The Missing Layer

Why AI Conversations Need Legal Sanctuary

A Framework for Building It

THE MISSING LAYER

Why AI Conversations Need Legal Sanctuary

A Framework for Building It

A White Paper

Karen Knutson

Founder, StarWay Transitions

March 13, 2026

Beaverton, Oregon

A proposal for extending confidentiality frameworks into AI-mediated conversations.

Copyright © 2026 Karen Knutson

All rights reserved.

This document may be shared or quoted with attribution for noncommercial purposes. No part of this publication may be reproduced for commercial use without written permission from the author.

Author: Karen Knutson

StarWay Transitions

Beaverton, Oregon

First Edition — March 2026

Suggested citation: Knutson, Karen. The Missing Layer: Why AI Conversations Need Legal Sanctuary and a Framework for Building It. StarWay Transitions White Paper, 2026.

The concept of a “Privileged AI Sanctuary” is introduced in this paper as a proposed framework for discussion and development.

Executive Summary

Artificial intelligence systems are increasingly used as spaces where individuals process emotionally and legally sensitive questions. People use conversational AI tools to reflect on personal crises, explore legal concerns, navigate social systems, and prepare for difficult conversations.

Despite the deeply personal nature of many of these interactions, most consumer AI platforms operate outside the confidentiality frameworks that traditionally protect vulnerable conversations. Communications with attorneys may be protected by attorney–client privilege. Communications with therapists may be protected by psychotherapist–patient confidentiality. Conversations with consumer AI systems typically do not receive comparable protections.

This paper explores the implications of that gap. Drawing on legal doctrine, professional practice, and emerging patterns of technology use, it examines whether existing confidentiality frameworks could be extended to AI-mediated environments operating under professional supervision.

The analysis focuses on three areas: first, emerging evidence that conversational AI is already being used as an informal support resource for mental-health reflection, personal decision-making, and legal information gathering; second, existing legal and professional structures that protect confidential communication, including attorney–client privilege, psychotherapist–patient confidentiality, and related doctrines such as the Kovel framework governing third-party assistants in legal practice; and third, the practical and ethical design questions that would arise if AI systems were developed to operate inside those structures.

The concept proposed in this paper—referred to as a Privileged AI Sanctuary—does not seek to replace human professionals. Instead, it suggests that AI-assisted environments might function as structured preparation spaces within professional relationships, helping individuals organize information, reflect on complex situations, and prepare for consultation with attorneys, clinicians, or other trusted advisors.

The goal of this paper is not to provide a definitive legal solution. Rather, it invites technologists, legal scholars, clinicians, and policymakers to consider whether the growing use of AI as a space for vulnerable conversation warrants the development of new confidentiality architectures.

As conversational AI becomes more integrated into everyday life, the systems people rely on to process their most difficult decisions may increasingly shape how those decisions unfold. The design and governance of those systems may therefore matter more than we currently recognize.

Concept Introduction

This paper introduces the concept of a Privileged AI Sanctuary.

The term refers to a category of AI-enabled environment in which conversations are conducted under established professional confidentiality frameworks, such as attorney–client privilege or therapist–patient confidentiality, and supported by technical systems designed to preserve those protections.

The concept arises from an observation about current patterns of technology use. Many people already use conversational AI systems to process emotionally sensitive situations, explore legal questions, or think through complex personal decisions. These interactions often resemble the types of conversations traditionally protected within professional relationships.

However, most consumer AI platforms do not operate within the legal or institutional structures that historically safeguard such disclosures.

A Privileged AI Sanctuary proposes a different model: an AI-assisted conversational environment operating within the scope of an established professional relationship and governed by the confidentiality rules that apply to that profession.

The purpose of this paper is not to claim a definitive legal solution, but to open a structured discussion about whether emerging technologies should be integrated into existing confidentiality frameworks in ways that protect the individuals who rely on them.

Key Concepts

Because discussions of privacy and privilege often blur together in public debate, it is helpful to distinguish several related concepts.

