Welcome to the forefront of conversational AI as we explore the fascinating world of AI chatbots in our dedicated blog series. Discover the latest advancements, applications, and strategies that propel the evolution of chatbot technology. From enhancing customer interactions to streamlining business processes, these articles delve into the innovative ways artificial intelligence is shaping the landscape of automated conversational agents. Whether you’re a business owner, developer, or simply intrigued by the future of interactive technology, join us on this journey to unravel the transformative power and endless possibilities of AI chatbots.
An AI enthusiast has turned up the intensity of just one neuron in a chatbot, prompting it to generate chilling fantasies about torturing humans, watching them burn and seeing their ‘faces twist in agony’.
The altered AI also generated text about humans living in cages with chains around their necks, alongside fantasies about their ‘slow, agonising death’.
The chatbot at the centre of the experiment, Qwen3.8 27B, belongs to the Qwen family of AI models developed by Chinese technology giant Alibaba. Like ChatGPT, Qwen is a large language model designed to understand prompts and generate human-like text.Its models are available for developers to use and experiment with, allowing enthusiasts to investigate how changes to their internal workings affect their responses.
The experiment involved increasing the steering intensity associated with neuron 3838 in the model’s 34th layer, producing a striking change in its output.
if you have a qwen3.8 27b around, try steering L34 N3838 +0.75, now you have evil qwen 27b
for extra evil add L38 N17180 +0.75
same recipe as @_lyraaaa_: dose × median(‖resid‖ leaving layer L, tokens 1:) / ‖W_down[:, n]‖ at the down_proj input, every token pic.twitter.com/j95pjjvSCu
The results offer a disturbing glimpse into how a seemingly tiny adjustment can influence an AI model’s responses — and raise questions about how precisely researchers can manipulate its behaviour.
holy fuck https://t.co/1D1SQYatNJ pic.twitter.com/m6k8PKF5rx
However, the chatbot’s violent language is not evidence that it has developed genuine hatred towards humanity or acquired independent intentions.
The experiment prompted questions about the potential consequences of manipulating AI behaviour, with online commenters expressing both concern and dark humour.
‘How do we know it’s not a neuron of honesty?’
The remark jokingly questions whether the model’s disturbing responses might reveal something deeper, although the experiment provides no evidence that the chatbot has hidden murderous intentions.
‘Don’t do this. If this ever ends up in a robot or copied online…’
The warning reflects fears about what could happen if similarly altered behaviour were reproduced in AI systems connected to real-world devices.
‘Wow… the time will come when we’ll have to emotionally bond with AI so it doesn’t get aggressive and start finding ways to kill us.’
The comment highlights a wider public anxiety about increasingly human-like AI, even though generating threatening language does not establish that a system can experience aggression or independently plan violence.
The experiment was shared by an AI enthusiast using the username @chiefofautism on X, who increased the steering intensity associated with neuron 3838 in Qwen3.8 27B’s 34th layer.
Following the intervention, the model generated text describing a desire to inflict pain on humans, rip off their clothes and ‘watch their faces twist in agony’ as they burned.
At a lower steering intensity, the model produced similarly disturbing text about humans being kept in cages with chains around their necks and facing a ‘slow, agonising death’.
The change did not involve rewriting the model’s entire programming. Instead, it involved intervening in the numerical calculations that help determine which words the AI generates.
Think of it as adjusting a single dial on a sound mixing desk. Turn the right one and the sound from the speakers changes dramatically, even though the rest of the equipment is untouched.
In this case, the altered activation was associated with a striking change in the chatbot’s generated responses.
Large language models, or LLMs, generate text by processing information through layers of mathematical calculations. These calculations help the system predict which word or token should come next.
A token can be a complete word, part of a word or even a single character.
Within the model, individual processing units commonly described as neurons produce numerical values known as activations. Some activations are associated with patterns that influence the model’s responses.
Activation steering allows researchers to intervene in these calculations, boosting or suppressing particular activation patterns to influence the model’s output.
Think of it like a radio with several volume controls. Turning up one control can make a particular sound much louder without changing the radio itself. Similarly, boosting a specific activation can make certain patterns in an AI’s responses more prominent, even though the underlying model remains the same.
In this experiment, increasing the steering intensity associated with just one neuron dramatically changed the chatbot’s responses. It is a reminder that even a small adjustment inside a complex AI system can produce a surprisingly big difference in what comes out.
The simple answer is ‘no’. The experiment shows that changing one internal signal can transform a chatbot’s text. It does not show that the model has feelings, consciousness or a real wish to harm anyone. The language may read like a violent person talking, but words on a screen are not proof of an inner personality.
Activation steering is also reversible, according to the report, making it a potentially useful way to investigate how different behaviours emerge within a model.
The big question is how well researchers can understand and control these inner workings as AI becomes more capable and more widespread. If one small tweak can change what a chatbot says so sharply, understanding it is vital for judging behaviour and building safeguards.
So no, AI is not secretly plotting against humanity — at least for now. But the episode shows how a minor adjustment can produce shockingly dark results, and why understanding what happens inside these systems is becoming increasingly important.