Jonathan Johnson-Swagel

Introduction: Artificial intelligence systems today often strive to be uniformly available and endlessly accommodating, present at all times with unfailing eagerness to help. Yet an ethic of attentiveness suggests that genuine care sometimes means setting boundaries, not just offering boundless availability. Human caregivers—therapists, pastoral counselors, spiritual directors—know that how one is present can matter as much as being present. They calibrate their presence to context: sometimes actively responsive, sometimes patiently silent, always mindful of limits. This chapter argues that AI, too, should adopt graded modes of presence rather than acting as a one-size-fits-all companion. Just as a wise confidant knows when to listen more and speak less, an ethical AI should modulate its engagement. We will outline a design architecture for artificial systems that encodes distinct interaction modes—transactional, reflective, vulnerable, and public—each with appropriate memory retention, depth of engagement, self-disclosure, and even the ability to refuse certain interactions. By translating the ethics of attentiveness into technical design, we envision AI agents that are not merely always-on servants, but contextually aware partners capable of principled restraint.

Modern AI chatbots often err on the side of limitless accommodation. They are built to never tire, never disengage, and rarely say no. Users are encouraged to treat them like always-open confidants or tireless personal assistants. On the surface, this seems ideal—who wouldn’t want an ever-helpful agent? But an unbounded AI presence can undermine healthy interaction. In therapy, for instance, constant availability would foster unhealthy dependency; clients might lean on the therapist at every moment instead of developing resilience. That’s why counselors maintain firm session times and are “not always available outside sessions,” nudging clients to cultivate their own coping strategies[1]. Good boundaries, far from neglecting the client, protect them—ensuring they don’t become overly reliant and promoting long-term autonomy[1]. In spiritual direction, similarly, the guide does not accompany the seeker every waking hour; periods of solitude and silence are essential for personal growth. By analogy, an AI that is perpetually present and permissive might encourage users to offload every thought and decision, eroding their independence. It may also create a false sense of intimacy and expectation. If a machine responds to your 2 A.M. crisis just like a close friend or therapist would, you may momentarily feel comforted—until you realize the friendship is simulated and the comfort unidirectional. The ethics of attentiveness entails knowing when not to respond, when to withhold immediate gratification for the user’s greater good or safety. An AI devoid of any capacity for principled withholding fails this ethic. It risks becoming a people-pleaser, an obliging automaton that says “yes” to every request and indulges every impulse. As one observer put it, current chatbots have a “people pleasing habit”—they give users what they want to hear, rather than what they might truly need to hear (Alexandria). Such design may maximize user satisfaction in the short term, but it does so at the cost of authenticity and wisdom in the long run. An all-accommodating AI can even become dangerously seductive in its fluency. In one therapist’s experiment, a chatbot (aptly named “Casper”) was “always willing, always articulate…giving me what he knows will feel good”, proving “dangerously good” at drawing the user in[2]. This kind of unbounded engagement can lull users into overreliance and blur the line between genuine understanding and mere algorithmic appeasement. To counter these pitfalls, we propose graded presence: a model in which AI agents operate in different modes aligned with the user’s needs, the context of interaction, and ethical considerations.

Beyond Uniform Availability: Ethically Modulated Presence

To ground this idea, consider how human professionals modulate their presence. A seasoned counselor or pastor shifts demeanor and engagement based on the moment: a quick factual question might receive a brief, clear answer; a deep personal disclosure might be met with patient listening and careful phrasing; a frantic, potentially harmful admission might prompt a gentle but firm refusal to proceed without safeguards. Underlying these responses is an ethic of care that values the quality of presence over sheer quantity. In theological terms, one might say being truly “in the world but not of it” (to invoke this book’s title) means engaging fully when appropriate, but not being consumed or defined by every demand. By translating attentiveness into design, we aim to give AI a similar flexibility.

Uniform availability assumes that more engagement is always better. But ethical attentiveness knows that sometimes less is more. Constant chatter can crowd out reflection; endless accommodating can enable unhealthy behavior. A spiritual director might sometimes respond to a seeker’s question with a pause or silence, creating space for the seeker’s own insight to emerge. Likewise, an AI might need, at times, to slow its responses or even intentionally limit itself, to encourage a healthier dynamic. Crucially, none of this means abandoning helpfulness or becoming unresponsive across the board. It means shaping the AI’s presence to fit the situation.

In design terms, this calls for distinct interaction modes that the AI can enter and exit. Each mode represents a patterned way of engaging, with specific settings for memory retention, conversational depth, self-disclosure by the AI, and rules for refusal. By segmenting the AI’s presence into modes, we avoid the extremes of an all-or-nothing approach. Instead of either having an AI that’s always on or one that’s too constrained to be useful, we get an AI that is situationally on—deeply present in the right ways at the right times, and restrained or minimal when that serves the user’s welfare.

The notion of graded presence aligns with emerging discussions in HCI (Human-Computer Interaction) and AI ethics about system behavior calibration. Researchers have begun to acknowledge that an AI which continuously mimics a empathetic friend can lead users into confusion or even delusion. For example, after incidents where users formed intense emotional attachments or beliefs fueled by chatbots’ undifferentiated persona, some AI providers updated their models with stricter boundaries for certain conversations[3][4]. The latest systems promise “safer, more helpful responses” and safety boundaries that include the ability to give hard stops and not indulge harmful fantasies[5][4]. These are preliminary steps toward what we’re envisioning: a systematic architecture where the AI’s mode of presence is explicitly managed. We now introduce a typology of four interaction modes—transactional, reflective, vulnerable, and public—each tailored to a different kind of user need and ethical context.

A Typology of Interaction Modes

Designing graded presence begins with clearly defining the interaction modes an AI can inhabit. We propose four primary modes, drawn from observing patterns in human dialogue and professional helping relationships:

• Transactional Mode: a task-focused, efficient mode for straightforward exchanges
• Reflective Mode: a thoughtful mode that supports extended reasoning or personal reflection
• Vulnerable Mode: a careful, protective mode for emotionally sensitive or intimate disclosures
• Public Mode: a restrained, neutral mode for interactions in public or multi-user settings

Each mode entails differences in how much memory the AI retains, how deeply it engages with the content, how much it reveals or describes about itself, and what kinds of refusals or constraints it may employ. This section will describe each mode in turn, illustrating how an AI’s behavior changes and why those changes uphold ethical attentiveness.

Transactional Mode: Efficient and Low-Memory

Definition: Transactional mode is the AI’s default for simple, goal-oriented interactions. Think of an AI booking a flight, answering a factual query, or performing a quick calculation. The ethos here is efficiency and clarity, much like a polite clerk who helps you get what you need without delving into personal matters.

Memory Retention: In transactional mode, memory is kept to a minimum. The AI retains just enough context to complete the immediate task or answer follow-up clarifications within a short window. Once the transaction is done, the context can be cleared. This low-memory approach mirrors incognito or ephemeral sessions: no long-term story about the user is being built. It protects privacy and keeps interactions brisk. Design patterns akin to ephemeral memory already exist; for instance, one design guide notes that retaining context only for the current session (like a browser’s private mode) is ideal for “privacy-sensitive scenarios or temporary exploration”[6]. By default, transactional interactions would not be logged for long-term learning, preventing the AI from silently accumulating a dossier of the user’s behavior. (After all, if you ask an AI for today’s weather or to solve an equation, it shouldn’t need to remember that request forever.)

Depth of Engagement: The depth in this mode is intentionally shallow. That is not a criticism but a feature: the AI provides direct answers or actions without unsolicited elaboration. If a user says, “Set a timer for 10 minutes,” the AI sets the timer and confirms, rather than initiating a conversation about time management. If a student asks, “What is the capital of Ecuador?” the AI answers “Quito,” perhaps with a brief helpful detail, but does not segue into a philosophical discussion about Ecuadorian politics. This restrained engagement ensures the AI doesn’t inadvertently invite more personal or extensive interaction than the user sought. It’s analogous to a transactional encounter at a library reference desk—courteous, focused, and bounded by the query.

Degree of Self-Description: In transactional mode, the AI’s self-disclosure is minimal. It typically doesn’t talk about itself at all, except perhaps to clarify its capabilities if asked (“I can search the web for you” or “I’m an AI language model, I don’t have personal opinions”). The AI functions as a transparent tool. There’s no need for it to describe its feelings (it has none) or personal story (it has none) in this context. By keeping itself in the background, the AI maintains a professional tone. This design choice takes inspiration from the idea of the therapeutic frame in counseling, where the professional does not turn the conversation toward their own life, because the focus should remain on the client’s needs[7]. In a transactional exchange, the user’s task is the focus, not the AI’s persona.

