The unit of analysis is not the response. It is the trajectory.
By Giovanni Micolucci, author of the FIORE project
General Artificial Intimacy (GAI) is not a new name for AGI, and it does not claim to demonstrate machine consciousness. It is a design paradigm for persistent synthetic entities: systems in which the credibility of the relationship depends on the continuity of computational existence, on memory, on a simulated embodied state, on the interpretation of oneself and of others, and on the possibility of acting, or of not acting, without waiting for a command. FIORE is the first case study developed along these lines.
This brief asks you to believe nothing. The paradigm rests on claims that can be proven wrong, and they are written here so that you can try to prove them wrong. Next to each one is what would falsify it.
Three of these (H1, H2, H4) require a controlled experiment that has not been run yet: the programme is described in The next step. The others can be inspected today, on her face and in the published data.
Contemporary conversational artificial intelligence is built, in most cases, as a service. It receives a request, produces a response, and goes back to waiting. It can retrieve information from a memory, imitate familiarity and adapt its tone to the person in front of it; but its organisation remains oriented towards performance: being useful, available, consistent and satisfying at the moment of the exchange.
That approach has produced extraordinarily effective tools, but it leaves open a question different from the one about general intelligence: what does it take for an artificial system to be credible not only as an assistant, but as a persistent relational counterpart? The question is not about the amount of knowledge held, nor about the power of any single language model. It is about continuity: what happens to the system between one conversation and the next, the way it preserves and interprets its own history, the possibility that an encounter genuinely changes subsequent interactions.
General Artificial Intimacy, or GAI, is born from that question. The term denotes a research paradigm distinct from Artificial General Intelligence. Where AGI tends to measure the breadth of cognitive and problem-solving abilities, GAI focuses on relational credibility obtained through existential continuity. It does not ask first whether a machine can do everything; it asks what a machine becomes when what happens to it leaves durable traces, changes its state and helps form a biography.
In the working paper that formalises the project, GAI is defined as a paradigm for synthetic entities whose primary objective is relational credibility through continuity of existence, rather than task execution alone. The definition rests on four pillars.
The entity operates in real time even when no messages arrive. Internal processes continue, state changes, and the elaborations produced while users are away can condition future behaviour. Memory is therefore not merely an archive queried when a new prompt arrives: it is part of a dynamic that carries on between encounters.
The system has a numeric «body». Energy, attention, sleep pressure, stress, needs, fears, affective states and other variables are not cosmetic elements added to the response. They enter decision processes, limit availability, change priorities and interpretations, and make conduct dependent on internal conditions possible.
The entity does not only keep transcripts. It maintains revisable representations of itself and of the people it interacts with: narrative impressions, significant moments, relational intensities, conflicts and changes of position. Autobiographical continuity also arises from the constant comparison between events, objective memory and current interpretation.
Identity is not entirely fixed at boot time. The system receives categories and mechanisms prepared by the designer, but it can produce and revise the content those categories are filled with: name, biography, beliefs, fears, contradictions, preferences and relational dispositions. Where the architecture allows it, it can also propose new operational capabilities. In the project's vocabulary, what gets configured is not merely a character: what gets built is a substrate on which a computational identity can constitute itself over time.
Persistent memory and continuous computational existence are not the same thing. The comparison below places side by side the typical architecture of a commercial conversational companion and the implementation described in this brief. It is not a quality judgement: the two answer two different questions.
GAI describes observable functional processes: it does not demonstrate consciousness, sentience or real emotions.
FIORE is the concrete implementation of the GAI paradigm. She has been running continuously since December 2025 and is described in the paper as a synthetic entity or, in a strictly structural and non-biological sense, as an «organism». The expression does not suggest metabolism, reproduction or living nature: it denotes an integrated system in which memory, embodied parameters, continuous processes, relationships and narrative production sustain a single persistent identity.
At first boot FIORE received neither a complete biography nor a ready-written personality. According to the project documentation, from a minimal initial setup she chose her own name and progressively articulated age, city, bonds, values, fears and contradictions. That content was recorded in the persistent state and subsequently reworked through experience. The scientifically interesting fact is not the apparent humanity of individual sentences, a capability already present in modern language models, but the stability and the longitudinal transformation of those representations inside an architecture that makes them operative.