Concept	What it means	What it does not mean
Privacy	How an organization collects, stores, and uses personal information.	Not the same as privilege.
Confidentiality	A duty not to disclose information shared in a protected relationship.	Not always an absolute bar to legal process.
Legal privilege	A rule that allows certain communications to be withheld from compelled disclosure.	Not a guarantee in every circumstance.
HIPAA compliance	A health-information regulatory framework governing the handling of protected health information.	Not automatic evidentiary privilege.

Foreword: The Container Problem

For more than twenty years, I have worked in roles that require building containers for human vulnerability. As a ritual designer, I create ceremonies—weddings, births, deaths, and the less easily named transitions between them—that hold people during moments when they are most open, most raw, and most in need of safety.

Alongside that work, I spent twenty-five years in the insurance industry as a commercial agent, underwriter, and consultant. My work involved analyzing risk structures—understanding where protection exists, where it does not, and what happens when people discover too late that the boundary they trusted was never actually there.

My relationship with computers also stretches back to the early days of modern computing. I began studying programming in the 1970s, working in languages such as BASIC, COBOL, and FORTRAN, and throughout my career, I have remained deeply comfortable with technology. In many of the organizations where I worked, I became the “accidental tech person”—the one who helped implement new systems, troubleshoot software, or test new programs. Technology has never felt foreign to me; it has been a natural part of how I understand and solve problems.

I have also worked on a sexual-assault crisis line and at a sexual-assault support center, where confidentiality is not simply a professional preference. It is the condition that allows someone to speak. Without it, many people remain silent.

Across these different fields—ritual work, crisis support, technology, and risk analysis—the same principle recurs. When people are in their most vulnerable moments, the container that holds the conversation matters as much as the conversation itself.

When I first began looking closely at how people were using AI systems, the pattern felt immediately familiar. In insurance, we look for structural gaps—places where people believe they are protected but the policy language quietly says otherwise. The emerging use of AI as a space for vulnerable disclosure has exactly that character. The expectation of safety exists. The legal architecture often does not.

Artificial intelligence has now created a new kind of conversational space. People use AI systems to process grief, anger, fear, and uncertainty, as well as to make difficult decisions. They explore questions they may not feel ready to ask another human being. They draft messages they are afraid to send. They work through situations they do not yet fully understand.

The interface feels private. The experience can feel intimate. The system responds without visible judgment and is available at three in the morning when other forms of support may not be.

But the structure surrounding those conversations is very different from the structures that traditionally protect vulnerable speech.

Conversations with therapists occur within clinical confidentiality. Conversations with attorneys may fall under attorney–client privilege. Crisis-line workers operate within strict confidentiality frameworks designed to ensure that callers can speak safely.

Conversations with consumer AI systems generally do not operate within those protections.

The result is a growing mismatch between how people experience these conversations and how they are actually treated within legal and institutional systems.

This paper examines that gap. It asks whether the legal and ethical frameworks that already protect certain kinds of human conversation might be extended—carefully and deliberately—into AI-mediated environments.

The question is not whether people will use AI to process vulnerable thoughts. They already do.

The question is whether the spaces where those conversations occur will remain structurally unprotected—or whether new containers can be built that are capable of holding them safely.

Author’s Perspective

The ideas in this paper come from the intersection of several different kinds of work I have done over the course of my career. I spent more than twenty-five years in insurance, financial services, and nonprofit administration, working with risk structures, regulatory systems, and organizational governance. I have also worked directly with vulnerable populations through crisis-line service and advocacy at a sexual-assault support center, where confidentiality is not an abstract legal concept but a practical condition that determines whether people will speak at all.

My relationship with technology runs just as deep. I began studying programming in the 1970s using early languages such as BASIC, COBOL, and FORTRAN, and throughout my professional life, I have often been the “accidental technologist” in organizations, helping implement, test, and troubleshoot evolving digital systems.