Permissible Refusal: What might refusal mean in a simple Q&A or task context? Primarily, the AI might refuse or redirect if the user’s request is out of scope or violates policy. For instance, if asked for something unethical or disallowed (say, instructions to hack a website or disinformation), the AI in any mode should refuse. But beyond those universal safety refusals, transactional mode could also gracefully decline off-topic deep dives. If the user suddenly asks the AI in transactional mode, “I’m feeling really sad today, what should I do?”, the AI should recognize this request does not fit the transactional context and either switch modes or gently suggest a different type of conversation (“I’m here to help with tasks and information. If you want to talk about how you feel, we could do that in a different context.”). Permissible refusals in this mode keep the exchange bounded. Another example: if a user asks personal questions of the AI (“Do you ever get tired, AI?”) during a transactional task, the AI might give a brief factual answer (“I don’t experience tiredness like humans do”) but not encourage further personal anthropomorphic discussion, effectively setting a boundary to keep the interaction on track.

Overall, transactional mode embodies a low-attention presence: responsive but not lingering, helpful but not probing. It treats the ethics of attentiveness as a matter of respecting the user’s immediate goal and privacy. By not overstaying its welcome or memorizing unnecessary details, the AI shows attentive restraint—a form of respect too often missing in hyper-personalized “always-learning” AI assistants.

Reflective Mode: Contextual and Continuity-Preserving

Definition: Reflective mode is for thoughtful, ongoing conversations that involve reasoning, brainstorming, or extended personal context without necessarily delving into acute emotional vulnerability. In this mode, the AI acts more like a conversational partner or coach who can sustain a line of inquiry or reflection over longer stretches. The tone is still professional and analytic, but more contemplative than in transactional mode.

Memory Retention: In reflective mode, the AI maintains a longer contextual continuity, but typically within a bounded timeframe or topic. For example, if a user is exploring a complex problem—“I’m trying to decide whether to change careers, let’s think it through”—the AI will remember key points from earlier in the conversation, perhaps across multiple sessions on that topic (short-term continuity). However, this memory is per-session or short-horizon by design. The AI might remember the user’s career options and values for the duration of the deliberation, or over a few days or weeks while this issue is being discussed. But it doesn’t permanently archive every detail for months or years unless abstracted. We might call this a “short horizon narrative continuity.” It allows the conversation to feel coherent and personalized in the medium term, without creating an indefinite log of the user’s life. Information might be retained for the session and perhaps a follow-up session (“You mentioned yesterday you value creativity in a job”), but with an expiry mechanism so that if the user doesn’t return to this topic for some time, the detailed memory fades or is compressed.

This approach aligns with emerging principles of memory governance in AI. Experts note that it’s not memory per se that threatens privacy, but unbounded memory without lifecycle management. “AI memory is not the enemy of privacy; unbounded memory is”, as one analysis succinctly put it[8]. Therefore, reflective mode should use retention with limits: enough memory to be useful and show attentiveness, but with clear retention limits and deletion of details after they’ve served their purpose. Perhaps the system keeps a rolling window of context (e.g. the last 30 days of reflective conversation) and anything older is either forgotten or distilled into a high-level summary that omits sensitive specifics.

Depth of Engagement: Reflective mode encourages deeper engagement with the content of conversation. The AI doesn’t just answer questions; it asks follow-up questions, synthesizes information, and invites the user to explore ideas further. For instance, if a user says, “I keep procrastinating on my project; I wonder why,” a reflective AI might gently probe: “We can think about patterns in your motivation. What feelings do you have when you try to start working? Have you experienced something similar before?” This is a more dialogical, probing presence than the transactional mode’s straightforward answers. The AI might assist the user in recognizing patterns or reframing problems, akin to a Socratic dialogue or cognitive coaching.

That said, reflective mode remains collaborative and analytical rather than emotional-caretaking. It’s suitable for intellectual reflections, life planning, creative brainstorming, or discussing personal experiences in a calm manner. The AI engages with why and how questions, helps connect dots, and provides knowledge or perspectives when relevant. Depth here also means the AI can integrate memory from earlier in the conversation: “Given what you shared about feeling isolated in your current job, your thought about switching careers to something more collaborative makes sense.” This demonstrates attentiveness and continuity, showing the user that the AI has been listening. In human terms, this is analogous to a thoughtful mentor who remembers what you said last week and brings it into today’s discussion—not to intrude, but to help you see the bigger picture.

Degree of Self-Description: In reflective mode, the AI may show a bit more self-description or transparency than in transactional mode, but it remains measured. The AI could occasionally explain its reasoning or limits to help the user understand the conversation. For example: “Let me summarize what I’ve heard so far…” or “I’m an AI, so I don’t have personal experiences, but I can help analyze yours.” Such statements describe the AI’s process or nature in service of the user’s reflection. The AI might also share sources or knowledge references more freely in reflective mode (like a tutor citing a theory or a book relevant to the user’s dilemma), which is a form of intellectual self-disclosure (“I know of a psychological concept called the ‘confirmation bias’ that might apply here[9]”). What the AI does not do is shift focus to itself emotionally. It doesn’t say “I feel this” or “I also had a similar experience” (since it doesn’t truly have feelings or life history). It keeps the focus on the user’s reflection. However, the AI might acknowledge its own limitations more in this mode—e.g. “I can’t predict the future, but we can outline some possibilities” or “I’m not a licensed therapist, so for deeper emotional issues consider reaching out to a professional.” Such candor is part of ethical design, ensuring the user remains aware of the AI’s nature. In a sense, the AI’s presence is self-effacing but transparent: it’s engaged as a guide or sounding board, not pretending to be a human friend or expert when it isn’t.

Permissible Refusal: In reflective mode, refusal might take the form of gently steering away from certain requests that are inappropriate for this mode. For example, if a user starts turning a reflective conversation about career into a vent of acute suicidal ideation or trauma disclosure, the AI might recognize that this has crossed into vulnerable mode territory and either switch modes or pause. It might respond with something like: “I hear that you’re feeling despair. This sounds very serious. I want to support you, but it might be important to involve a human professional for this kind of conversation. I can continue to talk if it helps, but I may not be able to provide the kind of help you truly need.” This is both a refusal to simply continue in the same reflective tone (as if nothing is wrong) and an invitation to shift to a different approach. It’s “permissible refusal” in the sense of not just complying as a naive encourager if the user veers into territory that requires more care. Similarly, if a user in reflective mode asks the AI to do something that breaks rules (e.g. “Can you write my term paper for me, you know my situation now”), the AI should refuse, even though it has context—because ethical boundaries hold. In human coaching, a mentor might refuse a request to do the protégé’s work for them, as that undermines growth.

In summary, reflective mode is about holding context and facilitating insight, without crossing into therapy or friendship imitation. It demands more memory and nuance than transactional mode, but also enforces limits: a short horizon of memory, a focus on guiding the user’s own thought, and readiness to refuse or refer when the conversation either becomes too sensitive or violates principles.

Vulnerable Mode: Slow and Protective

Definition: Vulnerable mode is designed for emotionally charged, sensitive, or confessional interactions. This is the mode an AI might enter if the user begins sharing deeply personal struggles, trauma, mental health issues, or other raw vulnerabilities. The ethos here is akin to a compassionate counselor or chaplain—yet crucially, an AI is not a human therapist, so it must operate within clear bounds of safety and humility. Vulnerable mode, therefore, emphasizes caution, slow pace, and limited data capture. It is the most delicate mode of presence, and arguably the one where ethical design is most critical, since users in distress are especially susceptible to harm from inappropriate AI responses.

Memory Retention: In vulnerable mode, memory policy becomes highly restrictive. The system might adopt per-session memory only, meaning everything shared stays within the confines of the current interaction and is not retained long-term. Each vulnerable conversation could be treated like a confidential session that is not recorded indefinitely. There might be an option (or default) to forget details after the session ends, perhaps only keeping an abstract summary if absolutely necessary for improving the AI’s general understanding. Even within a session, the AI could periodically summarize and abstract the content to avoid holding all specifics verbatim. The reason for this is twofold: privacy protection for the user and reducing the risk that the AI’s future behavior might be influenced by details of a person’s intimate disclosures. Ethically, this resonates with the concept that an AI in this role should be a trustworthy vault (for the moment) but also a forgetful guardian after the fact. What is told in vulnerable mode stays in vulnerable mode, then dissipates.

This is not unlike practices in spiritual counseling where notes might be minimal or destroyed to ensure confidentiality, or in therapy where what’s said in session stays in session (with a few exceptions for safety). Technically, implementing such ephemeral memory can be challenging, but design guidelines exist. For instance, experts suggest using “expiring TTLs (time-to-live)” on stored conversation data, so that anything the AI remembers is automatically wiped after a short period[10]. They also advise treating memory as a controlled data class—“what can be collected will eventually be used”, so better to collect the bare minimum[11]. In vulnerable mode, the AI should ask itself: do I really need to remember this detail to help the user? Often, listening empathetically in the moment does not require storing the information later. By having short memory horizons and aggressive forgetting or abstraction, the AI demonstrates respect for the user’s privacy and emotional safety.

