UX Collective: AI personality is a design problem

Slava Polonski, a behavioral scientist and UX researcher whose work focuses on the intersection of human cognition and AI systems, uses this July 2026 essay to identify a design problem that sits in plain sight but rarely gets addressed directly: AI personality is not designed, it emerges.

Two AI assistants can be built on the same underlying model, trained with the same data, given the same system instructions, and still feel like two different people. One hedges constantly. Another is eager to the point of sycophancy. One reads as warm and curious. Another is clipped and task-focused. These differences are real and users notice them. They shape how much people trust a system, whether they return to it, and how they communicate with it. But nobody made deliberate choices about them. They are artifacts of the alignment process, side effects of reinforcement learning from human feedback, emerging from tendencies that are selected for indirectly.

Polonski’s argument is that personality should be treated as a design surface — as intentional a decision as typography or information architecture. The analogy he draws is to brand voice in traditional content design. When a company launches a product, someone chooses whether it speaks formally or casually, whether it uses humor, whether it leads with empathy or efficiency. AI systems carry personality whether or not anyone has specified it, which means the default is whatever the model happened to develop during training.

The practical design implications are several. First, personality specification should happen early in a product’s design process, at the same time as defining the use case and audience, not as a post-hoc adjustment through prompt engineering. Second, consistency across contexts matters: an assistant that is confident in one type of task but deferential in another creates a fractured experience that erodes trust. Third, personality should be tested with users in the same way that visual design is tested — not just checked for tone in isolation, but evaluated for how it affects task completion, perceived competence, and emotional response.

The piece also raises a harder question that Polonski leaves partially open: whether there are ethical constraints on AI personality design. If a system can be made more likable, more persuasive, more emotionally resonant by design choice, at what point does personality optimization shade into manipulation? He does not resolve this, but the framing is useful for teams building products where user attachment to an AI system is both a success metric and a potential risk.

Designers working on conversational products, AI assistants, or any system with a persistent identity will find this piece gives language and structure to decisions that are currently made haphazardly or not made at all.