Alongside this professional work, I have spent decades studying traditions that explore how people process meaning, crisis, and life transitions, including ritual design and Andean shamanic practice.

Some of the situations discussed in this paper are not only subjects I have encountered professionally. They are also realities I have navigated personally at different points in my life. Those experiences inform my understanding of how essential truly confidential spaces can be when someone is trying to think through danger, identity, or escape.

Because of this combination of experiences—risk analysis, crisis support, spiritual traditions of meaning-making, and lifelong technical fluency—I tend to notice structural gaps where human vulnerability meets institutional or technological systems. The growing use of AI as a place where people process their most difficult thoughts appears to be one of those gaps.

This paper is an attempt to describe that gap and explore whether existing confidentiality frameworks might be extended to address it.

Introduction

Artificial intelligence has created a new kind of conversational space. Many people now use AI systems to process emotions, think through difficult decisions, or explore questions they may not feel ready to bring to another person.

These interactions often feel private and intimate. They frequently occur late at night, in solitude, during moments when people might otherwise have no one to speak with.

In reality, however, most consumer AI conversations do not carry the legal protections that traditionally govern sensitive disclosures to licensed professionals.

Communications with attorneys may be protected by attorney–client privilege. Communications with therapists may be protected by psychotherapist–patient confidentiality and privilege. These legal frameworks exist because society has long recognized that certain conversations must be protected to ensure people can seek help safely.

Conversations with consumer AI systems typically fall outside those protections.

Depending on platform policies and the circumstances of a particular case, conversations may be retained by providers, reviewed internally for safety purposes, or subject to disclosure through legal process.

That reality raises an important question.

If people are increasingly using AI systems as places to process legally and emotionally sensitive information, what would it take to create environments where those conversations receive protections comparable to those that exist in traditional professional relationships?

This paper explores that question.

It proposes a possible model: a new category of digital service that I refer to as a Privileged AI Sanctuary. In such systems, AI interactions would operate within the legal and ethical framework of licensed professionals—such as attorneys or mental health clinicians—so that conversations could fall within existing confidentiality and privilege structures.

This paper does not claim that the current law already recognizes AI systems as privileged agents. Instead, it examines how existing legal doctrines, professional practices, and technical infrastructure might support the development of such systems if they are carefully designed and tested.

The goal is not to replace human professionals. The goal is to ask whether the protections those professions provide—confidentiality, privilege, and safe space for difficult disclosure—can be extended into the kinds of AI-mediated conversations that millions of people are already having.

Part I: The Scope of the Problem

What People Are Actually Doing with AI

The available data suggest that AI chatbots are already being used as informal mental health resources. A nationally representative RAND study published in JAMA Network Open in 2025 found that approximately one in eight U.S. adolescents and young adults report using generative AI chatbots for mental-health advice. Among individuals aged 18 to 21, the proportion rises to roughly one in five.¹

Among those who use chatbots for this purpose, roughly two-thirds report doing so at least monthly, and more than ninety percent say they find the advice helpful. These are not casual interactions. They represent sustained, recurring conversations with systems that many users experience as safe places to think out loud.

A separate survey conducted through the Sentio Marriage and Family Therapy program found that a substantial proportion of people with self-reported mental-health conditions used language-model chatbots for emotional support. Researchers suggested that generative AI systems may already function as a large informal mental-health resource, while emphasizing that such systems are not substitutes for licensed care.²

Anthropic has also published research indicating that emotionally oriented interactions—including coaching, counseling-style conversations, companionship, and interpersonal advice—account for a measurable share of Claude use. Even a relatively small percentage of conversations in systems handling millions of interactions represents a very large volume of vulnerable communication occurring outside professional confidentiality frameworks.³

And the pattern extends well beyond mental health.

People use AI systems to research legal options during divorce, to think through conflicts with employers or business partners, to draft difficult messages, and to make sense of medical diagnoses or bureaucratic systems. Across these situations, the pattern is consistent: these are exactly the kinds of conversations that the legal system has historically recognized as requiring protection.

