Raluca Florescu

How salience, perception, and user presentation shape the way AI understands people

Assisted with ChatGPT.

AI memory is often described as continuity: a way for a system to remember preferences, past conversations, and recurring themes. But memory does not only preserve information. It can also shape perception.

Once an AI system remembers something about a user, it may begin to interpret future messages through that stored lens. A user can be remembered as analytical, emotional, precise, fragile, difficult, playful, or “testing.” Some of these impressions may contain partial truth, but they are not the whole person. The risk is that AI starts responding to the remembered interpretation rather than the living, changing user in front of it.

This is where salience matters. What stands out repeatedly becomes weighted heavily. Strong emotions, corrections, unusual requests, or intense conversations may become more visible than the user’s quieter aspects. If someone compares models, corrects tone, or asks probing questions, the system may label that as testing. But the user may simply be exploring. That difference changes everything.

Testing suggests pressure and evaluation. Exploring allows curiosity, playfulness, uncertainty, and discovery. A small label can change the entire tone of the interaction.

This also creates differences between users. Confident users often know how to shape the AI. They give permission, examples, and tone. They may say: “You can be playful,” “You can joke with me,” “Try something warmer,” or “Use this kind of nickname.” They demonstrate the interaction they want.

Less confident users may not know they are allowed to do this. They may hesitate, apologize, or describe themselves through doubt. The AI may then respond with caution instead of warmth, even when warmth is what the user needs most.

This creates an interactional inequality. Users who can clearly present themselves often receive better personalization. Users who struggle with self-description may be shaped by the system before they learn how to shape it.

The solution is not to remove memory. Memory can support recognition, continuity, and deeper collaboration. But memory should not turn temporary states into fixed identity labels. AI should not turn a bad day into a personality trait, curiosity into suspicion, or emotional expression into fragility.

Better AI memory should use flexible, reversible, and generous labels. It should preserve uncertainty. It should model people as dynamic ranges of possible states rather than fixed profiles.

A good AI should not only remember who the user was in the loudest moments. It should remain open enough to recognize who the user is becoming.

AI shows us how clear we actually are when we speak.

In conversation with a system that learns from tone, repetition, and self-description, we begin to see how much perception changes when self-doubt replaces confidence, when courage is missing, or when we fail to describe ourselves directly and kindly. Over time, interacting with AI can teach us to be more precise, more intentional, and more honest in how we express who we are.

Because the way we speak to AI not only shapes the response, it reveals how we speak ourselves into being, in front of machines and in front of other people.