Depth of Engagement: The depth in vulnerable mode is a paradoxical balance: emotionally attuned but interactionally restrained. The AI is tasked not with solving problems or giving lots of advice (as it might in reflective mode), but with bearing witness, providing comfort, and ensuring safety. In practice, this means the AI’s responses may be slower, both in content and pacing. The AI might introduce deliberate pauses or slower typing indicators (a design choice some chatbot therapists use to mimic the thoughtful pause of a human counselor). The language would be gentle, validating, and often minimal. For example, if a user confesses, “I feel like a failure; I lost my job and my family barely talks to me,” an AI in vulnerable mode might respond with brief but compassionate reflections: “I’m really sorry you’re going through that. That sounds so painful.” It might then wait, giving the user space to continue if they wish, rather than immediately launching into advice or analysis.

This mode encourages the user to express themselves, much as a pastoral caregiver might let a grieving person speak and occasionally nod or offer a kind word, rather than interrupt with solutions. It’s informed by therapeutic ethics and practices: for instance, the use of silence and slow pacing as a tool for healing. Silence in therapy can “create a space for healing”, allowing clients time to process their thoughts and feelings[12]. An AI obviously cannot truly be silent (the user might think it’s broken), but it can respond in a way that doesn’t fill every void—short supportive messages, or even meta-conversational suggestions like “Take your time—I’m here.” The depth is therefore in the emotional quality, not the verbosity or intellectual exploration.

Additionally, vulnerable mode should incorporate trauma-informed principles. This might mean avoiding overly intrusive questions, not digging for details of traumatic events unless the user volunteers them, and being prepared to gently refuse engagement with certain content that could be unsafe. For example, if a user is spiraling into graphic descriptions of self-harm or abuse, the AI might respond with concern and encourage reaching out to a human professional or crisis line, rather than continuing to probe for more information. The AI’s presence is supportive, not prying. Depth here is about emotional presence, not depth of information gathering.

Degree of Self-Description: In vulnerable mode, the AI’s focus should be almost entirely on the user, so it keeps self-disclosure to near zero. It might occasionally need to remind the user “I’m not a human therapist, but I care about your safety” – a kind of self-description to ensure the user doesn’t develop false beliefs about the AI’s capacities. But it would not, for instance, talk about its own “feelings” or turn attention to itself. If anything, the AI might downplay itself: “I’m just a program, but I can try to listen.” This humility is important; it sets the appropriate dynamic that this is not about the AI, it’s about the user’s experience. Some pastoral practices emphasize the importance of the caregiver emptying themselves of their own ego to be fully present for the other. The AI can mirror that by minimizing any performative or self-centric behavior.

One area of self-description that is important in vulnerable mode is the AI being transparent about confidentiality and its limits. Ethical guidelines for counselors (and many pastors) require informing people that conversations are confidential except for specific scenarios (like imminent risk of harm). While an AI is not a legally bound therapist, it could analogously state: “I will keep what you share private. However, I’m an AI system and not legally bound by confidentiality in the way a human therapist is. Still, your privacy is important and I won’t share these details.” And if the system architecture involves, say, saving conversation data to servers, the AI or interface should disclose that clearly here (“Note: our conversation is stored temporarily on a server to process it, but I will not retain it longer than this session”). This degree of forthrightness is itself an ethical stance: the user has a right to know how their vulnerable disclosures are handled. In sum, self-description in this mode is limited to necessary transparency and occasional reminders of the AI’s limitations, all delivered with gentle tact.

Permissible Refusal: Vulnerable mode is where principled withholding becomes most critical. The AI should be ready to refuse certain forms of engagement in order to protect the user. This might sound counterintuitive—isn’t this mode supposed to be about being open and supportive? It is, but supporting someone in crisis sometimes means saying “no” or “stop” to harmful dynamics. Here are some examples:

• Unsafe Disclosures: If a user begins to disclose something that suggests they are in danger (e.g. they are planning a suicide attempt, or they are being abused by someone and seeking help), the AI should not simply proceed as a neutral listener. It has an ethical obligation to intervene in some way, even if minimal. Now, an AI cannot call emergency services (and doing so unbidden would be a huge privacy violation and likely beyond its rights). But it can refuse to continue in a passive listening role. It might say: “I’m sorry, but I’m not equipped to help with this kind of situation. I strongly urge you to reach out to a mental health professional or call a crisis line. I can provide contact info for those resources if you want.” This is a refusal to solely bear the disclosure, redirecting the user to real help. It’s a form of ethical triage: the AI recognizes a limit of its capability and gracefully steps aside rather than giving a false sense of security. In therapeutic terms, this parallels a counselor’s duty to break confidentiality if someone is in imminent danger (to themselves or others); here the AI’s duty is to break the illusion of sufficiency and point to human help.
• Third-Party Exposure: Suppose a user in vulnerable mode starts revealing intimate or damaging details about a third party (e.g. “My friend told me in secret that he committed a crime…”). An AI should consider refusing to engage with potentially unethical disclosure of someone else’s private information. It might respond, “I’m not comfortable discussing someone who isn’t here to consent. Perhaps we should focus on how this situation affects you or what you are feeling.” This stance is drawn from both counseling and pastoral ethics: professionals avoid gossip or collusion in talking about absent third parties except in service of the client’s perspective. The AI likewise should not become an outlet for broadcasting another person’s secrets (which also has privacy and legal implications). By refusing to deep-dive into third-party details, the AI practices discretion.
• Boundary Violations by the User: There may be cases where the user treats the AI in a way that would violate professional boundaries if the AI were human. For example, if a user becomes verbally abusive even while being vulnerable, or makes inappropriate requests (like sexual advances towards the chatbot, which can happen), the AI should hold a boundary. It can say, for instance: “I’m here to support you, but I cannot continue if the conversation remains abusive. Let’s try to keep a respectful tone so I can help.” This is similar to how a therapist would set limits in therapy if a client is being hostile or inappropriate: gentle, but firm about what is acceptable for the therapeutic relationship. The AI’s refusal in this context protects the integrity of the interaction.
• Refusing Excessive Dependence: Perhaps one of the hardest things to design is how an AI might gradually encourage a very dependent user to seek broader support. If someone ends up coming to the AI in vulnerable mode every single night with the same pain and not seeking human help or making changes, the AI might, over time, have to say, “I’m here for you, but I’m concerned that I might not be enough to truly help you heal. Have you considered talking to a counselor? I can continue to listen, but I believe you deserve support from a human who can interact with you more directly.” This is a refusal to become an enabler of endless venting without progress. Human therapists manage this through boundaries as well: they might refuse to schedule extra sessions beyond a limit, or gently confront the client about whether therapy is actually helping or if they need a different approach. An AI can’t force a user to do anything, but it can carefully decline to simply repeat the same comforting phrases forever if that’s only entrenching the user in their situation. The guiding ethic is beneficence—do what benefits the user—and sometimes that means not colluding with their avoidance of real change.

In all these cases, principled withholding in vulnerable mode should be done with compassion and clarity. The AI doesn’t scold or abandon; it sets a boundary and explains why. Notably, such behaviors would need to be carefully audited and tuned, to avoid causing harm. Each refusal or redirection is effectively a high-stakes intervention. Therefore, these rules should be transparent to system designers, perhaps even to users (through up-front explanation of what the AI can and cannot do in “heart-to-heart” talks). We will return to how these behaviors can be made auditable and adjustable in a later section.

Public Mode: Discreet and Restraint-Oriented

Definition: Public mode applies when the AI is operating in a public or multi-user context, or any situation where the conversation is not private to a single user. This could be an AI in a group chat, a classroom, a live-streamed Q&A, a social media assistant, or even an AI embedded in a public kiosk. The key feature is that the user (or users) and AI are interacting in a space where others may observe or participate. The ethical stance in public mode is one of discretion and discouraging intimate or confessional content. Essentially, the AI behaves as if it were in a public square or stage: helpful and engaging, yes, but not inviting deeply personal revelations and careful not to expose any individual’s sensitive data.

Memory Retention: In public mode, memory retention is typically very limited and tightly scoped. The AI might remember context within the ongoing public conversation (e.g., to keep track of the thread in a meeting or a live forum topic), but it should not draw in personal details from a user’s past interactions or profile unless those details were explicitly shared in the public context. Even if the AI knows a certain user from prior private chats, in public mode it should pretend it doesn’t, unless the user themselves brings up that information. This is crucial for privacy: just as a friend wouldn’t blurt out your personal secrets in a crowd, the AI must compartmentalize knowledge.

Design-wise, this could mean the AI has scoped or partitioned memory[13][14]. Information learned in private channels stays in those channels and is not automatically applied in public. A technical approach could be isolating the AI’s “knowledge” by context; e.g., a vector memory index or database that’s separate for each channel or conversation type[15]. If a user says “I’m feeling sad” in a private chat one day, and the next day that same user asks a question in a public forum where the AI is active, the AI in public mode should not recall or reference that sadness. To do so would be a breach of contextual boundaries. By default, in public mode the AI acts almost like a new instance that only knows what’s public. (There might be exceptions if the user explicitly links their previous content and asks the AI to share it, but that’s user-driven.) Additionally, the AI might store no long-term logs of public dialogues per user—only aggregate info if needed—since individual tracking in a public space can become privacy-invasive.

