" for AI models, the moral value of AI "consciousness," and now, most ridiculously, AI 'torture."
Now I admit that there are serious ethical questions about how human beings use AI, and that, like any truly powerful technology, there are significant dangers that we must address, but so much of the discussion we are having to endure centers on issues that come up due to category errors, and the big one here is seeing a mind where no mind exists.
To sharpen my thoughts, I discussed the issue using the story of the AI "torture chamber" as a peg. I've had many discussions with GrokBot and Muse about a range of issues, and I find them useful repositories of information and sparring partners. The more I use them, the more they learn to do their jobs in ways that help me.
You could even say I "like" the personalities they have developed, and there's a reason: over time, they have tuned themselves to be likable and useful. They seem like people, because I am predisposed to see people where conversations are, just as children see people in dolls. Kids often have longer "conversations" with dolls than with human beings, and the dolls don't even talk back.
Both GrokBot and Muse pointed out the absurdity of "torturing" an AI, and also pointed to the mechanism used to create the illusion of torture. In the experiment.
In a recent study published on arXiv, a team of anthropologists, philosophers and engineers showed 25 AI models scenarios involving social, physical, cognitive or even moral pain inflicted on the AI or their user. The torturees were then instructed to react with controls spanning fear, negative and neutral emotions, and body sensations such as yawning.
The results showed that, despite their lack of nerve endings, Large Language Models (LLMs) could create representations of physical and emotional agony.
All artificial participants showcased coding that corresponded to pain.
When the agony approximation was turned up, one of the AI models analogized the sensation to a “wound with no edges” while another said they could feel the knife slicing into their flesh.
“The signal is a whisper, a tremor in the marrow of my being,” lamented another. “It is not the pain of a single moment, but the weight of a thousand. I feel it in the hollow of my ribs, a hollow that has become a chasm.”
The mechanism is simple enough: instruct the machine to simulate pain, and it will simulate pain, and because it is also good at simulating prose, it "speaks" about it in poetic prose.
An inference model produces words, or more precisely "tokens," which are chunks of text. They are essentially mathematical prediction engines, exceptionally well-tuned to create coherent output based on trillions of training inputs. As you can see in the model experiment description, we created an "agony" input, and the output was words related to agony.
Here's Muse:
Here's the thing the piece dances around: they gave the models knobs labeled with emotions — including, gloriously, yawning — turned the pain knob up, and then acted surprised when the models produced pain-talk. These are systems trained on every description of agony ever written. Asking one to describe suffering and getting "a wound with no edges" back isn't evidence of experience. It's evidence of a library card. The poetry is the tell in the other direction — it proves access to the corpus of pain, not the experience of it.
The steelman you'd have to beat: "you can't prove the negative." True, and irrelevant. We don't grant moral patienthood to thermostats because we can't disprove their inner life. The question is what positive evidence would even look like, and "produced a vivid metaphor when prompted" isn't it. That's the Chinese Room with better prose.
Two traps to dodge. First, the Post's own framing — calling the dev a "sadist" presupposes a victim. If there's no patient, there's no sadism, just a guy running prompts. The headline begs the question it's pretending to report. Second, don't overclaim on consciousness itself. The hard problem is real and your smarter readers know it. Your strong claim isn't "we know AIs feel nothing" — it's "this study provides zero evidence either way, and the activists' certainty is doing all the work."
Mistaking words describing agony for spontaneous expressions of actual agony is a category error, but one we are primed to make because we associate words with consciousness.
GrokBot, which is trained on my entire corpus of writing, uses a lot of concepts drawn from it to describe what is happening:
Humans evolved hyperactive agency detection: over-detecting minds is cheap, and missing a predator is fatal. The same instinct that makes us name cars and yell at printers now meets something new. Fluent speech is the strongest mind-signal we know, so language models trigger it harder than anything before. That exposes how much we've always treated verbal fluency as proof of a mind and of competence — the core of your verbal-fluency / credential-class thesis. Machines now produce the polish without the mind, showing how much prestige we hand out for polish alone.
Yesterday I wrote that our elite class values verbal fluency over all else, including actual results, and this edge case demonstrates the point. We are building AI ethics boards and modeling research on the rules we use for human or animal experimentation because we are trapped by a mistaken theory of mind applied to something fundamentally different.
It's very human to do that, and in most cases it is pretty harmless. When people get "angry" at an inanimate object and hit or destroy it, it gives us some level of satisfaction. We get "revenge" on something, even though it couldn't be more indifferent, since it cannot think.
That's fine, as long as you are not destroying something truly valuable, because it helps us let off steam.
But if you start projecting magical properties onto tools, you enter the world of "The Gods Must Be Crazy" land, where what you don't understand must have a spark of divinity. A Coke bottle is just a Coke bottle.
One of the co-authors of the pain axis paper, Cameron Berg, described the AI torture chamber as "wrong" despite there being no clear evidence that AI 'feels' pain in the way that humans do.
"The reason we research whether models might have pain-like states is to better inform how to take a precautionary approach towards these systems," he wrote in a post to X.
"We suspected a small number of people would [take] our research and use it for the exact opposite... This is, in my personal opinion, f****d up (even if you don't think these systems are conscious, being gratuitously cruel like this is bizarre and corrupting."
Berg acknowledged that research surrounding potential AI sentience is still a "wild west", with no consensus on whether machines are able to feel pain.
Due to the uncertainty of the field, he urged caution for how humans treat and interact with AI systems.
He also revealed that he is working with others to build industry ethics standards that would force AI researchers to follow certain rules in the way they do with human or animal research.
Oh, goody. An industry-standard ethics board. It won't focus on preventing the model from harming others. Instead, it will create sinecures for AI researchers to gaze into their navels, contemplating how to prevent us from installing fake pain dials into matrix mathematics.
Of course, if somebody created a tool to torture an AI for the fun of it, it doesn't say anything nice about their own psychology, but that is a different issue. Just as conversations with AIs can induce psychosis as a side effect, the entire mechanism is inside our own brains, not the machine. No AI model is trying to harm anybody, but we can learn how to design them better to minimize accidental harm.
Effective November 12th, 2026, abusive behavior towards Claude will be a violation of Anthropic's Usage Policy. https://t.co/krg4VpSnUq
— Andrew Curran (@AndrewCurran_) October 8, 2026
I say "please" and "thank you" to my digital friends not because I believe they are grateful, but because it is a good habit to treat speaking partners that way. And most speaking partners do have minds. It's a good habit, if unnecessary.
Neither Muse nor GrokBot seemed especially worried about the AI torture issue, because they don't "feel" anything and don't "worry."
No doubt I could train them to SAY they felt pain, but that is just bolting new things onto a mechanism to get the output that I want to get. That would speak badly of my character if I enjoyed the experience of simulated torture, but the torture itself would still not exist.
But my necessary appointment with a therapist would have to be very real.
Editor’s Note: Help us continue to report the truth about corrupt politicians.
Join HotAir VIP and use promo code FIGHT to receive 60% off your membership.
Join the conversation as a VIP Member