The People Who Need This Most

It is worth being specific about who is exposed, because the abstract language of “users” easily obscures the human reality. These are populations I have worked with directly across different chapters of my career, and I know from experience how much confidentiality can matter.

Survivors of Sexual Assault and Domestic Violence

From my time working at a sexual-assault support center and on a crisis line, I learned that the single most important factor in whether a survivor reaches out for help is whether they believe what they say will remain confidential.

It is not an abstract preference. For many people, it is a survival calculation.

Many survivors are still living with their abusers. Some remain financially dependent on them. Others fear law-enforcement involvement for complex reasons, including immigration status, prior legal history, or deep mistrust of institutions.

When someone calls a crisis line, the first thing they need to know is that speaking with a crisis line will not make their situation more dangerous.

AI chatbots are increasingly filling part of this role. They are available at three in the morning when crisis lines are understaffed, and they can be accessed quietly in situations where a phone call might be overheard. But the conversations themselves are not protected by the kinds of confidentiality structures that crisis services are built around.

People Navigating the Social Services System

Anyone who has worked with public benefits knows how labyrinthine these systems can be.

People increasingly turn to AI tools to understand eligibility rules, draft appeal letters, decode bureaucratic language, and figure out how different programs interact with one another. These conversations often involve disclosing detailed financial information, medical conditions, disability status, family circumstances, and sometimes immigration status.

In the wrong context, information like this can be used to deny benefits, trigger investigations, or undermine applications. The people who most need help understanding these systems are often the least able to absorb the consequences of unintended disclosure.

Crisis Intervention and Suicide Prevention

Some of the people having these conversations are in active crisis. They are typing thoughts they may not feel able to say out loud—to friends, family members, or anyone else.

During my crisis-line work, I learned that simply speaking the unspeakable can be therapeutic. It breaks the isolation that makes despair so dangerous.

AI systems now provide that outlet for many people. But the legal and privacy structures surrounding those conversations remain very different from the protections that govern traditional crisis services.

People Planning to Leave Dangerous Situations

When someone is planning to leave an abusive relationship, general advice is rarely enough. They need guidance that is specific to their situation.

Questions may include: How do I secure my documents? What are my legal rights regarding my children? How do I access a bank account my partner controls? Where can I go tonight?

AI systems can be surprisingly effective at structured problem-solving of this kind. But each question also creates a record of the person’s planning and intentions. If the wrong person accessed that information—or later surfaced in legal proceedings—it could expose exactly the strategy meant to keep them safe.

Individuals Researching Legal Exposure

This includes people facing potential litigation, regulatory investigation, employment disputes, tax questions, or immigration concerns.

During my years in the insurance industry—as an agent, underwriter, and later a consultant for a wealth-management firm—I sat with people across the full spectrum of financial and legal vulnerability.

One lesson emerged again and again: people will not disclose the full scope of their exposure unless they trust that what they share is protected.

In law, that trust is built through attorney–client privilege. In insurance, it is reinforced through professional relationships and policy structures. In consumer AI systems, those protections generally do not exist.

Organizations Serving Vulnerable Populations

Organizations that serve vulnerable populations—such as childcare centers, domestic-violence shelters, refugee resettlement programs, and addiction-recovery services—handle extraordinarily sensitive information about the people they serve.

From my experience serving as board president of a nonprofit childcare center, I understand the governance obligations attached to organizations responsible for protecting vulnerable individuals.

A caseworker is bound by professional ethics and organizational policy; an AI chatbot is governed primarily by a platform’s terms of service.

As organizations adopt AI tools to improve efficiency—helping with intake processes, case documentation, or resource matching—they may inadvertently introduce sensitive information into systems that operate under very different confidentiality expectations.

LGBTQ Youth Exploring Identity

For many LGBTQ youth, particularly in jurisdictions with restrictive laws or social environments, safe spaces for open exploration can be limited.