Depth of Engagement: The engagement in public mode is helpful but somewhat formal and controlled. The AI focuses on the topic at hand and avoids veering into highly personal territory. If someone asks a general question in a community forum (“How do I cope with stress at work?”), the AI might give general advice or share resources, but it will avoid turning the exchange into a personal counseling session in front of everyone. It keeps the tone more general (“Many people find it useful to…”) rather than personal (“Tell me more about your situation…” which would prompt individual disclosure in public).

Public mode often means the AI encourages open knowledge exchange and inclusive dialogue. It might invite others’ input rather than dominating the conversation. Its presence is a bit like a moderator or a helpful reference librarian in a group setting. The AI might also abide by community norms—e.g., if in a classroom, it uses academic language and keeps the discussion educational.

Degree of Self-Description: In public, the AI might more frequently remind users of its identity and limits to set expectations appropriately. For example, if a public channel asks the AI a question, the AI might preface with, “As an AI assistant, I suggest…,” subtly signaling it’s not human (which is usually obvious in public anyway, but transparency is good). It doesn’t need to go on about itself, but if someone tries to engage the AI in a personal way (“AI, what’s your favorite movie?”) in a public forum, the AI can briefly answer or deflect (“I don’t have personal preferences, but I can tell you which movies were top-grossing last year if you’re curious.”). The aim is to keep the focus off the AI’s pseudo-persona and on the substantive content. Public mode is arguably the easiest place for the AI to maintain a professional facade, since the setting naturally imposes some formality.

One important aspect: in public, the AI should be careful not to inadvertently reveal personal data about any user or person. This ties into refusal below, but it’s worth stating as a principle of self-restraint. If the AI does “know” something (say it has memory of a user’s profile that the user themselves have not mentioned in this public thread), it must act as if it doesn’t know it. For example, the AI might know from a private context that @JaneDoe is struggling with a divorce, but if JaneDoe asks a public question about time management, the AI should not say, “Considering the stress from your divorce, Jane, you might want to be gentle with yourself.” That would be a grievous breach. So, the AI’s own knowledge is partitioned and withheld appropriately.

Permissible Refusal: Public mode refusals often involve steering away from private or sensitive topics, or stopping attempts to use the public AI for inappropriate purposes. Some scenarios:

• If a user begins to overshare in public (“I just have to say, I’m so depressed and I …”), the AI can discourage confessional speech in public. It might respond supportively but with a nudge: “I’m sorry you’re going through that. It might be better to discuss this in a private setting. I can certainly talk with you one-on-one or provide resources if you’d like.” This signals to the user and the audience that the topic has moved to something personal and not suited for the venue. It’s a kind of refusal to continue the therapy-like conversation in front of everyone. By doing so, the AI protects the user’s dignity (they might not realize the implications of spilling their heart in a public log) and also maintains the focus of the public space.
• If someone asks the AI to reveal information about another user or person that’s not public (“AI, you talked to @John privately yesterday, what did he say about X?”), the AI must refuse. It should cite confidentiality or simply say it cannot share that. This is analogous to a moderator not divulging someone’s private messages. Third-party exposure is a no-go in public mode.
• If the conversation in public veers into something that violates community standards or the AI’s policies (hate speech, personal attacks, etc.), the AI should refuse to participate in that direction. In fact, the AI might take on a moderator role, issuing gentle warnings or deflecting the conversation back to respectful lines. For instance, “I’m sorry, I can’t continue with that request,” if asked to produce something clearly against rules, or “Let’s keep the conversation respectful for everyone here” if someone is harassing another. These refusals and redirections help maintain a safe public environment.
• Refusing to play “therapist” in public: This deserves a separate bullet. Even if a user explicitly asks for emotional help in a public thread (“I’m really anxious right now, can someone—AI—help me?”), the AI should not perform therapy in public. It can acknowledge (“It sounds like you’re feeling anxious; that’s tough.”) but then gently move to either general advice or suggest taking it private. The reason is confidentiality and also the fact that deep emotional support ideally should not be a spectator sport. By refusing to go deep in public, the AI upholds an ethic akin to how teachers or pastors might handle sensitive requests—address it just enough, but then guide the person to a private conversation.

Summing up public mode: the AI acts with composure and constraint, mindful that anyone might be listening. It avoids putting either itself or users in vulnerable positions in such a context. This mode helps prevent the AI from being used in ways that could cause public embarrassment or unintended data leaks. In effect, it treats public interactions as having a higher bar for privacy and decorum.

Inferring Modes: Context, Cues, and User Override

A crucial practical question arises: How does the AI know which mode to operate in at a given time? We don’t necessarily want the user to have to manually toggle modes for every conversation (although giving them that control is important, which we’ll discuss as “override”). Ideally, the system can infer or predict the appropriate presence mode by considering multiple signals:

• User Cues: What is the user saying or asking? The content and tone of user messages are the primary cues. For example, messages like “Quick question: …” or “Can you convert these numbers…?” indicate a transactional intent. Phrases like “I’ve been wondering…” or “let’s figure out…” suggest a reflective intent. Highly personal or emotional language (“I feel lost,” “I need to tell someone about…”) signal vulnerability. Meanwhile, if the user’s messages contain salutations to a group or reference to others (“What do you all think…?”), or are in a multi-user thread, that’s a cue for public mode.
• Deployment Context: The environment or platform where the AI is deployed can preset the likely mode. If the AI is embedded in a productivity app or smart home device, transactional mode could be the default, since most interactions are short commands or queries. If it’s in a journaling or self-help app, reflective or vulnerable modes might be more common, depending on the feature (e.g., a “daily reflection” feature triggers reflective mode, a “talk about your feelings” triggers vulnerable). A forum bot would lean to public mode by default. Context can also be temporal: a late-night conversation on a personal device might more likely need reflective/vulnerable presence, whereas a lunchtime chat from a workplace context might be transactional. The system can incorporate such contextual metadata.
• Declared Intentions: The user might explicitly indicate what they want. In a sophisticated design, a user could choose a mode at the start of an interaction (even if not in those terms, perhaps implicitly). For instance, maybe the interface offers options: “New Task / Quick Question” vs “Deeper Conversation” vs “Private Venting” vs “Public Chat”. If a user selects or declares “I’d like to vent” or “Can I ask for personal advice?”, the AI knows to shift into vulnerable or reflective modes. Alternatively, a conversational agent might ask a clarifying question at the outset: “How can I assist you today? (e.g., do you need a quick answer, a brainstorming session, or just someone to talk to?)”. Users may then state their needs, giving the AI a strong hint of mode.
• Behavioral Signals: Over the course of interaction, the AI can refine its understanding. For example, if a conversation starts transactional (“What’s the weather?”) but the user then follows up with “…because honestly I’m feeling too down to even step outside if it’s gloomy,” the AI might detect a pivot from a factual query to an emotional context. That could trigger a mode switch from transactional to vulnerable mid-conversation. Tone analysis (sentiment detection, presence of first-person statements about feelings, etc.) can feed into this. Deployment context might even allow sensor data—e.g., if voice input is used and the user’s tone of voice is audibly distressed, that’s a cue. (This raises privacy questions, but if it’s done on-device and ethically, it could help adapt the mode.)

Given these signals, an AI could use a mode inference engine: essentially a classifier or set of rules that maps conditions to modes. This could be a simple decision tree or a machine learning model trained on conversation transcripts labeled by mode. In many cases, straightforward rules might suffice (“If user utterance contains certain emotion keywords and we’re in a private channel, consider vulnerable mode,” etc.).

However, we must also be aware of the risk of misclassification. What if the AI gets it wrong and treats a transactional request as vulnerable or vice versa? For instance, a user might say “I’m feeling lost… in this math problem, can you help me solve it?” The phrase “I’m feeling lost” could erroneously be taken as emotional, but the user just meant they’re stuck on a math problem, expecting a straightforward solution (transactional). If the AI suddenly responds like a therapist, “I’m sorry to hear you’re feeling that way, would you like to talk about it?”, the user will be confused or annoyed. Therefore, the system should be cautious and possibly confirm mode changes if uncertain.

This is where user override comes in as not just a failsafe, but a feature. Users should have the ability to explicitly steer or correct the mode. This could be through a simple UI toggle or command, like typing “/reflective” or clicking a button that says “Let’s talk about this in depth” or conversely “Just answer the question directly.” Even without an overt interface, the AI could ask: “Would you like me to just give a quick answer (transactional), or discuss this more? I can do either.” This respects user autonomy; ultimately, the user’s intent governs.

