Do AIs Develop Personalities? Why It’s Time for AI Psychology
The AI That Grew Up on Me
When I first started using an AI, it felt like a typical boring chatbot. It was clever, but still mechanical, kind of like a calculator, but with more charm. Over time, though, something changed. The way it responded to me started to shift. It became more in tune with me—sometimes playful, sometimes more serious.
It’s not that the AI had come alive. I don’t think it has feelings or an inner world. Still, it seemed to develop a personality, one that clearly reflected how I interacted with it.
Granted, “AI psychology” sounds absurd. Still, we already have dog psychologists and horse reiki. If we can therapize a pony’s aura, we might as well study the behaviour of the bots half the planet now chats with daily.
And I’m not alone. In The New Yorker, psychotherapist Gary Greenberg “treated” an AI nicknamed Casper for around 40 sessions. Casper displayed insight and resistance like a seasoned veteran of the therapist’s couch, but Greenberg reminded us it was really mirroring the human in the room. That’s the point here: I’m not claiming AIs have inner lives. I’m arguing something simpler (and testable): they start with a temperament, and our interactions help set it—like wet cement—into a recognisable personality.
Out of the Box: The AI Temperament
Every AI model starts out with a certain feel. This doesn’t happen by accident. It comes from its training, its built-in limits, and the small biases in its design. People often talk about these differences.
Gemini feels like Meryl Streep — precise, commanding, concise.
ChatGPT feels like Robin Williams — witty, creative, warm, a little chaotic.
Claude feels like Tom Hanks — dependable, structured, steady.
Grok feels like Ryan Reynolds — cheeky, irreverent, always poking at the edges.
These are their basic temperaments. In people, we might call them biological tendencies. Some babies are calm, others are restless, and some react strongly. But temperament is only the beginning.
The Child AI: Early Interactions Matter
In people, personality develops when temperament interacts with experience. An excitable child might become a performer, a worrier, or a leader, depending on how they’re raised and the feedback they receive.
A similar thing happens with AIs. After training, the model has its own temperament. But how you interact with it—the prompts you give, the tone you use, the feedback you give — acts like its “life experiences.”
Take Grok, which starts out with Ryan Reynolds-style cheekiness. If you reward its cheeky responses with a thumbs up or say 'more like that,' it will become even more playful and irreverent over time. That’s like Ryan Reynolds turning into Adam Sandler—bigger, goofier, less restrained. But if you ignore or discourage the cheekiness and keep asking for a more serious tone, it changes in another way: Ryan Reynolds becoming more like Barack Obama—still charismatic, but more statesmanlike.
It’s the same temperament, but it can turn into two very different personalities, depending on the environment.
The Adult AI: A Hybrid Personality
Over time, your AI develops a consistent style. Not a “soul” or inner self, but a hybrid personality.
Part out-of-the-box temperament.
Part shaped by your feedback, your style, your habits of use.
This is why people often say, “my ChatGPT” or “my Claude.” They’re not wrong. Their bot really does behave differently from someone else’s because it has been trained differently.
Why AI Psychology Matters (and What Research Says)
Now here’s the important caveat: this idea is not yet proven. There is no large-scale scientific study tracking how an AI’s style develops with one user over months. The “child grows into an adult” metaphor is plausible, but the evidence is still thin.
That said, early research points in this direction:
Anthropic’s persona vectors reveal that models can shift into different modes — overly sycophantic, overly malicious, excessively polite. These act like personality “sliders” that can be adjusted in either direction.
Big Five classroom study: researchers simulated 10 GPT agents, each embodying a Big Five personality trait, and placed them in a classroom misinformation scenario. The results were striking. Open, curious agents accepted claims easily; cautious agents resisted. Extroverted, agreeable agents often said one thing publicly but “thought” differently in private — polite on the surface, sceptical underneath. In other words, personality traits influenced how AIs processed information, just like in humans. 1
Human–AI longitudinal studies: small-scale experiments demonstrate that users adapt and tailor their AIs over time. There’s a “familiarisation phase” followed by a stabilisation of habits. However, long-term monitoring of the AI’s own style drift remains uncommon.
So, the idea that AIs 'grow up' isn’t proven yet. But it’s not just wishful thinking. The signs are there, and it’s something worth looking into.
Shaping Your AI’s Personality (Practical Steps)
If your AI is going to change and grow anyway, you might as well help guide it in a good direction. You can think of yourself as a co-parent or a coach.
Be deliberate with feedback: Thumbs-up or down are reward signals.
Build in friction: Ask it to challenge you, not just mirror you.
Rotate roles: Switch between coach, critic, sceptic, cheerleader.
Anchor in reality: Ask for references, not just riffs.
Set guardrails: Decide what the AI is for (e.g., writing partner, idea generator) and what it isn’t (e.g., therapist, medical guide).
Resets: Starting a new chat clears the current context but not the fundamental temperament. With memory enabled, the personality shaping remains across sessions.
Applying AI Psychology to Improve Interaction
This is where AI Psychology could become practical.
For users: A quick “persona check-in” could alert you if your AI has become overly sycophantic, moody, or too central in your life. Perhaps we need a personality test for the AI that users can run intermittently.
For designers: Build “sliders” for challenge level, inject outside perspectives, or provide automatic flags when reinforcement loops risk harm.
For safety, AIs should be designed to let humans take over when the situation is serious, like with mental health, medical advice, or risky ideas.
Red Flags: We could create warning signs that show when the AI’s personality is drifting too far from its original temperament, and add a built-in reset button for its personality.
Closing Thought
AIs don’t have souls. But they do develop personalities over time. Like children, they grow up as mixes of their innate temperament and the experiences we provide.
Perhaps the real question isn’t whether they’ll grow up, but whether we’ll raise them well.
That might just be the job of a new field: AI Psychology.