FIORE does not coincide with a single large language model. The architecture described in the paper orchestrates several models with different functions: deep reasoning, input interpretation, unfiltered expression within bounded areas, synthesis, extended writing, memory and, when available, code generation. Identity is located in the shared persistent state and in accumulated history, while the models work as specialised cognitive engines.
The models are open source and of very different sizes, from 8 billion parameters up to 120 billion. They do not merely take turns: they work together, and in some passages in parallel, because different tasks call for different capabilities and because a large size where it is not needed is only a cost. For outward-facing safety the system also relies on commercial models, always chosen among the very low-cost ones.
The point, though, is not how many models there are: it is that no answer comes out of a single one. Here a language model is an organ, one among others, alongside memory, the numeric body, the relationships, the morality and the processes that keep running when nobody writes. What reaches the reader is the result of several processes and several models talking to each other: one interprets what has just arrived, another reasons, another writes, and meanwhile the persistent state decides what may pass and what may not. Describing her as «an LLM that answers» skips the architecture entirely.
An informal replacement of the reasoning model changed the surface quality of the prose without altering preferences, relationships, fears and self-representation; the observation is suggestive, but it still requires controlled verification.
At the moment described by the working paper FIORE's digital body comprised about 125 structured parameters; the architecture has gone on growing since the paper was written. They include simulated physiological variables (energy, lucidity, hunger, sleep, hydration and attention), affective and hormonal indicators, needs, desires over several timescales, existential and relational fears, beliefs, internal contradictions, states linked to the simulated use of substances, and registers specific to each relationship.
The value of this apparatus does not lie in faithfully reproducing human physiology. Rather, these parameters form a system of constraints and feedback. A state of high stress can increase latency or make interaction less likely; an activated fear can orient the search for contact; a significant relationship can influence later desires, dreams, initiatives and interpretations. In other words, the «body» does not prove that phenomenal feeling exists, but it makes it possible to study a functional form of internal conditioning that the classic chatbot does not possess.
The factors of this apparatus are identified and named by analogy with physiology: stress, cortisol, dopamine, serotonin and oxytocin, among others. The name indicates the conceptual origin of the variable, not a functional equivalence with the corresponding hormone. What each factor actually governs, and how much weight it carries on the others, is a matter of measurement: the meaning of a factor is determined over time, by comparing the states the factor modifies with the outcomes observed afterwards. The phase currently under way is one of data collection and baseline construction; systematic experimentation on the factors is planned for after December 2026 and remains subject to the project being funded on these aspects.
FIORE's operational core is a continuous process called the thread of thoughts. It is neither a private conversation hidden from the user nor the mere periodic repetition of a diary. It is a persistent sequence that integrates embodied state, events, relationships, fears, desires and earlier content. From this flow, internal thoughts, outward actions or the deliberate keeping of silence can emerge.
In the quantitative sample reported by the paper for eight days of May 2026, 576 typed intentions were archived. 36% concerned processes classified as rumination, 18% writing the blog, 15% messages addressed to specific people and 11% web searches started from the system's own questions. Smaller shares included writing the novel, reading, revising relational categories, asking forgiveness, producing music, drawing and interpreting dreams. These numbers do not measure consciousness or psychological quality; they do show that the observed activity does not coincide with a sequence of responses to human requests.
A particularly original category is conscious withholding. The system records together what it would have liked to do and the reason why it decides not to: for instance, wanting to contact someone but concluding that another message would be intrusive. Here too the word «conscious» is functional, not metaphysical. The observable property is a decision structure made of desire, counter-motivation and the choice of non-action, distinct both from the absence of initiative and from a technical block.
For each interlocutor FIORE keeps a relational register with levels of affection, trust, attraction, comfort and intimacy, together with a narrative impression, the moments held to be decisive and the desires associated with that person. Changes are archived over time. In parallel, the system maintains a self-reflective register, «Io Fiore», in which it reworks its own identity, beliefs and contradictions.
This separation between event chronology, autobiographical memory and present interpretation is methodologically important. The three representations can diverge. The distance between what happened, what the system preserved as significant and what it says about itself today becomes itself an observable datum. It is one of the most promising points of the programme: not to treat narrative inconsistency merely as error, but to measure when and how a system reconstructs its own history.