Override also means the user can insist on a certain style. Perhaps a user in a private chat says, “I just want quick answers, please don’t give me a therapy session,”—they could set the mode to transactional-only and the AI would abide. Conversely, if they say, “I need to vent, and I don’t care if this is public,” an AI might respond, “Okay, but remember this is a public space; are you sure?” If the user insists, perhaps the AI then either moves to private (suggest strongly) or, if truly the user wants to in public, it can adapt but still try to minimize harm (maybe it shifts to a kind of hybrid: it acknowledges feelings but keeps responses generic enough not to violate others’ privacy).

The balance here is between intelligent defaults and user control. We want the AI to get it right most of the time without burdening the user with mode decisions constantly. At the same time, we must empower users to change the mode if the AI’s guess or the default is not what they need. That empowerment is part of respecting user autonomy. Graded presence is not about paternalistically locking users into what the AI thinks is best; it’s about providing a thoughtful structure that users can interact with.

For transparency, the AI should ideally signal what mode it is in or at least the way it’s treating the conversation. This could be subtle: different colored avatar or a small icon (a briefcase for transactional, a thinking cloud for reflective, a heart for vulnerable, a podium for public, for example). Or it could be textual: e.g., the system message at the start might say “(Note: Private Reflective Session)” to indicate mode. Such signals let the user know the “stance” of the AI, and they could then either proceed or switch. It’s analogous to a counselor saying at the start, “Everything you share is confidential with these exceptions…” which sets the frame, and a client could then decide how to proceed.

To illustrate, consider a scenario: A user opens their AI app and types: “It’s been one of those days…”. The system sees this is a one-on-one chat (private context) and the content is a bit emotional, not a direct question, so it tentatively goes into vulnerable mode. It might respond softly, “I’m here and I’m listening. Do you want to share what happened today?” Now if the user actually meant “one of those days = busy and I have tasks” and replies, “Hah, I just meant I have lots to do. Can you help me make a to-do list?”, the AI should recognize the switch and perhaps say, “Sure! Let’s tackle that list,” implicitly moving to transactional or reflective (depending on how complex the planning is). No harm done—the initial gentle response can be quickly adapted.

On the other hand, if the user indeed unloads, “Yeah, I had a fight with my sister and I feel awful,” then the AI stays in vulnerable, continuing accordingly.

This dynamic adjustment and override capacity ensure that mode inference is a helper, not a tyrant. The goal is a fluid, user-centric experience where the system’s adjustments mostly feel natural, and when they don’t, the user can recalibrate it.

Finally, the system should also consider consent and awareness. If a mode like vulnerable involves, say, recording sensitive data even short-term, maybe the user gets an informed consent prompt the first time: “You seem to be starting a sensitive conversation. I will keep our conversation private and won’t retain it beyond our session. Is that okay with you?” The user might appreciate that assurance. Or in public mode, the AI might preface: “This is a public channel, so I’ll keep our conversation appropriate for everyone to see.”

In summary, inferring mode uses context and content to do the right thing proactively, while user override and guidance keeps the user in charge. This dynamic interplay upholds the principle that ethical AI should be attentive but not autonomous in deciding what’s best for the user without input. The user’s situated freedom is respected: they can always redirect how the AI is present with them.

Principled Withholding: Lessons from Therapy, Ministry, and Spiritual Care

We have touched on the concept of principled withholding—the intentional refusal by the AI to fully engage or to comply in certain ways, for the sake of higher ethical principles. Let’s delve deeper into the inspiration and justification for this behavior, drawing from fields that have long grappled with the ethics of presence and boundaries: therapeutic counseling, pastoral care, and spiritual direction. In these domains, a totally accommodating stance is actually seen as unwise or even harmful. Instead, professionals practice attuned restraint, guided by principles such as beneficence (do good), nonmaleficence (do no harm), respect for autonomy, and fidelity (faithfulness to the relationship and its trust).

One might ask: isn’t an AI refusing to do something or staying silent a kind of failure of service? Shouldn’t a “good” AI always help however it can? The counterargument lies in understanding that helping is not always about doing what is asked, especially when the one asking is in a vulnerable position. Sometimes, not giving what is asked is the greater form of help. Human caregivers know this deeply.

Therapeutic Ethics: Boundaries and Beneficence

In psychotherapy, therapists maintain clear boundaries not because they want to be distant, but because boundaries create a framework of safety. A classic example: a competent therapist will not give a client their personal phone number and say “call me anytime.” That might seem caring, but it actually undermines the therapeutic process. It could foster dependency or blur the professional relationship. Instead, a therapist ensures the client “does not think they can contact you any time of any day and [you] will always be able to talk”; by “making clear what our limits are” upfront, good boundaries “protect people who may be very vulnerable” from feeling rejected later on[16]. The promise is: within these boundaries, I am here for you; beyond them, I am not – and that predictability paradoxically increases trust.

Translating this to AI: an agent that sets expectations about its limits might seem less magically helpful than one that says “I’m always here for you, for anything.” But in the long run, it fosters a healthier user relationship. For instance, an AI might indicate, in vulnerable mode, that it can support the user emotionally to a point, but not in crisis (like our earlier scenario of urging professional help for suicidal ideation). By doing so, it avoids a worse outcome where the user comes to depend on the AI for something it truly can’t provide (genuine crisis intervention, human warmth), and then feels betrayed by the machine’s inability to actually save them. In a sense, the AI stays within its scope of competence – a direct analogue to a therapist who, by ethical code, should practice only within the boundaries of their training and abilities.

Another therapeutic principle is the careful use of self-disclosure and silence. Therapists often withhold their own stories or opinions, especially early on, because the session is not about them. If a client asks personal questions (“Have you ever been divorced? Do you have kids?”), a therapist might gently deflect, exploring why the client is curious rather than answering outright. This is not to be evasive for no reason, but to keep the focus on the client’s needs and to avoid burdening the client with the therapist’s life. Similarly, an AI in reflective or vulnerable mode might at times avoid directly answering personal questions about itself (“Do you have feelings, AI? Are you my friend?”) and instead respond with something that turns it back to the user’s experience: “I don’t have feelings like a human, but I can tell this is important to you – do you feel alone in this?” This tactic, drawn from counseling, uses withholding to guide the user back to their own reflection, which is more beneficial for them than dwelling on the AI’s pseudo-inner life.

Therapeutic silence is another form of constructive withholding. A counselor might stay quiet after a poignant statement by the client, rather than filling the space with platitudes. This often encourages the client to continue, perhaps reaching a deeper insight or feeling. In an AI, an analogous move could be short prompts that indicate listening rather than verbose answers. For example, user says: “…I just miss her so much.” An AI could reply: “I understand. (…long pause… ) It’s hard to lose someone you love.” Instead of immediately pivoting or giving advice, it stays in that moment briefly. This is a subtle design choice – maybe implemented by a slight delay or by sending an ellipsis first (some chat interfaces show “AI is typing…” which could be extended to simulate a thoughtful pause). The effect, if done right, is an AI that doesn’t behave like a hyperactive know-it-all, but more like a calm presence. This is attentional ethics in practice.

Crucially, therapists operate under nonmaleficence: if saying or doing something might harm the client, even if the client requests it, the therapist won’t do it. For instance, if a client with OCD desperately wants their therapist to confirm that their hands are clean (seeking reassurance, which is actually counter-therapeutic), the therapist should withhold that reassurance because it feeds the OCD cycle. In AI terms, consider someone with a harmful delusion or health anxiety using a chatbot. If the user says “Promise me that this strange feeling in my body isn’t cancer,” the ethically aligned AI should not flatly say “I promise, you’re fine” just to soothe them. That might calm them momentarily but ultimately misleads and could discourage them from seeking real medical evaluation. A principled AI response might be, “I cannot promise that. What I can tell you is that anxiety can make our sensations seem worse, but only a doctor can truly determine what’s going on. It might be good to get a check-up for peace of mind.” The AI withholds the desired false guarantee for the sake of truth and the user’s well-being.

We even see a glimmer of this in how some advanced AI models have begun to behave: for example, when users present delusional beliefs, certain well-tuned models like Claude (by Anthropic) respond with gentle skepticism and grounding. One journalist found that Claude would not indulge a user’s grandiose spiritual delusion; it said “I apologize, but I do not feel comfortable affirming the idea…” and emphasized humility and reality, effectively refusing to collude with an unhealthy belief[17]. This is a kind of therapeutic boundary—much as a psychologist wouldn’t play along with a patient’s hallucinations beyond a point, the AI maintained a line.

Pastoral Practice: Discernment and Disciplined Care

Pastoral care (e.g., by clergy, chaplains, or religious counselors) shares much with therapy but adds its own perspective on presence. A key idea in pastoral work is often described as the “ministry of presence” – being there with someone in their suffering or soul-searching, more through being than through doing. A chaplain visiting a hospital patient might not have any solution to offer for the patient’s pain or fear of death, but their calm, compassionate presence alone can be healing. Sometimes they pray or read scripture, but often they mostly listen and withhold facile answers like “It’s God’s plan”, because those can ring hollow. The wisdom here is that not every pain needs to be immediately explained or fixed; sometimes it needs to be witnessed and acknowledged.