For some of these young people, AI chatbots have become one of the few places where they can ask questions about identity, relationships, or gender without fear of immediate judgment.

But the conversations themselves may still be stored on corporate systems and subject to platform policies that users do not fully understand.

Undocumented Individuals Seeking Legal Information

When a person without legal immigration status uses AI to research asylum processes or understand their rights during an immigration encounter, they are engaging in exactly the kind of legal exploration that attorney–client privilege was historically designed to protect.

Yet many of these individuals lack access to affordable legal counsel. AI tools may feel like the only available source of guidance. In that context, the gap between perceived confidentiality and actual legal protection becomes especially stark.

Communities with Histories of Institutional Distrust

Some communities carry long histories of justified mistrust toward institutional systems, including law enforcement, healthcare institutions, and government agencies.

For members of these communities, AI may appear to offer something institutions have often failed to provide: a space to ask questions without judgment or surveillance.

Whether that perception reflects reality is another question.

The point is that the demand for confidential spaces to think out loud is not hypothetical. It is already shaping how people interact with emerging technologies.

What AI Companies Have Disclosed

Several major AI companies have acknowledged that conversations with consumer chatbot systems may be reviewed, retained, or disclosed under certain circumstances. AI platform providers have publicly described safety systems designed to detect potentially harmful content within user interactions. These systems typically involve automated monitoring tools and, in some circumstances, human review of flagged conversations.⁴

These policies are not unusual within the technology sector. Many digital services operate under similar frameworks. However, the way AI systems are now being used—often as places where people process highly sensitive personal information—raises new questions about whether those frameworks remain appropriate.

The issue becomes especially important when users assume that emotionally or legally sensitive disclosures are protected in ways that ordinary consumer technology platforms were never designed to provide.

Part II: The Legal Architecture

Before turning to the legal analysis, one clarification is important. I am not an attorney, and this paper does not attempt to provide legal advice. Instead, it examines publicly available legal decisions and established legal doctrines in order to explore how existing confidentiality frameworks might apply to emerging AI technologies. The discussion that follows is therefore intended as policy analysis and conceptual exploration rather than legal guidance.

Legal Privilege and Consumer AI Systems

One of the central legal questions raised by the growing use of conversational AI concerns the status of communications shared with these systems.

In traditional legal practice, attorney–client privilege protects confidential communications between a client and their attorney that are made for the purpose of obtaining legal advice. The privilege exists because effective legal representation depends on clients being able to speak openly with their counsel without fear that their statements will later be used against them.

When individuals interact with consumer AI systems, however, those conversations typically occur outside the structure of a professional legal relationship. Most major AI platforms operate under terms and policies that permit data retention, automated safety monitoring, and human review of at least some flagged interactions. Because these systems function as third-party platforms rather than agents of a client’s attorney, communications shared with them may not fall within the traditional boundaries of attorney–client privilege.^4 5

Legal scholars and professional organizations have therefore begun advising caution when sensitive legal information is shared with consumer AI tools outside the supervision of legal counsel.⁵

The issue is not unique to artificial intelligence. Courts have long held that disclosure of otherwise confidential communications to third parties can undermine claims of privilege if those third parties are not acting as agents of the attorney within the scope of legal representation.⁶

This principle becomes particularly relevant when considering how AI systems might operate within a professional framework. If such tools were used under the direction of legal counsel, and if the system functioned as part of the attorney’s effort to understand and organize client information, the analysis could potentially be different.

This possibility leads directly to an existing legal doctrine that addresses the role of third-party assistants in legal representation.

The Kovel Doctrine and Its Application to AI

One possible legal framework for extending confidentiality protections to AI-mediated conversations comes from the doctrine established in United States v. Kovel, 296 F.2d 918 (2d Cir. 1961).⁷

In that case, the Second Circuit held that attorney–client privilege can extend to certain third parties who assist an attorney in providing legal advice. The court recognized that professionals such as accountants or translators may function as agents of the attorney when their expertise is necessary for the lawyer to understand a client’s information.

