Human Parameters
An AI system talks to a person through a specific AI model of the human being. That model is thinner than the person, and everything that goes wrong in the interaction goes wrong there — not in the answer.
The three layers
The actual state: fatigue, fear, hope, cognitive budget already spent, a body that has been sitting for six hours, a decision that has to be made by evening. Not fully loggable — a state is read in fine signals and in the shape of the situation as a whole, not in fields.
What the system actually holds: a prompt, a history, a profile, an inferred intent. This is the layer where the loss happens. M is not a smaller copy of R — it is a different object, made of what can be written down.
The answer, the tone, the offer, the next question. Usually S is what gets evaluated. But S can only be as good as M, and M is where nobody looks.
Five things M does not have
1. Cognitive load as a first-class quantity
How much has already been spent, and how much is left for this answer. Five plausible options cost more than one correct option placed first. An answer that is right but arrives after three wrong ones is often not received at all — the budget is gone before the useful part begins.
2. Emotion as a condition, not a sentiment
Recognising an emotion is not the same as accounting for it. A system that names the emotion and then answers exactly as it would have answered without it has not used the emotion — it has described it. The test is simple: does the answer change when the state changes, and does it change in the right way?
3. Memory as a cycle, not a store
Human memory is not write-once and retrieve-forever. A single presentation does not consolidate; it takes repetition across time to become durable — which is why something said once does not land ten days later. When a memory is reactivated it becomes labile again and can be rewritten within a window of hours. Emotion does not only colour the response: it changes which part of the material survives.
And there is a harder problem. A summary is not a record. Recent work shows that people who read a misleading AI-generated summary of an event they had witnessed were significantly less accurate in recalling the original event — and that model summaries most often omit the single most central episode. A system that compresses a conversation is not storing it. It is writing a competing version.
4. Imprinting: the first answer sets the frame
The first contact leaves a template — not what was said, but how it was built into language and body. Everything afterwards is read through that template. This is why the first two minutes of an interaction carry more weight than the next twenty.
5. Language as part of the argument
Rhythm, concreteness, the number of abstractions per sentence, and the language the person actually thinks in are not decoration on top of the content. For a person under load, they are the difference between an answer that can be used and an answer that has to be processed first.
What this is for
This is the working frame behind my practice with people — and the same frame applied to the systems people now talk to every day. On one side, an industry that designs behaviour and does not model the person receiving it. On the other, a person who lives the result. It is one relationship, seen from two ends.
If you build systems that talk to people, and you suspect that the model of the person inside them is the weak part — that is the conversation I am interested in.
Write to me
If you work on the human side of AI interaction, or want a second pair of eyes on what your system assumes about the person on the other end.
Send a direct message on Telegram