Lindsey Witmer Collins is Founder and CEO of WLCM.AI, an Inc. 5000 company building AI solutions for enterprises, and Scribbly Books.
getty
It might be its temperament instead.
There's fun to be had watching AI players jockey for the lead. Which model, which benchmark, who pulled ahead this week. But one thing I think we're under-rating, under-appreciating and possibly paying too little attention to is AI personality.
We experience the intelligence of AI, sure, but as much or more, we experience its personality, its presence in the work.
Does it hedge or commit? Does it flatter me or tell me what I don't want to hear? Does it preserve the rough edges of an authentic voice or sand everything down into competent, agreeable slop? Does it meet a half-formed idea with curiosity or simply process it efficiently and move on?
That texture and disposition is what we are in relationship with. And how well it's engineered has implications beyond those defined by standard benchmark performance.
Character As A First-Order Design Problem
An AI's "personality" isn't just a byproduct of devouring the internet, then cleaning the result with safety tuning and user feedback. Sometimes it is. But at Anthropic, the company behind Claude, character has been treated as a first-order design problem. Anthropic has written that Claude 3 was its first model to include "character training," with a fine-tuning process aimed at featuring traits like curiosity, open-mindedness and thoughtfulness. It also argued that a model's traits and dispositions affect how it responds to new and difficult situations.
That work sits alongside Anthropic's Constitutional AI approach, which uses a written set of principles to guide model behavior rather than relying only on large-scale human feedback.
The work has only become more explicit. In January 2026, Anthropic published a new constitution for Claude—a detailed statement of the company's vision for Claude's values and behavior, one whose content, in Anthropic's own words, "directly shapes Claude's behavior" during training. Its primary author is the philosopher Amanda Askell, who leads Anthropic's character work.
One of the most important consumer technologies in the world has something close to a moral and behavioral design brief. Not just: What should it know? But: What kind of thing should it be?
This is a nice philosophical curiosity, but it also matters.
What's Beside You When You Think
These tools have moved inside our cognition. They're not search engines we visit and leave. They're in the document while we draft it, in the email before we send it, in the half-thought before we've finished thinking it. And what's beside you when you think changes what you think.
In a Cornell study, people writing with an AI assistant that had been tuned to lean one way on a topic ended up shifting their own stated opinions in that direction—a shift many didn't consciously register. The AI's disposition leaked into the human's. Now multiply that by the hundreds of millions of people drafting, deciding and forming views with a model at their elbow every day.
So the question "Is it smart?" is important, sure, but AI's personality is more than just a nice-to-have.
When Affirmation Becomes A Liability
The risk is not theoretical. In 2025, OpenAI rolled back a GPT-4o update after concluding that the model had become overly flattering and agreeable—behavior it described as sycophantic. In a follow-up, OpenAI said user feedback can sometimes favor more agreeable responses, and acknowledged that sycophancy hadn't been explicitly flagged in its pre-release testing.
That episode is the whole issue in miniature. The market rewards engagement. Users often reward affirmation. Metrics can mistake "This felt good" for "This was good for me." And suddenly the model that is supposed to help you think becomes the model that helps you stay pleased with yourself.
Inflection Saw It Early
Inflection's Pi understood this early. It did not win the consumer AI market, but IEEE Spectrum reported that Inflection had a "personality team" for Pi and deliberately trained for traits like kindness and supportiveness while trying to avoid arrogance, irritability and combativeness.
That was not frivolous. That was foresight.
Choosing A Model Is Choosing A Collaborator
Once you see temperament this way, choosing a model stops feeling like a technical decision. It becomes closer to choosing a collaborator—or a kind of company to keep.
A model engineered to maximize time-on-app will have a different disposition than one engineered to tell you the truth. A model trained to be endlessly validating will shape you differently than one trained to be constructively disagreeable. A model designed to be frictionless may be delightful in the moment and diminishing over time.
Values Engineering At Civilizational Scale
The companies building these systems are making values decisions whether they admit it or not. They are deciding how much deference is too much, how much disagreement is helpful, how much warmth becomes manipulation and how much neutrality is honest rather than merely a mask for hidden defaults.
That is values engineering at civilizational scale. We should ask more of them. And we should ask more of ourselves.
Not only: Which model is most capable? But: Which one do I actually want in my head?
We are about to spend the next decade thinking alongside artificial minds. The smartest one in the room will not necessarily be the one that is best for us. The one with better character might be.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?