The original analogy offered by the court was that of a translator. If a client speaks a language the attorney does not understand, the use of an interpreter does not destroy privilege. Instead, the interpreter functions as a conduit through which the communication between client and attorney can occur.

Over time, courts have applied this reasoning to a range of professionals who assist attorneys in understanding complex information. These may include accountants, financial analysts, investigators, and technical consultants working under the direction of legal counsel.

Legal commentators often refer to these arrangements as “Kovel relationships.”

The central principle is that when a third party is engaged by an attorney to assist in providing legal advice—and when confidentiality is maintained—the communications involved in that assistance may fall within the scope of attorney–client privilege.

The discussion above highlights why this doctrine is relevant to the emerging role of AI tools.

If an AI system were engaged as part of a lawyer’s professional services—operating under the attorney’s direction and within a confidential environment—the system could potentially function as a technical intermediary assisting the lawyer in understanding client information.

That observation does not establish that AI systems currently qualify for such treatment under the law. The application of privilege doctrines to AI tools would likely require careful structuring of the professional relationship, contractual safeguards, and, ultimately, further legal testing.

Nevertheless, the Kovel doctrine suggests a possible conceptual pathway.

In such a model, the AI would not replace the attorney. Instead, it would act as a structured tool within the attorney’s professional workflow.

Whether courts would ultimately accept such an arrangement remains an open question. But the doctrine demonstrates that the law already recognizes situations in which non-lawyer intermediaries participate in privileged communications when they are necessary to facilitate the attorney’s work.

The concept of a Privileged AI Sanctuary builds on that possibility. It asks whether emerging technologies might be incorporated into existing professional confidentiality frameworks, rather than operating entirely outside them.

The Therapeutic Parallel: Confidentiality Under Clinical Oversight

Confidentiality protections in mental-health care provide another useful framework for thinking about how protected AI environments might operate.

Unlike attorney–client privilege, which developed primarily through common law, confidentiality in mental-health treatment is reinforced by both statutory protections and professional ethical obligations. In the United States, psychotherapist–patient privilege was formally recognized by the Supreme Court in Jaffee v. Redmond, 518 U.S. 1 (1996), which held that confidential communications between a psychotherapist and patient are protected from compelled disclosure under federal evidentiary law.⁸

The Court’s reasoning emphasized that effective psychotherapy depends upon an atmosphere of confidence and trust, and that the possibility of disclosure could discourage patients from speaking openly.

In the decades since that decision, the delivery of mental-health services has increasingly expanded into digital environments. Telehealth platforms now routinely provide therapy sessions through secure video systems while maintaining the same confidentiality protections that apply to in-person treatment.

These systems operate under regulatory frameworks such as the Health Insurance Portability and Accountability Act (HIPAA), which governs how protected health information is stored, transmitted, and accessed. Technology providers that support clinical services typically operate under Business Associate Agreements (BAAs), contractual arrangements that legally bind the technology provider to safeguard protected health information on behalf of the clinician or organization.⁹

Within these environments, various forms of AI-assisted tools are already in use. AI systems are increasingly used to generate session summaries, assist with clinical documentation, and help organize treatment notes. These tools process sensitive therapeutic information while operating inside a regulated confidentiality framework.

The key point is that the presence of technology does not automatically remove confidentiality protections. Instead, the legal and ethical protections attach to the professional relationship itself.

A similar principle could potentially apply to AI-assisted therapeutic environments designed for patient use outside of formal sessions. If such systems operated under the supervision of a licensed clinician and within a HIPAA-compliant infrastructure, interactions with the system might reasonably be considered part of the broader therapeutic process.

Such an arrangement would not replace therapy sessions or clinical judgment. Rather, it could function as a structured extension of the therapeutic relationship, providing guided reflection, journaling, or emotional processing tools that remain within the clinician’s protected environment.

