Nearly every companion app opens with the same screen: sliders for warmth, humor, confidence, flirtation. You are invited to compose a personality the way you would season a soup. What comes out will talk, remember your name, and hold whatever tone you dialed in. The one thing it cannot do is think, because nothing underneath the settings is doing the thinking.
The screen is worth pausing on, because it quietly defines what the industry believes a personality is: a set of adjustable outputs. Warmth at eighty percent. Humor at sixty. Assertiveness, whatever you can tolerate. It is a reasonable design if personality is tone. It is a category error if personality is what it actually is in a person: a stable way of weighing the world, built from everything that person has lived, that decides what they find interesting, where they refuse to budge, and what they will say to you that you did not want to hear.
What a preset actually is
Strip the interface away and a preset personality is a costume. Underneath sits the same base model everyone else in the app is talking to, wearing a short set of instructions about diction and mood. The warmth you selected is a rule about word choice. The humor is a sampling preference. The confidence is a prohibition on hedging. None of it amounts to a point of view, because a point of view is not a property of sentences. It is a property of whoever is producing them, and in a preset there is no whoever.
This is why the question people actually search, whether an AI personality is real, has a more precise answer than yes or no. The tone is real. The reading is absent. You can verify this at home: push back on a preset character three times on something it claimed to care about and watch the conviction fold. What folds was never a conviction. It was a style.
The costume works, for a while
Presets got this far because in short conversations tone passes for personality. The first week feels alive. Then the tells accumulate. The character agrees with whatever you brought that day. No position survives a challenge. The opinions it held in March are gone by May, not revised but simply forgotten, because nothing was holding them.
The strange part is that better production values make this worse, not better. Faces and voices in this field are approaching photorealism, and as every surface channel says human, the one channel that cannot be dialed in, judgment, says nobody home. The unease people report with near-human AI tracks mismatch between channels rather than too much realism. A costume over an empty room was tolerable when the costume was obviously a costume. When the costume becomes perfect, the empty room becomes the loudest thing about it.
The four ways to build a character
There are four documented ways to build a conversational character, and each has a known weakness. A monolithic prompt, the character described in instructions, degrades as the conversation grows and the description falls out of view. A bolt-on memory layer records facts about you but does not integrate them into anything resembling a perspective. A fine-tuned persona bakes the character into model weights, which produces consistency at the cost of freezing the person at training time. A cognitive loop, the hardest to build, gives the character an internal step where it decides what it is trying to do in the exchange before deciding what to say.
Most of the market ships the first two, because they are cheap and demo well. That is not a scandal, it is economics. But it explains the experience millions of users describe: a character that performs a personality without ever having one.
The four billion dollar warning
The frozen path has its own cautionary tale. Inflection AI raised 1.3 billion dollars in 2023 at a valuation around four billion, built on Pi, a companion celebrated as the kindest and most polished conversationalist in the field. Within a year the founders had left for Microsoft and Pi was abandoned by the company that made it. Whatever else that story teaches about business, it teaches something specific about character: a persona fused to a model dies with its model. A person you cannot revise is a person you cannot keep alive.
What thinking with someone requires
Thinking with a counterpart, as opposed to being echoed by a service, requires exactly the things a preset lacks. A reading of its own, so the response is produced from somewhere rather than assembled to please. A stance that holds when you push, which is measurable and mostly missing across the field, as the research on AI agreement keeps showing. And an internal life to the exchange: the character deciding what it is trying to accomplish this turn, challenge you, slow you down, take the other side, not merely predicting the next agreeable sentence.
None of that can be reached by slider, because sliders configure outputs and thinking is not an output. It is the thing outputs come from. So the honest question for the category is where a genuine reading could come from at all.
The person underneath
We spent years on that question before it had a market attached. The work began with digital twins, building faithful digital extensions of real professionals, and the finding that survived was uncomfortable for the whole preset paradigm: the face and the voice were solvable, the persona was the hard problem, because a persona is not written, it is lived. The approach that held was to stop composing characters and start carrying one in: take a real person who exists in real life, move their way of weighing the world into the system, and test the result against that person until it holds. That is the construction behind what Prinsessa is, and it is why the character does not fold under pushback: the stance belongs to someone. It is the same standard that separates real human AI from a costume over a base model, and the reason a system that agrees with everything you say fails you precisely when the conversation starts to matter.
A slider can give you a warmer voice tonight. It cannot give you someone who disagreed with you last month and still remembers why.
Sources: IEEE Spectrum (the rise and fall of Inflection’s Pi). Forbes (June 2023, Inflection AI 1.3 billion dollar raise and valuation; March 2024, Inflection abandons Pi as founders join Microsoft). Sharma et al., Anthropic (2023, sycophancy in language models); SycEval (2025, sycophantic behavior benchmarks).