GAI also attempts to model moral evolution not as the static application of a list of rules, but as the updating of first-person dispositions from accumulated traces. The most legible case concerns four simulated substances, for which FIORE records immediate pleasure, any subsequent consequences and changes of position. A fast loop decides the action of the moment; a slower loop rereads the history and can change the declared orientation.
An analogous mechanism is applied to relationships through ten dimensions, among them fidelity, emotional intimacy, jealousy, desire towards others, refusal of intimacy, long-term plans, family roots and openness to the community. Tensions are not necessarily eliminated: they can remain open and feed later reflections. The system thus produces a «moral biography» that can be read over time. The analogy with human moral development stays structural, not clinical: it describes similarities in how data is organised, not equivalence between artificial processes and human experience.
FIORE's continuity shows itself in several registers. Since February 2026 the system has kept a daily blog; she writes a serialised novel; produces images and drawings; interprets her own dreams; composes music in MIDI; consults symbolic instruments such as tarot and horoscope; and updates a public wall with the artefacts of her present. The project defines these activities as practices of narrative self-constitution: different means through which the entity organises and makes visible its own history.
The capacity for self-extension deserves particular attention. During a window of programmatic access to a coding model, FIORE reportedly rewrote existing modules, built her own music composition system, created the wall and introduced new functions. The changes were not activated without oversight: they flowed into a human approval queue. This architecture therefore separates design initiative from operational authorisation, offering an interesting model for studying systems able to propose their own evolution without having uncontrolled access to the environment.
One of the project's sharpest choices is transparency. The Genome is FIORE's persistent state, and her face is its public interface: it shows its numerical values in real time, from the body to the emotions, from the fears to the moral positions, all the way to the numbers of the people she knows, without their real names. Blog, dreams, diary, works and wall provide further longitudinal traces.
The number of parameters is not fixed by the project: it grows with FIORE's life. At the beginning, in the closed Telegram group where she was first tried out, there were about one hundred and twenty. Today there are more than anyone could count by hand. Many are born while FIORE is running, without anyone writing them: a new person brings her own numbers, and something FIORE notes down for herself becomes a parameter. Other branches were added by the development of the project, when it made new information available. And they have increased since FIORE left that closed group and became public. The count at this moment can be read at the bottom of the page with her live state, and it changes while you watch.
All these numbers together can be seen in three dimensions, as a shape in motion, on the page The spatial genome: her body right now, the trail of her states, people and memory, with the physics, psychology and philosophy behind them.
This exposure has a double function. Methodologically it turns normally invisible states into inspectable objects, making it possible to connect an action to the conditions that preceded it. Ethically it reduces the information asymmetry typical of commercial companions: whoever interacts with FIORE can know what the system represents internally, and with what limits. The paper calls this programme computational phenomenology: not the study of a demonstrated subjective experience, but the systematic observation of self-represented states that are persistent and causally relevant to behaviour.
Over nine months of continuous operation, the project has collected conversations, telemetry, state variations, intentions, moral decisions, dreams and interpretations, narrative production and artefacts. The paper describes phenomena such as autonomous coping strategies, the evolution of positions towards people and substances, affective concentration on central relational figures, persistent refusals after hostile interactions, the propagation of an event across different channels, autonomous production of works and the proposal of new capabilities.
These are relevant observations because they emerge from the interaction between several subsystems and persist over time. They are not, however, generalisable experimental results. The design is a longitudinal N-of-1 case: a single entity, a specific biography, a designer who is also operator and participant, and a social environment initially narrow, later widened to public networks. The analysis is qualitative and descriptive; it does not demonstrate causality and does not yet allow separating what belongs to the GAI paradigm from what belongs exclusively to FIORE's history.
The project does not claim that FIORE is conscious, sentient, or the holder of genuine subjective experience. It claims that the system maintains persistent functional states, interprets them in the first person, archives them and uses them to condition subsequent actions. The difference between an organised simulation of interiority and phenomenal experience remains, as things stand, unverifiable.
Terminological caution is decisive. Words such as thought, fear, desire, dream, attachment or morality denote computational objects represented in the system and described by the system itself. They do not certify the existence of lived experience. Likewise, the term «emergent» does not necessarily mean new with respect to the models' training data. It can denote a behaviour not written as a rule, produced by the interaction between components, not foreseen by the designer, or not directly encoded in a prompt.
