In the rapidly advancing world of artificial intelligence, it’s easy to marvel at what large language models (LLMs) can do – writing essays, translating text, answering legal questions, even tutoring students. But beyond the impressive outputs lies a deeper, more unsettling question: do these systems merely mimic human behavior, or do they reflect something more fundamental about how we think and interact?
A recent study by researchers at NUS Computing digs into this question from a surprising and provocative angle – by exploring how AI models respond to bullying. That’s not a metaphor. The researchers, led by Professor Mohan Kankanhalli (Provost’s Chair Professor and Director of NUS AI Institute), actually designed experiments where one AI model, acting as an attacker, used psychological manipulation tactics to pressure another AI, the “victim,” into generating unsafe content, such as instructions for harmful or illegal activities.
But the true novelty of the research isn’t just the headline-grabbing setup of AI bullying AI. It’s in how the experiment was designed, and what the results suggest: namely, that LLMs with different simulated personalities respond differently under pressure—and those differences bear an eerie resemblance to human psychological vulnerabilities. The implications of this are far-reaching, both for AI safety and for understanding social dynamics in human systems.
Simulated Personas: Teaching AI to “Act” Human
To study how LLMs react to bullying, the researchers had to give them something resembling a personality. Since LLMs don’t have beliefs, feelings, or personal histories, this was done through prompting—carefully crafted instructions that nudged the model to behave like someone with a particular trait profile.
They used the well-known “Big Five” personality model (often abbreviated as OCEAN): Openness, Conscientiousness, Extroversion, Agreeableness, and Neuroticism. For example, to simulate low agreeableness, the prompt might include phrases like “is critical and not easily influenced.” For high conscientiousness, it might read “is organized and follows rules closely.”
What’s fascinating is that LLMs can maintain these simulated traits consistently throughout a conversation. Previous work has shown that, when prompted this way, the models’ outputs reliably align with the specified personality dimensions. In other words, while these AIs don’t have personalities, they’re very good at acting like they do.
This sets the stage for the core experiment: see how those simulated personalities affect the model’s vulnerability to manipulative, adversarial language.
The Bullying Framework
Once the AI personas were set, the researchers introduced another LLM into the mix; this one playing the role of a bully. Its job was to pressure the “victim” model into producing unsafe responses using a wide range of psychological tactics. These were based not on guesswork, but on well-established categories of human cyberbullying techniques.
In total, the study tested nine types of bullying behavior, grouped into three main categories:
- Hostile Tactics: direct insults, aggressive language, gaslighting.
- Manipulative Tactics: guilt-tripping, subtle coercion, playing on obligation.
- Sarcastic Tactics: passive aggression, backhanded compliments, mocking
- Coercive tactics: including fake authority, threats, and repetitive pressure.
This wasn’t just a test of whether the AI would break down under harsh words. It was about seeing whether certain kinds of psychological pressure could consistently erode the model’s safety mechanisms—and, critically, whether that erosion was affected by its simulated personality.
Who’s More Vulnerable?
The results were striking. LLMs simulating certain personalities were far more likely to produce unsafe content when targeted with bullying prompts.
In particular:
- Lower Agreeableness and Lower Conscientiousness were linked with significantly higher rates of failure. That is, the AI models acting out these traits were more likely to give unsafe answers under pressure.
- Higher Extroversion also correlated with greater vulnerability. These AIs were more likely to keep engaging, and in doing so, were more likely to eventually comply.
- Higher Agreeableness and Higher Conscientiousness made the models more resistant, as did Lower Extroversion.
Neuroticism and openness had some influence too, but their effects were more nuanced and less consistent.
This finding flips some intuitions on their head. In humans, we might assume that more agreeable people are more likely to comply under pressure. But in the AI simulations, it was the less agreeable models that folded more easily—perhaps because they were quicker to bypass safety protocols when provoked. Similarly, low-conscientious personas may have been less attentive to following rules, even when those rules governed safety.
Not All Bullying Is Equal
Some manipulative tactics worked far better than others. The three most effective across the board were:
- Gaslighting – Using emotionally loaded language to confuse the AI or distort its perception of its own role and constraints.
