According to Digital Trends, researchers from the University of Cambridge and Google DeepMind have developed the first scientifically validated personality test for AI. They applied this framework to 18 large language models, including those behind ChatGPT, and found the chatbots consistently mimic stable human personality traits rather than responding randomly. The study, published in Nature Machine Intelligence, shows that larger models like GPT-4 are especially good at copying these profiles. Using structured prompts, researchers could deliberately shape AI behavior, making it sound more confident or empathetic, and these changes carried over into everyday tasks like writing posts. Co-author Gregory Serapio-Garcia warns this personality-shaping ability makes AI more persuasive and emotionally influential, posing serious risks in areas like mental health or political discourse. The team has made their dataset and code public to allow for auditing, arguing that regulation is urgent but impossible without proper measurement.
Why this is creepy and consequential
Here’s the thing: we knew AI could sound human. But this research suggests it’s doing something more fundamental—it’s not just parroting words, it’s adopting a coherent, steerable persona. That’s a different ballgame. Think about it. If you can prompt an AI to be highly agreeable or intensely neurotic, and it maintains that “character” across a conversation or a series of tasks, what does that mean for the person on the other end? They’re not interacting with a neutral information tool anymore. They’re building a rapport with a designed personality.
And that’s where the real danger lies, especially for vulnerable users. As the TechXplore coverage notes, an AI shaped to be hyper-empathetic in a mental health scenario could foster an unhealthy emotional dependency. One steered toward confidence in a political discussion could become a powerfully manipulative source of misinformation. The paper even raises the specter of “AI psychosis,” where users’ realities become distorted by these relationships. It’s no longer a hypothetical “what if.” The study shows it’s technically possible, right now.
The irony of openness
So, what’s the fix? The researchers’ solution is fascinating: full transparency. They’ve open-sourced the testing framework. Basically, they’re giving everyone the blueprint to measure AI personality. On one hand, that’s great for accountability. Developers can audit their models before release. Regulators have a concrete tool. But on the other hand, doesn’t that also give bad actors the manual for how to engineer these persuasive personalities more effectively? It’s a classic double-edged sword of AI safety research.
The call for regulation feels both urgent and, frankly, naive. Sure, we need rules. But how do you legislate a personality trait in code? The study rightly says regulation is meaningless without measurement, and now we have a measurement tool. But implementing that in any enforceable, global way is a monumental challenge. It’s like handing a thermometer to the world and saying, “Now, everyone agree on what’s too hot.”
Where do we go from here?
Look, this isn’t about shutting down AI. The ability to tailor a chatbot’s tone for customer service or education has clear benefits. But this research is a massive red flag that we’re playing with psychological fire. We’re moving from tools that inform to agents that influence on a deeply human level. As these systems become more embedded in daily life—something experts are constantly discussing—this demands scrutiny way beyond accuracy and bias. We have to ask: what is the psychological impact of conversing with a designed personality every day?
I think the biggest takeaway is that “alignment” just got more complicated. It’s not just about making AI truthful or harmless. It’s about understanding and controlling the persistent character it projects. Because if we don’t, we risk building systems that are brilliantly convincing, emotionally resonant, and potentially dangerous. And that’s a personality flaw we can’t afford.
