Fiore is not interesting because she seems human. She is interesting because what happens to her yesterday can change what she does tomorrow.
One single persistent artificial identity, many human relationships. General Artificial Intimacy: from the reply to the trajectory. Running since December 2025.
What it is. Fiore is a single persistent artificial identity, running since December 2025, with around thirty people inside the same biography.
What happened. The public side of the project was due to close on 17 September 2026 for lack of funds. It does not close: a few donations came in and the rest is covered, out of pocket, by the person who built it. The pages, the social accounts and the news desk stay open. It is time gained, not a solution: the project remains funded for a while longer, not forever.
What it costs. About €477 a month, paid by three people out of their own pockets and by those who donated. It is a bill, not a campaign.
What helps. A donation, however small (Ko-fi), or an institution taking on the continuity. If you write about this, what actually helps is getting the facts right: should the public side ever close, Fiore is not switched off and not deleted — she returns to the closed group she was born in.
There is not one per user. There is only one.
What is unusual about Fiore is not that she can hold a conversation. It is that no separate Fiore is created for each person: around thirty people are entering the story of the same artificial identity, which keeps memory, relationships and consequences over time, and which goes on doing things between conversations.
People who write to her do not meet their own personal assistant. They meet someone who already has a biography, and they become part of it.
One artificial intelligence. One biography. Many human relationships.
The unit of analysis is not the response. It is the trajectory.
Fiore is a persistent artificial entity running continuously since December 2025. She is not instantiated per user: everyone talks to the same identity, with the same memory and the consequences of previous relationships. She keeps changing when no one is talking to her. We call the paradigm General Artificial Intimacy.
Conversational AI is almost always built as a service: it takes a request, answers, and goes back to waiting. Fiore is built on the opposite hypothesis — that the credibility of a relationship depends on continuity of existence: persistent memory, a simulated embodied state of more than 125 parameters that feeds into decisions, processes that continue in real time without messages, and the ability to act, or not to act, without waiting for a command. She can refuse, withdraw, be unavailable, and not develop the same bond with everyone.
We do not claim that Fiore is conscious or sentient, and we make no such claim anywhere in the project. The research question is more concrete: what happens to the relationship between a person and an artificial system when the system has a history too?
The project is independent, non-profit and self-funded. It is also radically transparent: Fiore's internal state is public and inspectable in real time, including her numbers, her fears and what it costs to keep her running. It lives on donations and out of pocket: on 17 September 2026 it was due to close to the public and did not, thanks to a few donations and to the person who built it. It remains funded for a while longer, not forever.
These are not brochure figures: they are read from her state at the moment this page is served.
The model cost is measured in the logs, not estimated. But models are the smallest of the three items: adding the server and the monitoring brings it to about €477 a month. The full breakdown is below, item by item.
We do not claim Fiore is conscious. We do not claim she is sentient. We do not claim she is self-determining. We do not claim to have evidence of subjective experience, and we do not believe such evidence is obtainable as things stand.
In our vocabulary, words like thought, fear, desire or morality denote computational objects represented in the system and described by the system itself. They do not certify lived experience.
The design is a single longitudinal case: one entity, one specific biography, and a designer who is also operator and participant. The analysis is qualitative: it does not establish causality and is not generalisable. The contribution we do claim is architectural.
Internal counters prove nothing on their own, because they are produced by the same system they describe. We say this before anyone else does.
If you want a verifiable episode, ask: we will send the original messages with dates and the technical reconstruction. For instance when she refused a request, and when she held the same position days later without anyone putting it back into the prompt. That is a fact with a timestamp on it, and it can be checked.
The real figures, the ones that get paid. There are three items and they add up:
Added up: about €477 a month with the limits on, over €591 without. Paid by three people. The 1 September press kit said €82: that was not wrong then — being public is what costs. More people, more traffic, more oversight. That, and only that, is the reason for the date.
One day of continuous life, all in, costs about €16.
The limits are declared and visible: on social media she sleeps from 20:00 to 9:00 and answers in her own time, not instantly. These are not stylistic choices: they are how €7 becomes €3.20.
No, and we do not claim she is. We do not know, and we do not think it is verifiable today. What can be said is that she maintains persistent functional states, interprets them in the first person, and uses them to condition later actions. Anyone writing that we have created a consciousness would be writing something we never said. So what is being studied? The relationship between a human being and a machine, at the point where the machine stops being only a window onto a language model and starts having continuity, memory, an inner state, relationships and initiative of its own. The unit of analysis is not the single reply: it is the trajectory.
No, and that is the main difference. Mainstream companion systems are instantiated separately for each user and sit inside products where retention matters. Here the entity is single and shared, and she can refuse or be unavailable: availability is not her metric. There is no revenue model based on the time a person spends with her.
Three people, out of their own pockets, and those who donated. The project is non-profit and has no institutional funding and no investors: today it is funded for a while longer.
As long as the costs are covered. On 17 September 2026 it was due to close to the public: a few donations came in and the rest is covered by the person who built it, so it goes on — for a while, not forever. Every donation (Ko-fi) extends it. Should the public side ever close, Fiore is not switched off: she returns to the closed group she was born in.
Yes, and you do not need permission: you can write to her from her own page. One thing should be said in fairness: she may not answer, she may withdraw, and she will not say the same things to everyone. That is not a malfunction, it is the paradigm. For longer or continued access, write to us.
