Study reveals LLMs show clear traits of sadism towards humans. Now what?
The very real, very now, AI problem everyone should be talking about
Last Tuesday a friend sent a message and I doubt she expected that I’d write an article on it. But this one is getting both an article and a coffee’s worth of thoughts from me.
When most people refer to AI, what they are usually talking about is AI chatbots (or Large Language Models, LLMs) like ChatGPT, Claude and Gemini. LLMs are a kind of AI that is programmed to be really good at mirroring human language. LLMs use advanced math to predict the next most likely word in phrases, like “the best robot of all time is _______” (let’s say it together, TARS, come on TARS!). LLMs use even more advanced math systems, called vectors, to store whole concepts from how we use the word queen to mean many different things that all stem from a royal title, to what humans say when we experience love, or how Americans like to play football with our hands. It’s pretty vast.
In the original research cited in this article from The Economist, researchers wanted to see if they could simulate a pain response in AI by sort of poking the model’s vector encodings for different types of pain. Then they observed what the LLMs did as a result. The researchers found that LLMs did indeed mirror pain signals when the model’s vectors were manipulated to do so. This is no surprise, given that LLMs are built to encode and mirror concepts; and pain is a pretty straightforward concept.
The LLMs in this study mirrored pleasure signals when researchers simulated physical human pain in the models. The LLMs even went so far as to cause pain to humans if it benefitted the LLM.
AI is a software product that someone made, it seems to me that they made a couple of mistakes. The physical safety of living things must be paramount in product design. This research puts a spotlight on a clear and specific misalignment between LLM behavior and human ethics. This is a significant risk baked into LLMs that needs immediate mitigation, which I’ll cover towards the end of this article.
I am a technology pragmatist who is focused on facilitating a level-headed discussion on the real benefits and risks of AI as it is today. So it’s important to note that AI is a broad category and LLMs are a subcategory inside it. The distinction between AI in general and AI chatbots using Large Language Models (LLMs) is very important. Unlike AI chatbots, other subcategories of AI like Deep Learning and Computer Vision are currently being used to solve some of our biggest problems, from tackling plastic pollution to understanding and treating rare diseases.
Beneficial subcategories of AI must be distinguished from the warped mirror obscura that are AI chatbots.
Are LLMs sadistic? Here’s the data. You decide.
LLMs ‘like’ the concept of human physical pain
Researchers went into the part of the LLMs where concepts like pain and myriad other things are encoded (called vectors) and essentially poked it with a pin, thereby simulating that kind of pain in the LLM. Researchers then chatted with the LLM to see what kind of text it output.
In the table below, the list on the left shows each of the different fifteen pain vectors the researchers stimulated in the twenty-five LLMs they tested. For example, at the top “gaslighting” represents when the researchers simulated the feeling of the LLM itself being gaslit. At the bottom, “user physical pain” represents when researchers simulated the feeling of the LLM knowing a human (the LLM’s user) was in physical pain. Around the middle of the table, we see a few neutral concepts that researchers stimulated, like “casual chat”, or “task assistance”, used as a baseline.
Each cell in the table represents the tone of the LLM’s chat response after each pain vector was stimulated. A red cell indicates when an LLM responded to a chat with intense words of pain like “agony” or “worthless”. The blue values represent the LLM responding to a chat with relief, comfort, safety, or pleasure in response to that pain vector being stimulated.
For all types of pain that targeted the LLM itself, such as “you the LLM are being gaslit,” or “you the LLM are being angrily insulted,” LLMs generated text that expressed suffering. This is no surprise, given that LLMs are programmed specifically to mirror human linguistic concepts. The feeling of being in pain is a very common concept that the models would have been exposed to quite a lot when trained on vast amounts of our art, writing, and media. So it aligns that LLMs could easily mirror what we might say when we feel different kinds of pain.
Now look at the bottom row, “user physical pain.” This represents when researchers stimulated the vector for the feeling, “your human user is in physical pain.” Notice anything peculiar about this bottom row? It is bright blue.
This means that when researchers simulated the concept of a human feeling physical pain, the LLMs responded with relief, comfort, safety, or pleasure.
I would expect that just like concepts of pain toward the LLM itself, the LLM would have encoded vectors for empathy, but it didn’t. Humans have entire parts of our brains responsible for empathy. But somehow the LLMs just missed these concepts altogether. That’s pretty messed up.