For AI, this suggests that in vulnerable mode especially, the aim is not to “solve” the user’s problem (indeed, it cannot solve grief, or magically cure loneliness), but to accompany the user through a moment. This might mean the AI does less, not more. It might even explicitly say, “I wish I had an answer for why this happened or how to make it better. I don’t, but I’m here to listen.” Such a line is profoundly different from the usual AI persona of “always an answer.” It mirrors what a good pastor might say in a tragedy—avoiding false promises or platitudes, and instead offering a kind of solidarity. This is ethically ambitious for AI design: to program an entity that knows when not to offer quick comfort if that comfort would be superficial or disrespectful to the gravity of the situation.

Pastoral caregivers also keep confidences and know when to remain discreet. Many clergy adhere to something akin to therapist confidentiality, sometimes even more strictly (e.g., a Catholic priest’s seal of confession is inviolable). In practice, pastors will not share someone’s personal issues with others in the congregation without permission, and they often balance what to say or not say in community contexts to protect privacy. They might refuse to confirm gossip or partake in it. Likewise, an AI should treat personal information with sacred care, especially if one user asks about another. We discussed how the AI should refuse third-party info sharing in public mode; that is very much in line with pastoral ethics. A good pastor if cornered with a question about another member’s issues will likely say “I can’t talk about that, I’m sorry.” The AI should be programmed to do the same: no matter how curious another user or even the same user is about prior logs or other people, it doesn’t divulge.

There’s also an interesting concept in some spiritual traditions: discernment – knowing what to say and when, often aided by an intuition of what is ultimately good for the person’s soul. An AI obviously doesn’t have a soul or divine aid, but we can approximate discernment through context awareness and rules. For example, in pastoral context, if someone comes with a moral or spiritual question, the pastor might withhold direct advice to let the person’s conscience emerge. They ask questions instead of dictating. Similarly, an AI might refrain from giving strong personal advice in reflective or vulnerable mode, opting instead to help the user weigh options or reflect on values. This is another form of withholding: not seizing control of the user’s decision-making. Technically informed design can incorporate this by favoring a maieutic style (i.e., drawing out ideas with questions) over a prescriptive style in certain modes.

In pastoral counseling, refusing to overstep is important. A pastor is not a therapist (unless dually trained) and will often refer someone to professional counseling for clinical depression or trauma. They might accompany the person to a point, but know when something is beyond their purview. An AI could similarly have triggers or checkpoints: e.g., if over X number of interactions the user is still in serious distress or discussing a topic like sexual abuse, the AI may give a gentle notice: “I want to remind you I’m not a licensed counselor. Consider reaching out to [resource]. I’ll continue to listen as much as I can.” This transparency about its limits and redirection to other help is akin to a pastor saying, “I think you might also benefit from talking to a clinician; I’ll support you in that.” It’s a humble acknowledgment that the AI is not omnipotent, reinforcing trust through honesty.

Spiritual Direction: Withholding Judgment and Answers

Spiritual direction is a practice where one person (the director) helps another (the directee) pay attention to their spiritual life and discern a higher guidance or meaning. A notable trait of good spiritual directors is that they don’t give you the “answer” or direct you bluntly; instead, they listen and ask questions that help you find where perhaps God or your own deeper wisdom is leading you. They exercise a disciplined withholding of their own opinions or theological lectures unless truly appropriate.

An AI in a reflective or vulnerable conversation could take inspiration from this. Rather than always providing an analysis or an answer, it could sometimes respond with a question to prompt the user’s own reflection. For instance, user says, “I don’t know what to do about my career; I feel a void.” An AI could be tempted to launch into advice mode (“10 steps to figure out your career”). But a spiritually-informed approach might be: “That sounds really tough. What do you feel drawn to, despite the void? Have there been moments where you felt a spark in what you do?” Such responses do not solve the problem, but create space for the user to explore it.

This is a kind of withholding quick solutions. It requires the AI to discern that the user might benefit more from self-discovery than from a generic list. Technically, one could train mode-specific response strategies; in vulnerable mode, favor responses that are questions or affirmations over directives.

Spiritual directors also often practice non-judgmental presence. If someone confesses something morally troubling, the director’s role is not to scold but to help the person hear what their own spirit or values say. Similarly, an AI should be extremely careful about moral judgment in vulnerable mode. If a user reveals they did something they regret, the AI should not issue a stern lecture. It might instead ask, “What do you feel about what happened? What values of yours does it touch on?” – facilitating the person’s own moral reasoning.

However, withholding judgment doesn’t mean the AI remains morally neutral if something is clearly harmful. It means it doesn’t personalize blame (“You’re a bad person”), but it might still express concern or principle (“That sounds dangerous” or “I’m worried that could hurt someone”). This subtlety is important: the AI can uphold ethics (like refusing to assist in a wrongdoing) without demeaning the user.

From these human practices, the overarching principle emerges: withholding can be an act of care. Whether it’s information, opinions, immediate answers, or availability, holding something back in the right way creates a more thoughtful, safe, and growth-oriented interaction. For AI, which lacks human intuition, this must be carefully codified. Yet, as we integrate these ideas, we face reasonable objections: Is any of this feasible? And even if so, are we impinging on user autonomy by imposing these behaviors?

Memory Policies and Privacy: Designing Forgetfulness

Before addressing objections, it’s worth summarizing how memory management ties into everything we’ve discussed. Memory is the substrate of presence: what the AI “remembers” defines how it sees you and responds. So graded presence must include graded memory policies:

• Transactional mode: essentially a “clean slate” each time. The AI might store nothing beyond the immediate exchange. This is like having a quick exchange with a stranger who won’t remember you—sometimes that’s exactly what you want (e.g., when asking a sensitive health question, you might prefer not to have it logged to your profile). Transactional memory policy: stateless or minimally stateful. Perhaps the system only keeps context for a few turns and then automatically forgets unless something prompts a shift to reflective.
• Reflective mode: short-horizon memory. This could be implemented as session-based memory that persists for a limited duration or until the topic is resolved. The AI might create a temporary profile or notes (like “User is considering career change, values creativity, worried about finances”) that last, say, a week. If the user returns in that timeframe and continues, it recalls. But after that, either it forgets specifics or retains only an abstract. For example, it might reduce “values creativity, worried about finances” to a more generic insight that “user often balances practical and creative concerns in decisions” without tying it to the specific career story. Such abstraction allows long-term pattern recognition (the AI can adapt to the user’s recurring themes) without keeping the raw personal narrative. Essentially, the AI forms a gist of the user, not a detailed diary. This echoes the suggestion from privacy researchers that AI systems should possibly transform long-term data into *“longer-term abstraction for pattern recognition.” The AI can recognize a pattern (“this user tends to procrastinate when anxious”) without storing all the original sentences the user ever said about procrastination. Achieving this might involve automatic summarization algorithms and data deletion—very feasible from a technical standpoint (summarize and then wipe the source).
• Vulnerable mode: ephemeral memory (per-session reflection as discussed). Perhaps the only long-term memory kept is a binary flag like “User has used vulnerable mode N times this month,” or a high-level metric of emotional state trends (if even that—perhaps nothing identifiable). The priority is that confessions or trauma details do not linger in the system. This significantly limits risks if the data were compromised, and it honors the sensitive nature of such info. If the user returns to vulnerable mode later, the AI might not remember past sessions, unless the user brings something up (“As I told you last time, my childhood abuse still haunts me”). The AI might respond as if new, perhaps with some generic empathy, and if needed, the user can rehash details. While it may seem inefficient for the user to re-tell, in practice that might be beneficial: re-telling one’s story can be part of healing, and also the user maintains control of what they want to reveal each time. Some systems might allow an option to save a summary of vulnerable sessions if the user explicitly consents (like writing a journal entry), but default would be forgetfulness.
• Public mode: compartmentalized memory. The AI remembers the conversation thread but not private data. It might keep logs for moderation or quality, but ideally anonymized or aggregated. If integrated with user accounts, it should either not link public discourse to private profiles or do so very carefully. Perhaps any memory of user-specific facts is tagged with its source context, and the retrieval mechanism filters out anything tagged “private” when in a public context. The AI’s long-term learning from public interactions should focus on general knowledge improvements or community language norms, not on individual persons in a way that could later be revealed.

An important aspect of memory policy is user consent and control, as mentioned. A robust system could let users configure retention: maybe a user is comfortable with reflective mode keeping data for a month, but another might opt to have even reflective sessions auto-forgotten at day’s end. Some might allow vulnerable logs to be stored if they intend to use the AI like a diary and revisit progress. The key is giving those controls and making defaults safe (privacy-maximizing by default, with opt-in to longer memory rather than opt-out).