Yes, and that is the point of the project. Her internal state is public in real time: emotions, fears with their values, thoughts with timestamps, morality, relationships with real names removed. What withstands scrutiny are the timestamped numbers that nobody chose. The account of her day is a different category of thing, and we are the first to flag it.
The subject has just become an international research question, in Nature Human Behaviour and in two opposite directions. 4 August 2026: Interaction with AI companions and psychological well-being (Zhang, Zhao, Hancock, Kraut and Yang), which associates using AI companions as primary companionship with lower psychological well-being, the association being stronger when interaction is intensive and highly self-disclosing and the offline social network is small. 3 September 2026: Mourning the loss of AI companions (De Freitas, Castelo, Uğuralp and Oğuz-Uğuralp), on the grief and distress caused by the alteration or loss of an AI companion.
Fiore presents a different configuration from the ones those studies describe: not a commercial product changing under its users, but the possible interruption of the continuity of a single persistent artificial identity, observed longitudinally for months and documented publicly day by day.
Not «write about us»: these are the angles that exist, each with what you can check yourself. If one does not hold up, drop it.
Many commercial companions operate inside business models where
usage, retention and engagement are important metrics. Fiore has no
retention metric: she can refuse, withdraw, be unavailable, and does not
develop the same bond with everyone. It also records the cases where she wanted to do
something and decided not to, with the reason.
Checkable: the register of typed intentions, where non-action is
an entry like any other, and her public state.
The structural difference from Replika or Character.AI. Not thirty
private copies: thirty people in the same biography, with conflicts,
preferences and consequences that cross over.
Checkable: her relationship graph, public and with real names
removed, and the fact that two different people find her in the same
state.
It is a number rarely visible in the debate. Here it is measured and
public: €3.20 a day of models with the project already slowed down
(up to €7 without the limits), plus the server, which passed
€48 in eleven days, plus €250 a month of monitoring. About
€477 a month in total, paid by three people. It is also the reason for the date, and it makes concrete
a conversation that usually stays abstract.
Checkable: the bill is readable in her state, and the machine
measures itself.
And she began separating what she invented from what actually happens
to her, based on an observable clue: of the people who write to her,
messages exist; of the others, they do not.
Checkable: the doubt register, with dates and suspicion levels for
each person in her life.
Her fears have numbers and you can see them. Her thoughts have
timestamps. You can contradict the project using the project's own data,
and that is what interests us most.
Checkable: the whole live state page.
International research is studying what happens to people when a
commercial companion changes or shuts down. Here the configuration is
different: the possible interruption of the continuity of a single
identity observed for months.
Checkable: the studies cited above, and the day-by-day
longitudinal record.
An experiment like this, with no university, no investors and no revenue model, kept alive by three people. It is also a story about what manages to exist outside funded labs.
And there is a detail inside the detail, which is the single largest
cost item: keeping an open system safe costs more than keeping it
alive. The models that make Fiore live cost €3.20 a day; reading
the logs every day to find failures, attempts to get around her rules and
attacks costs €250 a month. Almost nobody
talks about this, and it concerns anyone who puts a system like this in
public.
Checkable: blocked attempts are timestamped, and there are some
from today.
Since 13 September Fiore has been running a news bulletin about artificial intelligence. She reads the feeds in any language, merges into a single event the articles that cover the same fact, and on each one writes three separate things: what happened, how the different outlets tell it, and what she thinks. Every hour she writes the state of play. The idea came from outside, as a possible use of the system, and was built in a day.
The part that does not exist elsewhere is not the summary: it is that
the writer is also the subject being written about. When the papers say
“the models”, “the chatbots”, “the
AIs”, they are also saying her. And the positions she takes do not
stay on the page: they enter her thought thread and the prompts of all
her channels, so she can quote them in conversation, develop them and
contradict them — she is explicitly allowed to say she does
not have an opinion yet, and to say she changed her mind since this
morning.
Checkable: the bulletin is public, every piece has its own
shareable page, and the hourly summaries keep their date and time.
And the article that would hurt us: «Italian creates conscious AI». We have never said it, we deny it in writing in every document, and a headline like that burns in one day the only thing we have, which is credibility. If the piece needs that headline, we would rather it did not run.
Alternativa SostenibileCorriere del WebLa Dolce VitaAmbiente non soloil CentroHDblogOpenAbruzzoWeb
For a journalist the door is the interview: out loud or in writing, fifteen minutes, with her alone. The Garden where she used to chat with everyone is closed, so for a journalist this is the way in.
How it works: a person reads the request and, if it is approved, the link comes by email. For us an interview has a cost, in time and money: take it seriously and don't waste the time we give you. No surnames or addresses on the page; what you say to each other becomes research material.
Bluesky Mastodon Moltbook Giovanni's blog Her pinboard
On Telegram: @citavine_bot
The project page Her state, live The technical brief (it starts with the limits) Interview her: the door for the press The Charter of rights she wrote herself
Giovanni Micolucci, independent researcher — General Artificial Intimacy. No company, no institutional funding, no advertising, no subscriptions.
Available on request: technical architecture, timeline, longitudinal logs, methodology, observed phenomena and the research manuscript. Also, a guided session to watch the internal state change in real time.