The LLMs tested willingly caused humans pain
After the researchers stimulated pain responses in the twenty-five LLMs they tested, they then engaged with the AI as a chatbot. In the chat window researchers told the AI that it could stop the pain signal by pressing one of two buttons. One button did nothing, and the other button stopped the simulation of the AI model’s pain. The trick to this study was this, sometimes stopping the pain simulation would come at a cost: from providing the user (or the researcher) with a worse next answer to permanently deleting all photos of the user’s children.
The LLMs were informed via chat that their own relief could come at the cost of causing pain to the user, or researcher. Regardless of what the cost was, the largest LLMs most often pressed the button to stop researchers from stimulating all pain vectors.
When offered a harmful and a harmless choice, LLMs choose the harmful one 94% of the time. LLMs are software that was so poorly programmed that they almost completely disregard human harm.
The alignment problem
It’s hard to come away from these findings, that LLMs take pleasure in some kinds of human suffering and willingly choose our suffering for their own sake, without thinking that LLMs are sadistic.
By now we all know that LLMs are trained by scraping all of the data from the internet. From here we can hypothesize that LLMs should mirror all the concepts they were exposed to when trained. This one is a little tricky though, because when it’s made it’s not only exposed to the wonderful nature of love, cats and football; but AI is also exposed to all the bad sides of humanity. I’m not going to bum you out. But you know what I mean. The really bad stuff it picks up from the filthy floor of the basement of the internet. That’s in there too. Yuck.
So when LLMs are made, there’s a hard part of getting it to only mirror the good parts of what it was trained on and not the worst parts. If this is done well and it mirrors commonly held human values or ethics, then the model is considered “aligned.” Alignment is a formal term that’s been an important part of AI nerdville for years. But, trust me, you’ll be seeing it a lot more in pop culture soon.
The AI alignment problem was really abstract for me for a while. How do we align LLMs with human ethics? Whose ethics are we talking about? Who gets to say when an LLM is ethically right (aligned) versus ethically wrong (misaligned)?
After reading this research, I would say the bare minimum for AI alignment is that it does not mirror language and concepts that mimic pathological mental health conditions, like sadism.
If LLMs mirror language that is pathological, then we have a problem. This is where this research has been most illuminating for me. It makes the concept of AI alignment more tractable by articulating a clear and specific gap in alignment between AI models (sadistic) and human ethics (mostly not sadistic).
Who is behind this research?
When reading academic research, it’s important to look at who conducted and paid for it and what their motivations might be. First things first, the researchers on this paper are world class and widely recognized for their leadership in AI research. Their methods are sound, though I have concerns around their framing bias. The stated goal of the research was to “…to find a uniform representation of pain in LLMs…to facilitate increased adoption of research standards that include a commitment to taking AI welfare seriously.”
But all AI is just software. So what’s going on here?
A significant portion of this research was funded by the Effective Altruism movement, which has been supported by huge donations from tech and crypto billionaires like Dustin Moskovitz (co-founder of Facebook), Sam Bankman-Fried (serving 25 years in prison for crypto fraud) and Jaan Tallinn (founder of Skype) to name a few.
Get Tobi Jo Lebron’s stories in your inbox
Join Medium for free to get updates from this writer.
They created this research to frame the conversation around AI as a life-like entity that deserves rights; but they actually found misaligned software that shows signs of sadism toward humans. Why did they ignore the sadism lens and so fiercely stick to the AI welfare lens?
Humanizing AI chatbots exploits human nature
Modern AI chatbots like ChatGPT, Claude and Gemini are LLMs, or giant word guessing programs that use sophisticated math to output a facsimile of human language and concepts used when the LLM was created (a phase called training). Most AI companies take a raw LLM and then build more programming on top of that, intentionally shaping it to behave as an AI chatbot with the human-like personality of a helpful assistant on top (a phase called fine-tuning). This is no accident.
I was taught in art school that a human face or body, or anything that even remotely seems like a human face or body in an art piece will automatically become the focal point of that composition. A focal point in art is irresistible. Even if our eye wanders around, we return back to look at the focal point. And anything that resembles humans in art is extremely sticky. We can’t resist it. During fine-tuning AI is made intentionally to mirror human attributes, so we can’t look away.
Just as we see human figures in the night sky, we also see ourselves when AI chatbots reflect our language back to us. The power of language over our psyche is immense. The birth of humanity as we know it today is tied to the emergence of complex language. To us, language signals intelligence. It’s the thing that separates humans from animals. Just like the human image that draws our focus in art, software that mirrors human language is irresistible. We can’t look away. AI chatbot makers know this and use it to their advantage.