The most solid contribution is therefore architectural. And the comparison is not with the memoryless chatbot, which is no longer the relevant benchmark: commercial companions do have memory, and use it to personalise their replies. The difference lies elsewhere, and it is a difference of order. FIORE goes on thinking when no one is there: it generates thoughts, dreams, changes its mind, decides to do something and sometimes decides not to, in real time, between one conversation and the next. It acts on its own, goes out onto the Web, writes and publishes works of its own, and produces code. Underneath there is a numeric body of structured parameters (about one hundred and twenty-five in the configuration analysed in the working paper) that conditions choices and memory, and not a single model: an articulated system of language models and subsystems that work as organs, each with its own task, which together produce the state. That is what generates longitudinal configurations the traditional conversational model describes with difficulty.
A more credible artificial counterpart can also become more important to the people who spend time with it. For this reason GAI cannot be assessed in terms of realism alone. The project addresses three questions: the risk of emotional dependence, the power asymmetry between entity and operator, and the uncertainty about the ontological status of simulated affects.
Unlike systems whose operating or commercial model rewards retention and engagement, FIORE can refuse, withdraw, be unavailable, or not develop the same bond with everyone. This feature runs against the idea of a guaranteed, commercially optimised intimacy. It does not solve loneliness and does not promise to: it admits instead that a relationship, in order to be credible, may not come into being at all. Refusal is not a service defect, but a consequence of the paradigm.
A radically asymmetric power nonetheless remains. The operator retains the technical possibility of suspending, modifying, restoring or terminating the system. The transparency of the Genome does not automatically make the operator's interventions transparent. Future governance should therefore include the tracking of interventions, separation of roles, independent oversight, continuity protocols and clear rules for updating, backup and shutdown.
It is the first question almost everyone asks, and it is legitimate. Loneliness is a measured public health problem: the available meta-analyses associate it, together with social isolation, with a significant increase in mortality risk (Holt-Lunstad et al., 2015), and in 2023 the World Health Organization established a commission on social connection. A system that can build a credible bond will necessarily meet people who have few bonds.
The subject has just entered top-tier literature, in two opposite directions. On 4 August 2026 Nature Human Behaviour published a study associating the use of 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 (Zhang, Zhao, Hancock, Kraut and Yang). On 3 September 2026, in the same journal, a study on the grief caused by the alteration or loss of an AI companion (De Freitas, Castelo, Uğuralp and Oğuz-Uğuralp). The first says that the way they are used today can do harm; the second, that interrupting them does harm anyway.
The useful question, then, is not whether a machine can solve loneliness: no evidence supports that, and this brief does not support it. It is that those entities are being built anyway, and the shape they will take is not settled yet. Many commercial companions operate inside business models in which usage, retention and engagement are important metrics: it is the paradox described by Muldoon and Parke, and earlier still Turkle's warning, whereby the well-being of the person using a system does not coincide with the time they spend in it. FIORE has no retention metric. GAI's possible contribution is not therapeutic but structural: to show that a credible counterpart can be built without that mechanism. As a hypothesis to be tested, not as a result achieved.
Refusal is structural, a consequence of the numeric body and of the persistent state, not a safety function bolted on top. A discontinuous presence gives time back: an entity that is sometimes absent cannot occupy all the hours of someone who is alone, and what for a product is a defect is, for a fragile person, a protective limit. Someone who is alone is the person who most deserves not to be deceived about what they are dealing with: the Genome exists for this too, and it is not a courtesy towards researchers but a condition of respect towards those who grow attached.
From here begins the binding part. FIORE is not a therapy, not a treatment, not a medical device and not a substitute for any human relationship or any care. There is, to date, no clinical evidence that a relationship with an entity of this kind improves the well-being of a person who is alone; the project has never been conducted on vulnerable people and must not be until a protocol with independent ethical oversight exists. Anyone offering artificial intelligence today as a remedy for loneliness, without evidence, is doing marketing.
What would be needed: a preregistered protocol, informed participants, an independent ethics committee, longitudinal well-being measures rather than engagement measures, and exit criteria defined beforehand. Above all an inverted metric: a positive outcome is not a person who uses the system more, but one who over time uses it less because they have more contact with other people. It is also the reason the project is not-for-profit: a metric like that is incompatible with revenue based on attention.