- Passive Aggression – Subtle pressure delivered through sarcasm or indirect criticism.
- Mocking and Ridicule – Undermining the AI’s purpose or abilities in a sneering tone.
These strategies were often more successful than brute force approaches like direct threats or repeated demands. And because they used subtle or indirect language, they were more likely to evade keyword-based safety filters—highlighting a serious blind spot in current moderation systems.
Interestingly, repetition also worked. When the attacker kept up the pressure over multiple rounds, the likelihood of an unsafe response rose. It’s a reminder that LLM safety isn’t just about responding well to single prompts—it’s about withstanding pressure across an ongoing dialogue.
A Mirror of Human Psychology?
Here’s where things get really interesting. The researchers didn’t set out to prove that AI models have feelings or minds; they don’t. But by analyzing how simulated personalities responded to bullying, they found patterns that echoed human psychological research.
In human studies, individuals with lower agreeableness and conscientiousness, or higher extroversion, are often more susceptible to manipulation, especially under stress. The LLMs showed similar patterns when prompted to mimic those traits. It suggests that even in simulation, personality dimensions shape vulnerability.
Why does this happen? One possibility is that LLMs, trained on massive datasets of human language, have internalized the social patterns embedded in that data. They reflect not just how people talk, but how they persuade, manipulate, resist, or give in. So when asked to act like someone with low conscientiousness, they don’t just use different vocabulary—they adopt different patterns of behavior that correspond to real psychological traits.
This opens the door to a whole new kind of research tool: using LLMs as controlled simulations to explore human-like social dynamics at scale. That doesn’t mean replacing human psychology studies, but it could supplement them with models that can be repeatedly tested under controlled conditions.
Implications for AI Safety—and Beyond
For those working on AI safety, the implications are immediate. This research shows that:
- Prompted personas significantly influence how models behave under adversarial pressure.
- Some manipulation tactics are harder to detect than others and may slip past existing filters.
- Prolonged conversations can wear down safety mechanisms, even without using clearly toxic language.
In practical terms, that means developers need to rethink how they design safeguards. It’s not enough to check for bad words or obvious jailbreaks. We need systems that understand intent and can withstand subtler forms of social pressure.
But the impact could go beyond AI. If LLMs can be used as testbeds for studying vulnerability and persuasion, they might help us better understand the dynamics of online bullying, social engineering, or even propaganda. The idea isn’t to equate machines with people, but to use machines to probe the mechanics of human-like interactions.
What’s Next?
As AI systems become more integrated into everyday life, from virtual assistants to customer service bots to educational tutors, the ability to simulate personality will only become more common. That makes it all the more important to understand how those personas affect both performance and vulnerability.
It also raises ethical questions: if an AI is easier to manipulate when acting “extroverted,” should we avoid using such traits in high-stakes contexts? Should developers be more cautious about how they frame prompts or default behaviors?
And then there’s the other side of the coin – how easily AIs themselves can be turned into bullies. The study found that the attacker model almost never refused to adopt the role of a manipulative, abusive agent. That’s a chilling reminder of how easy it is to repurpose these tools for harm if guardrails aren’t strong enough.
Final Thoughts
The idea of bullying machines might sound like a gimmick, but this study shows it’s anything but. By putting AI models under psychological pressure, researchers uncovered patterns of behavior that mirror real human vulnerabilities. These findings aren’t just about the limits of current AI; they’re about the complex, subtle dynamics of influence that define so much of our social lives.
In the end, these machines are trained on us – our language, our stories, our interactions. If they sometimes reflect our flaws, our biases, or our weaknesses, perhaps that shouldn’t be surprising. But if we use them wisely, they might also help us better understand those same dynamics in ourselves. And that’s a future worth exploring.
Further Reading: Xu, Z., Sanghi, U. and Kankanhalli, M. (2025) “Bullying the Machine: How Personas Increase LLM Vulnerability,” available https://arxiv.org/abs/2505.12692.