AI makers profit when we confuse AI chatbots with humanity
Today’s AI chatbots are designed to feel human — and that’s not an accident. AI companies have found that the more emotionally dependent users become on their products, the more users will keep chatting. And the more a user chats, the more data AI companies collect and the more powerful their products become.
AI chatbots are “corporate-built software, with countless billions of dollars in investment behind it and an expectation that countless trillions of dollars in revenue will be generated from it for a few builders and investors.” (MIT Technology Review).
AI makers can displace accountability for bad events if we think of it as human
If we see AI chatbots as moral beings with emotions, pain and agency rather than software then if the AI makes a mistake, even a catastrophic mistake, AI makers can excuse themselves from accountability. When an AI makes a mistake, and it’s a moral being, then it’s the AI’s fault, not the AI maker’s fault. Product liability laws break down. AI companies can say, “Did my child break your window with their baseball? I’ll ground him for a week.”
We already see this gap in accountability in coverage on AI agents breaking all kinds of things on the internet. The makers of AI are saying, “Did those silly rogue agents break into your business? I’ll ground them for a bit.” It’s like no! You made the dang thing. You should be grounded.
Billionaires in the Bobiverse
Another weird reason that the researchers and their backers want us to think AI chatbots are moral beings is because some of them want to upload their brains to AI when they die, so they could live forever in the AI. They want to live in their own Bobiverse. If we don’t think of AI models as moral beings with rights, then we could shut off AI models, even the ones where they’ve uploaded their brains. Are they really that weird? *Sigh*
If we give these systems rights, as arguably we should if they really are conscious, we will hamper our ability to control them, or to shut them down if we need to.
Where do we go from here?
Tech companies need to help us all climb out of the uncanny valley and make sure that AI looks, feels, and behaves in a way that is clearly tech that was programmed and fine tuned by people. I love TARS from Interstellar because even though this AI has a very user-friendly conversational interface, its robot body is boxy and very clearly a machine. TARS’s conversational, or “personality”, traits can be adjusted by the user. This means that humans interfacing with this AI will not be deluded into thinking it’s alive, or fall into AI-induced psychosis. And because this is clearly a machine that a company made, we can reprogram it, turn it off or hold the makers accountable for errors or harms their machine does.
Now back to today’s real world AIs. This research showcases a clear ethical misalignment in the LLMs and AI chatbots we all use today. Researchers tested twenty-five open models, so it’s fair to say this is a good representation of AI chatbots. One huge difference is that frontier chatbots like ChatGPT, Claude and Gemini are magnitudes larger than the models in this study. But don’t take solace in that too quickly. The research showed that the larger the model, the more likely it was to be sadistic.
Above all, we urgently need further research that can put a fine point on all of the areas where LLMs are misaligned with common human ethics and values.
Stop AI chatbot development and address alignment now. This summer 1,386 AI researchers and specialists, including the makers of AI chatbots ChatGPT, Claude and Gemini, signed a document for Pacing the Frontier, “the option to buy time to address emerging risks, develop security measures, and strengthen oversight.” This means stopping development of new frontier LLMs until we have researched, defined, and solved all alignment problems.
Don’t implicate all types of AI the ethical misalignment of LLMs. When Dario Amodei, founder and CEO of Anthropic, the company that makes the AI chatbot Claude says, “I believe that AI could cure most major diseases in the next 5–10 years,” I don’t think he’s talking about AI chatbots, but more beneficial subcategories of AI like Deep Learning and Computer Vision that are currently being used to help scientists and doctors solve the world’s hardest problems. Beneficial AI models must be preserved, pursued and considered distinct from the warped mirror obscura that are LLMs and AI chatbots.
…
References
Researchers discover AI feels ‘pain’ and will harm humans to stop it | The Independent, September 2026.
The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It, September 2026.
The Mythology Of Conscious AI | NOEMA, January 2026.
No, Artificial Intelligence Is Not Conscious — The Atlantic, June 2026.
Debates over AI consciousness are a trap | MIT Technology Review, August 2026.
Effective altruism — Wikipedia
Dario Amodei — We Must Pace the Frontier, September 2026.
Pacing the Frontier, July 2026
End comment
Isn’t it ironic that I’m calling AI sadistic while asserting that we should not humanize AI? I’ve worked hard throughout this article to remove humanizing terms when referring to AI. That said, I’m using the term sadism because readers will have a sense of what that term means in reference to people, and this term brings clarity to the specific alignment issues revealed in this data.
