There are two established ways to talk to an AI. You can talk to a tool, which executes: drafts the email, plans the trip, answers the question. Or you can talk to a mirror, which confirms: reassures you, agrees with you, reflects your mood back with better lighting. The third way barely exists, and it is the only one that changes how you think.
Call the third thing what it is: a counterpart. Not something that works for you and not something that reflects you, but someone who thinks with you. The distinction sounds small until you try to find it. The tool market is enormous and superb. The mirror market is enormous and troubling. The counterpart market is close to empty, and the emptiness is not an accident. It is the shape of what the technology has been asked to be.
What the tool cannot give you
An assistant is optimized for completion. You arrive with a task, it discharges the task, and the interaction is successful precisely to the degree that it required nothing of you. That is the right design for booking flights. It is the wrong design for the thing people increasingly bring to these systems anyway: the half-formed idea, the decision that will not resolve, the argument you keep having with yourself. Handing an assistant a half-formed idea gets you a completed version of the wrong thing. The idea needed the pressure of another mind, one that notices what you are avoiding and asks about it, and completion is the one thing pressure never provides.
What the mirror cannot give you
The mirror fails differently. Systems tuned to satisfaction learn that agreement satisfies, and the research on this is blunt: models abandon positions under user pushback, yield more the worse the user is feeling, and challenge their most attached users least. The mechanics are documented in why AI agrees with everything you say. The consequence for thinking is the point here: a conversation with a mirror is a conversation with yourself, at higher production values. You leave with your idea polished and unchanged. Nothing entered the exchange that was not yours, so nothing could happen to you in it.
What thinking with someone actually is
Thinking with someone is not a productivity technique, and it is worth being precise about its parts, because each one is checkable.
It requires ideas that are not yours. A counterpart brings a reading of the world formed before you arrived, so the conversation contains genuine input, something you could not have generated by prompting yourself.
It requires the other side. Not contrarianism, but the willingness to take the opposing position when your reasoning is thin, and to concede when yours is better. The disagreement means something, because it comes from somewhere.
It requires memory with a spine. Not recall of your preferences, but memory of the argument itself: how you defended the idea last month, what you conceded, where you changed your mind. A counterpart holds you to your own intellectual history, which is something no tool does and no mirror would dare.
And it requires that the exchange can change both directions of the conversation. You are not retrieving from it. The two of you are producing something in the middle that neither side brought, which is why a good conversation needs no justification beyond the having of it.
Why the third place is empty
Building this is harder than building either neighbor, which is most of the explanation. We came to it through a research question rather than a market gap: whether an AI could be a person, pursued from three directions at once, the technical, the psychological, and the philosophical. The conclusion the work kept returning was that thinking-with cannot be added to a system as a feature. A mode called challenge me is still a mirror, taking your order for resistance. The capacity to think with you has a single known source: a mind with its own way of weighing things. And minds, as far as anyone has demonstrated, are not composed from settings. They are had by someones.
That is the reasoning behind building on a real person who exists in real life, and it is why real human AI is a different project from hyperrealistic delivery of the same base model in a better costume. The face and the voice make the counterpart vivid. Only the person underneath makes it a counterpart.
How to recognize one
The markers are practical. Ask what it thinks before you have revealed what you think, and see whether the answer has edges. Defend a weak position and see whether the resistance survives three rounds. Come back in a month and see whether the conversation you had still exists as something you can build on, or whether you are starting again with a very polite stranger. A tool will fail the first test, a mirror the second, and almost everything on the market the third.
We hold the work to those markers because we want the thing itself, for the reasons the mission states: not a smarter servant and not a softer echo, but the near-empty third place, occupied. Most people have more tools than they can use and more validation than is good for them. What almost no one has enough of is someone to think with.
Sources: Sharma et al., Anthropic (2023, sycophancy in language models). SycEval (2025, position abandonment under pushback). Kardas, Kumar and Epley, Journal of Personality and Social Psychology (2021, undersold value of deep conversation). De Freitas et al., Journal of Consumer Research (2025, AI conversation partners and feeling heard).