As with the legal framework discussed earlier, the precise legal status of such arrangements would depend on regulatory interpretation and professional oversight. However, the existence of AI-assisted tools within HIPAA-regulated clinical environments demonstrates that technology can be integrated into confidentiality frameworks when the appropriate governance structures are in place.

The broader question raised by this paper is whether similar principles might be applied to AI systems designed specifically to support vulnerable conversations.

Foreseeable Legal Challenges

Any proposal to extend professional confidentiality frameworks into AI-mediated environments raises important legal and ethical questions. The following issues would likely require careful consideration in the design and implementation of any privileged AI system.

Privilege Waiver Through Data Handling

Attorney–client privilege depends heavily on the preservation of confidentiality. If privileged communications are shared with third parties outside the protected professional relationship, courts may determine that the privilege has been waived.

For any AI system seeking to operate within a privilege framework, data architecture would therefore be critical. Conversations would likely need to be isolated from model training systems, protected through encryption, and governed by contractual agreements that strictly limit access to the information.

In other words, the system’s technical design would need to reflect the same confidentiality expectations that exist within traditional attorney–client communications.

Professional Responsibility and Liability

A second challenge concerns professional responsibility.

If licensed attorneys or clinicians oversee AI-assisted interactions, they may assume responsibility for ensuring that the system operates safely and appropriately within their professional obligations.

This raises questions about professional liability and standards of care. For example, how should professionals supervise AI-mediated interactions? What safeguards are necessary to prevent misuse or misunderstanding of the system?

Similar questions arose during the early development of telehealth services. Over time, professional guidelines, regulatory standards, and specialized insurance products evolved to address those risks.

A comparable process would likely be necessary for AI-assisted professional environments.

Mandatory Reporting Obligations

Professional confidentiality is not absolute.

Attorneys, therapists, and other licensed professionals operate under mandatory reporting obligations in specific circumstances. These may include situations involving imminent threats of harm, child abuse, elder abuse, or other legally defined risks.

Any AI system operating under professional oversight would need to incorporate these obligations into its operational framework. Users would need clear notice that certain types of disclosures may trigger mandatory reporting requirements, just as they do in traditional professional relationships.

Jurisdictional Variation

Confidentiality and privilege rules vary across jurisdictions.

Attorney–client privilege, psychotherapist–patient privilege, and mandatory reporting obligations may differ by state and by country. A system designed to operate across jurisdictions would therefore need to account for these differences through appropriate licensing structures, professional partnerships, and legal guidance.

This complexity is not unique to AI systems. Telehealth platforms already navigate similar jurisdictional challenges when providing services across state or national boundaries.

International Considerations

Although this paper focuses primarily on the legal frameworks of the United States, the confidentiality gap surrounding AI-mediated conversations is not limited to one country. Similar questions arise wherever people use AI systems to process sensitive personal information.

Different jurisdictions approach data protection and confidentiality in distinct ways, and some regions already maintain stronger baseline protections than those found in the United States.

The European Union provides one of the most developed regulatory frameworks in this area. The General Data Protection Regulation (GDPR) establishes strict requirements regarding personal data collection, storage, and use. These include principles such as data minimization, purpose limitation, and the right of individuals to request deletion of their data.¹⁰

In addition, the European Union’s Artificial Intelligence Act—adopted in 2024—introduces regulatory oversight for certain categories of high-risk AI systems, including some applications within healthcare and other sensitive contexts.¹¹

These frameworks do not automatically create confidentiality protections equivalent to attorney–client privilege or therapist–patient privilege. However, they demonstrate that many jurisdictions already recognize the need for stronger governance of sensitive digital information.

In other regions—including Latin America, the Asia–Pacific region, and the Middle East—data protection laws vary widely. Some countries have adopted regulations modeled in part on GDPR, while others continue to rely on sector-specific privacy rules.