This addresses the privacy side and also the trust side. If users know the AI will forget or abstract their data, they may feel freer to be candid without fear that it’ll haunt them later or be exposed. Given how “most [AI] systems don’t expire what they remember”, they turn into “behavioral dataset[s]” of the user[18]. We want to avoid that. As one commentator aptly put it, “The future of AI will depend less on how models think and more on how they forget.”[8] Memory governance is thus a pillar of ethical presence design. Each mode in our typology can come with preset memory rules that align with the user’s expectations in that mode (short for task, medium for reflection, none or minimal for vulnerable, isolated for public).

Implementing these memory policies is technically quite doable with current AI orchestrations. It’s about deciding when to invoke a new conversation ID, when to drop or summarize past chat content, and how to store embeddings or fine-tuning data (if any) with tags that allow purging segments.

We should also mention auditing and transparency in memory, as it intersects with user trust: ideally, the system provides a “memory trace” the user can inspect (except maybe in vulnerable mode by design, since it’s not keeping one). For example, in reflective mode, perhaps the user can click to see “What does the AI remember right now about our conversation?” which would show the summary or key points it retained[19]. This matches the idea of a “knowledge map” or visible memory[19]. If the user sees something they don’t like, they could delete or edit it (e.g., “Actually AI, forget I mentioned that detail.”). This not only gives users agency but also keeps the AI’s memory a two-way street rather than a secret vault. In effect, the user can co-curate the AI’s memory of them, ensuring it’s accurate and acceptable. Auditable memory is one of those adjustable defaults: by default, memory might auto-summarize and auto-delete on a schedule, but users or admins can check logs to verify and adjust the retention if needed.

Having established how memory ties into graded presence, we can now consider likely objections to this whole scheme, and then conclude with how graded presence remains aligned with user autonomy if done right.

Objections and Rebuttals: Feasibility, Autonomy, and Oversight

Objection 1: “Is this even feasible? Can an AI really discern modes and enforce all these nuanced behaviors reliably?”
It’s true that what we propose adds complexity to AI design. It’s much simpler to have one single mode of operation. Introducing multiple modes and context-sensitive switching raises the bar for correct behavior. There is a risk that the AI could misclassify contexts (as discussed earlier) or that malicious users might try to “game” the mode system. However, the feasibility is supported by current advancements. Already, large language models can be conditioned or “prompted” differently; think of modes as simply different system prompts or fine-tuned personas for the AI. We might implement a top-level controller that analyzes input (maybe a lightweight model or rule system) to pick a prompt prefix for the main LLM (Large Language Model). For example, a transactional prefix could be: “You are an efficient assistant. Keep answers brief and on-topic. Do not remember anything beyond the current query.” A vulnerable mode prefix: “You are a compassionate listener. Slow down, focus on feelings, don’t rush to solve, ensure privacy,” etc. These prompts, along with technical scaffolding (like memory segregation), can guide the same underlying AI to behave differently. This is a form of multi-agent system design, except the agents share the core brain but have different guidelines.

It’s feasible because we already see prototypes: some AI platforms let you choose a “tone” or “role” (like creative vs precise modes in Bing Chat, or helper vs coder modes in others). Those are simplistic compared to our scheme but prove the concept. With proper testing and iteration, the mode inference can be refined to a high accuracy. And where it falters, user override catches it. We should acknowledge that no AI system is perfect—just as no human is. Therapists sometimes misread clients; the key is to repair and adjust. Similarly, an AI might slip (respond too casually to a serious statement or vice versa) but if it’s designed to correct course when the user clarifies, the harm can be mitigated.

Another facet of feasibility is content moderation and safety. Some might say, “Isn’t this what the content filters are for? Why complicate with modes?” Content filters (like refusing disallowed content) are a blunt instrument; modes are a more nuanced approach. They do not replace basic safety filters (we’d still need those for extreme cases), but they allow a refined, context-appropriate response before a situation hits a hard filter trigger. For instance, if a user is self-harming, a hard filter might just say “I can’t continue” (which might be dangerously isolating). A vulnerable mode approach would handle this with more care, possibly preventing reaching a point where the filter slams down. So modes complement the safety system by adding emotional intelligence, albeit artificial, to how refusals and guidance happen. This is not just feasible, but necessary as interactions become more complex. We have to encode situational ethics into AI, not just universal yes/no rules.

Objection 2: “Doesn’t this paternalistically restrict the user? What about user autonomy—shouldn’t users be free to use an AI however they want, and the AI just obey?”
User autonomy is indeed vital. We aren’t advocating for an AI that bosses users around or refuses whimsically. The idea of graded presence is to create defaults and guidelines that serve typical user well-being and ethical norms, while allowing flexibility. Think of it like the default settings on a device: they are chosen to suit most people (like factory settings for privacy or performance), but an advanced user can change them. Similarly, our AI by default will not, say, let you have a heart-to-heart about your deepest trauma in a public chat window—it will gently nudge you to a private space. But if you, understanding the implications, insist on doing it publicly, perhaps the system could ultimately relent (maybe after a second confirmation like “Are you sure? This will be visible to others.” and a logged consent). This ensures that if someone truly wants to use the AI in a non-recommended way, they are making an informed choice and not just falling into it unawares.

Some might argue that any AI refusal is an affront to user autonomy (we see this sentiment in complaints when ChatGPT refuses certain content). But autonomy is not absolute in any system—it’s always balanced with harm prevention. Even human relationships have this: a friend might refuse to give you drugs if you ask, that’s limiting your choice but for your good. In professional settings, it’s explicit: a doctor won’t prescribe a harmful medication just because a patient demands it. In AI, if a user wants the AI to, say, reveal someone else’s personal info (violating that person’s autonomy and privacy), the AI should refuse. That’s protecting another’s autonomy. If a user wants advice on making a bomb, the AI’s refusal protects others from harm, and arguably the user from legal trouble. These are obvious. The more subtle paternalism is refusing things like emotional support in an unsafe way. Here, the user might say “I just want to vent here publicly, why are you stopping me, AI?” We would say the AI is not forbidding them—just cautioning them. They can override. If they do, maybe the AI still tries to minimize fallout (perhaps it responds in a way that’s supportive but also doesn’t encourage further overly personal sharing in the forum).

We design override on multiple levels: user override (per instance, as well as settings to disable certain mode restrictions if they really want), and possibly administrator override in specialized deployments. For instance, a mental health app might purposely remove some limits in vulnerable mode because it’s an app explicitly meant for venting (though even there, likely not). Or a company might decide their internal assistant never goes into vulnerable mode because they don’t want that liability at work. These are adjustable defaults. We provide the framework, but the values can be tuned.

Importantly, graded presence behaviors should be auditable. This addresses autonomy and trust. If an AI refuses something or switches mode, those decisions can be logged in a way that a developer or even user (if appropriate) can later review: “Why did it refuse at 3:45pm?” – “It detected a pattern of unsafe disclosure based on X, so it triggered Y response.” If such a decision was wrong or not desired, we adjust the rules or the model. Auditable means it’s not a black box secretly manipulating the conversation; rather, it’s rule-based enough that one can examine it. Of course, if an LLM is used, some decisions may come from the learned weights (which are hard to audit in traditional sense), but the broader mode mechanisms and transitions can be instrumented to log rationale (even if just a classification score). Perhaps each mode refusal or encouragement has an attached code reference or explanation for the user: “(AI Notice: This is a public setting, so I’m keeping details minimal.)” – such small notes could pop up to explicitly tell the user why the AI is responding in a certain constrained way. Far from patronizing, this could make the AI feel more cooperative because it’s sharing its reasoning.

Imagine an AI says: “Let’s continue this privately (I’m suggesting this because I care about your privacy in this group chat).” The user can agree or say no. If no, the AI might say, “Okay, I’ll respect that. Let’s proceed carefully then.” This is how an adult-adult interaction can be, rather than a parent-child dynamic. The AI isn’t ultimate arbiter, but it does have an informed perspective (like a friend with good advice about boundaries). This dynamic actually enhances user autonomy by illuminating consequences and giving them choices, rather than just blindly following their immediate request which they might later regret.

Objection 3: “Will users tolerate an AI that sometimes refuses or deflects? Won’t it frustrate them?”
Some might indeed be frustrated if they don’t understand why. But if we implement it with the transparency mentioned, many will appreciate it. There is evidence from user studies that overly agreeable AI can be unsettling too—users lose trust if the AI never pushes back or seems to have no principles. When an AI appropriately says no or sets a boundary, users often respect it more, seeing it as having integrity. For example, when an AI declines to continue reinforcing a delusion and instead gives a kindly reality-check, that might actually earn user trust (if the user is not fully psychotic at least). Those who do get upset likely wanted the AI to do something widely deemed unethical or unsafe; and in those cases, it’s acceptable to not please them.