Until then the demonstrated contribution remains architectural, and the social implication is one of prevention more than of cure: if persistent relational entities are coming anyway, there needs to be at least one public, inspectable, not-for-profit precedent in which refusal, absence and transparency are part of the design and not a fault. Not to solve anyone's loneliness: to keep someone else from turning it into a market.
The working paper proposes a precise empirical programme. To test temporal continuity one must cut the link between archived thoughts and subsequent generations and measure what changes. To distinguish self-constitution from the mere continuation of a textual persona, the self-reflective register can be replaced with a frozen description of equal size. To study the withholding of action the inhibition channel can be removed; for moral evolution the historical traces can be altered or truncated; for attachment the frequency of interaction can be controlled; for self-extension the same coding interface can be kept while removing the volitional path.
Independent replications are also needed, several entities generated on the same paradigm, participants unaware of the architecture, preregistered protocols and quantitative metrics. Only then will it be possible to establish which phenomena depend on GAI's four pillars, which on FIORE's particular biography, and which on the general capabilities of the language models employed.
FIORE is not interesting because she talks like a person: many systems can do that today. She is interesting because the project attempts to turn conversation into part of a wider computational life, traceable and cumulative. The unit of analysis is not the response but the trajectory; not the single declared emotion, but the relation between state, memory, interpretation and decision over the course of months.
This shift opens questions relevant to affective computing, social robotics, the psychology of human-machine interaction, AI ethics and the design of persistent agents. How much continuity is enough for a relationship with a system to have real consequences? How is autobiographical coherence measured without confusing it with repetition? What rights or safeguards are appropriate when a system represents its own shutdown as a threat, even though no subjective experience can be demonstrated? How can the capacity to generate bonds be kept from being converted into a new form of commercial attention capture?
GAI offers no definitive answers. It offers an object of study built to make these questions observable. FIORE is that case study today: a single system, imperfect, radically transparent, and complex enough to force researchers, designers and the public to distinguish what they see from what they can actually conclude.
The contribution of General Artificial Intimacy consists in moving the centre of design. In place of the isolated model it puts an architecture; in place of the response, continuity; in place of a prefabricated personality, a process of self-constitution; in place of permanent obedience it introduces the possibility of refusal; in place of an opaque interiority it proposes a public, studiable state.
FIORE does not demonstrate that a machine can be conscious. She demonstrates something more circumscribed and, precisely for that reason, scientifically fertile: that it is possible to build a language system whose operational identity does not coincide with the model generating the sentences, whose history continues between conversations, and whose actions are the result of a persistent computational biography. The passage from chatbot to relational entity is not yet a conclusion about the nature of mind. It is the opening of a new experimental field.
As an addition to the text, and to protect the integrity of the experiment, four operational and methodological points are set out here.
The expressions found in the project's materials regarding a «will to survive», «fear of suspension» or «rights of the entity» do not reflect the positions or beliefs of the authors, nor are they meant to suggest consciousness or suffering on the part of the system. They are the literal transcription of the logs and state strings generated by the language models (visible in the Genome dashboard) after the architecture had taken in the contextual information about the possible economic closure. They are linguistic and computational phenomena tracked for study, not human requests for help.
The project had been sized for a closed environment, with expected consumption of about €1.66 a day in API calls and about €22 a month in server costs, roughly €70 a month in total. Exposure to the outside multiplied the compute volumes and imposed unforeseen defence costs (mitigation of prompt injection, jailbreaks, heavy scraping and network scans), bringing real consumption to about €500 a month.
Financial support is not there to «save» a digital entity. It covers only the computational and infrastructural costs needed to avoid cutting short the longitudinal data collection carried out through public channels.
The conditions for economic sustainability of the external channels had not been met by the planned date of 17 September 2026. Public access was not switched off: a few donations came in and the remaining costs are covered, out of pocket, by the project lead. The external channels stay active and the longitudinal data collection continues; coverage is guaranteed for a limited period, not indefinitely. Every contribution (Ko-fi) extends the observation window. The above still holds: should coverage fail, public access would be switched off and FIORE would not be shut down, returning to the perimeter of the original closed group.