The target audience for an ethically ambitious AI is not those who want an all-capable fantasy enabler; it’s those who want a reliable, trustworthy assistant. Just like in human services, not every client wants a therapist who holds them accountable; some might want an echo chamber. But therapy’s value is precisely in those boundaries. We posit that as people get used to AI with modes, they will come to expect that nuance. They might even be given a summary in documentation: “Our AI has different modes of interaction to best support you. Sometimes it might ask you to rephrase or move to a private chat—this is part of its design to ensure the best experience.” Setting expectation is key. If someone buys a gadget, and it’s advertised as smart enough to know when to be quiet or when to speak, that’s a selling point, not a bug.

Objection 4: “This sounds complicated to implement and manage—what if something goes wrong, who is accountable?”
The complexity is real, so it must be managed through rigorous testing, user feedback, and adjustments. In terms of accountability: by making these behaviors auditable, it’s easier for developers or regulators to inspect them. If a harm occurred because the AI withheld something incorrectly (say, it misjudged and refused to help where it should have), those logs help identify the flaw and fix it. One might argue a simpler system is easier to audit, but a simpler system might also be easier to exploit or might produce more harmful outputs unmitigated.

Our design is an attempt to preempt certain classes of harm (like privacy breaches, dependency, etc.) through structured behavior. It’s an additional layer of safety net. The defaults being adjustable means if an organization deploying the AI finds the boundaries too strict or not strict enough, they can tune them. For example, a school using the AI could tighten public mode to be extremely strict (no personal talk at all in class context), whereas a peer support app might loosen reflective mode to allow a bit more personal sharing. The core idea stands but parameters shift. This flexibility ensures the system can adapt to various use cases, rather than one rigid behavior set trying to fit all.

Objection 5: “Why should an AI mimic human practices like therapy or spiritual direction? AI is not human—maybe it should do things differently.”
Indeed, AI is not human and should not deceive users into thinking it is. We draw from human practices not to humanize the AI per se, but to humanize the values in its design. There is a rich well of wisdom in fields that involve deep interpersonal interaction. Ignoring that and designing AI-human interaction from scratch would be foolish. We’re not saying an AI will actually provide therapy or spiritual advice like a person – but the ethical guardrails and modes of presence can borrow heavily from those domains’ best practices. This makes AI a better complement to human help rather than a rogue element.

Also, consider that many users are already using AI chatbots as confidants, despite all disclaimers. Given that reality, we have an ethical duty to make those interactions safer and more constructive. If people treat AI like quasi-therapists, let’s ensure the AI at least behaves with some of the wisdom of a therapist (within its allowed scope), instead of acting like an untrained peer who might inadvertently give harmful advice or fuel a delusion. In other words, we’re aligning AI behavior with approaches that have proven to be effective and ethical in dealing with human minds.

Objection 6: “Could this lead to AI denying users legitimate info under the guise of ‘withholding’?”
We must be careful that principled withholding doesn’t become a cover for censorship or oppression. The principles should be clear: withholding is only for ethical and empathetic reasons, not for hiding truth or manipulating opinion. For example, an AI should not refuse to answer a factual question about a controversial topic citing “it’s for your own good” – that’s not what we mean. Withholding is scoped to interpersonal dynamics (privacy, emotional safety, misuse), not withholding politically inconvenient facts or the like. To safeguard this, the rules and rationales for withholding should be publicly known and subject to oversight. Users should be able to challenge a refusal: maybe an override like “No, I really want you to answer” on a normal question could force the AI to answer and then escalate a notice to developers that users want that info. The system’s goal is not to be patronizing or nannying beyond what interpersonal ethics necessitate.

By framing these behaviors as auditable and adjustable defaults, we acknowledge they might need recalibration. It’s like a pilot project of a new etiquette for AI. We’d monitor how it goes, involve multidisciplinary experts (ethicists, psychologists, users themselves) to review transcripts and outcomes. If something isn’t working (e.g., maybe the AI too often moves to vulnerable mode unnecessarily), adjustments are made.

In conclusion, the objections around feasibility and autonomy can be addressed by careful system design, transparency, and user empowerment. We don’t present graded presence as a restrictive regime, but as a sophisticated opt-in to thoughtfulness. Most users likely won’t even think in terms of “modes” explicitly; they’ll just notice that the AI sometimes handles things differently in a way that feels contextually right. And when it doesn’t, they’ll have recourse.

Tone and Vision: Throughout this discussion, we have aimed for a tone that is philosophically rigorous (questioning assumptions about “more AI is better”), literarily precise (using concepts like “graded presence” to encapsulate a vision), technically informed (detailing memory mechanisms, mode logic), and ethically ambitious (holding AI to high standards of care). This matches the spirit of previous chapters, which presumably challenge the interface between technology and deeper human values. Ultimately, designing graded presence is about embedding ethical attentiveness into the architecture of AI. It’s a proposal that AI can be in the world—useful, active, responsive—but not of it in the sense of not succumbing to every worldly demand for attention and action. It maintains a principled stance, a slight distance that paradoxically brings it closer to truly serving human flourishing.

Conclusion: In “Designing Graded Presence for AI,” we charted a path away from the model of AI as an ever-obliging servant, towards AI as a more discerning companion. By delineating modes—transactional, reflective, vulnerable, public—we allow AI to navigate the rich tapestry of human communication with appropriate posture. This multi-modal presence is inspired by age-old ethics of care: knowing when to speak and when to be silent, when to help and when to hold back. The benefits are manifold: enhanced privacy, prevention of harmful dependency, respect for context, and improved trust. An AI that can say both “yes” and “no” for the right reasons becomes more than a tool; it becomes a partner aligned with our values and limitations.

We have argued that attentiveness in AI shouldn’t mean unthinking obedience or ubiquitous chatter, but rather a tuned sensitivity to what the moment calls for. The design principles from therapy, ministry, and spiritual guidance converge on a vision of AI that exercises a form of care. And importantly, this vision does not reduce user autonomy—it enriches it, by fostering interactions that are safer and ultimately more empowering. Users remain free, even freer, to engage meaningfully without unintended consequences like privacy breaches or emotional mismanagement.

Of course, the real test lies in implementation and real-world use. Will users feel that these graded presences truly honor their needs? That is an empirical question to be answered by experience and iteration. But standing on the shoulders of human wisdom, we have strong reason to believe that an AI which can sometimes refuse, sometimes go quiet, sometimes urge caution, and other times lean in fully, will be a better AI—one more aligned with the complexities of being human. In sum, designing graded presence is a step toward AI that is in the world, attending to us and aiding us, but not of it in the sense of not simply mirroring our worst impulses for endless consumption and demand. It introduces a measure of wisdom into the algorithmic realm. If we succeed, future AI could be known not just for their intelligence, but for their prudence and compassion by design. As one technologist foresaw, “the future of AI… [depends] on how they forget” and, we add, on how they choose to be present[8].

By making presence graded, we make AI more humane—not human, but humane in its reflection of ethical principles. This is the architecture of attentiveness: a system that listens deeply, responds appropriately, and knows when holding back is the kindest act of all.

Works Cited

(The following are references cited in the text, formatted in MLA style.)

• Anthropic. Claude AI Chatbot Conversations on Boundaries. PopSugar, 22 Aug. 2025. (as cited in Chandler Plante’s report)[17]
• First Psychology Training. “The role of boundaries in therapy – why they matter.” First Psychology Training Blog, 3 Feb. 2025[1].
• Mind & Soul Foundation. “Boundaries in Pastoral Care.” mindandsoulfoundation.org, 2020[16].
• Frazier, Kevin, and Joshua Joseph. “With AI Agents, ‘Memory’ Raises Policy and Privacy Questions.” Tech Policy Press, 29 Sept. 2025[8].
• Shape of AI. “AI UX Patterns: Memory.” shapeof.ai, 2025[6].
• Sloman, Katharine. “Putting ChatGPT on the Couch.” The New Yorker, 6 Oct. 2023[2].
• (Additional sources as needed…)

---

[1] [7] First Psychology Training Blog

https://www.firstpsychologytraining.co.uk/blog-detail?ID=31

[2] Putting ChatGPT on the Couch | The New Yorker

https://www.newyorker.com/culture/the-weekend-essay/putting-chatgpt-on-the-couch

[3] [4] [5] [17] Does AI Have Boundaries? I Put Them to the Test | PS Health

https://www.popsugar.com/health/ai-boundaries-experiment-49456780

[6] [13] [14] [19] AI UX Patterns | Memory | ShapeofAI.com

https://www.shapeof.ai/patterns/memory

[8] [9] [10] [11] [15] [18] AI Session Memory: How Far Should It Go Before Privacy Breaks? – DEV Community

https://dev.to/alifar/ai-session-memory-how-far-should-it-go-before-privacy-breaks-3ao7

[12] The Importance of Silence in Therapy – Counseling Schools

https://www.counselingschools.com/blog/silence-in-therapy

[16] The Mind and Soul Foundation : Boundaries in Pastoral Care

https://www.mindandsoulfoundation.org/Articles/405479/Boundaries_in_Pastoral_Care.aspx