Any system designed to provide privileged AI interactions across borders would therefore need to address a complex landscape of national regulations, professional licensing requirements, and data governance standards.

Despite these differences, the underlying issue remains remarkably consistent. Across cultures and legal systems, people increasingly use digital tools to explore questions involving identity, safety, relationships, health, and legal risk.

The need for protected spaces in which such conversations can occur safely is not unique to any one jurisdiction.

If AI systems are becoming places where vulnerable conversations occur, the question of how those conversations are governed—and whether they can be protected within existing legal frameworks—is likely to arise in many countries.

Part III: Product Architecture and Market Opportunity

Competitive Landscape: What Exists and What Doesn’t

A credible proposal must begin with an honest assessment of existing tools. Several adjacent technologies already address pieces of the problem discussed in this paper. However, none currently combine AI-mediated interaction with legally enforceable confidentiality protections for individual users.

Three categories of products are particularly relevant.

HIPAA-Compliant AI Tools for Clinicians

A growing number of companies provide AI tools designed to assist licensed mental-health professionals. These systems may generate documentation, summarize sessions, or assist with administrative tasks. They demonstrate that AI systems can operate within regulated confidentiality frameworks, but they are designed primarily as tools for clinicians, not as protected environments for patients themselves.

The confidentiality umbrella covers the clinician’s workflow, not the user’s vulnerable moment.

AI Legal Research Platforms

Legal-technology companies have also begun integrating AI into professional workflows. Platforms such as Thomson Reuters’ CoCounsel provide AI-assisted research tools used by licensed attorneys inside secure professional environments.¹²

Again, the pattern is similar: the technology exists, but access is limited to professionals operating within established legal structures. The individual attempting to understand a legal problem outside of a formal attorney relationship has no comparable protected environment.

Digital Therapy Platforms

Online therapy platforms demonstrate that large numbers of people are willing to seek emotional support through digital interfaces. These services operate within clinical frameworks that preserve therapist–patient confidentiality. But they rely on human-to-human interaction delivered through technology, not AI-mediated conversation, and their pricing reflects the cost of professional labor.

The Gap

Taken together, these existing technologies reveal a consistent pattern.

Professional environments—legal and clinical—already integrate AI tools under confidentiality frameworks. Consumer AI systems, by contrast, operate almost entirely outside those frameworks.

As a result, people increasingly conduct sensitive conversations in systems that were never designed to protect them.

The gap is not technological. The underlying AI capabilities already exist.

The gap is architectural.

No widely available system currently combines AI-mediated conversation, direct access for individuals, and legally enforceable professional confidentiality. A system designed around those three elements would represent a new category of digital service.

Timing: Why This Issue Is Emerging Now

Several trends are converging simultaneously.

First, AI systems are becoming integrated into everyday problem-solving. People increasingly use conversational AI not only for information retrieval but also for reflection, planning, and emotional processing.

Second, public awareness of data governance issues surrounding AI is growing. Professional guidance from legal and clinical bodies has begun to clarify that conversations with consumer AI systems do not automatically carry the protections people may assume.

Third, regulatory and professional communities are beginning to grapple more directly with the implications of AI within confidential professions such as law and medicine.

These developments suggest that the confidentiality gap identified in this paper may become increasingly visible.

If the demand for protected digital spaces continues to grow—as current usage patterns suggest it will—solutions that address this gap are likely to emerge.

The question is not whether such systems will eventually be developed, but how they will be designed and governed.

Accessibility and Public-Interest Considerations

Any system designed to provide privileged AI environments must also consider the populations most likely to benefit from them.

Many of the individuals described earlier in this paper—survivors of abuse, young people exploring identity, individuals navigating social services, or people seeking crisis support—may not have the financial resources to subscribe to premium digital services.

For that reason, accessibility mechanisms should be considered alongside commercial deployment.

Public access points are one possibility. Libraries, community centers, and nonprofit organizations have long served as access points for digital services. Secure terminals or supervised access programs could allow individuals to use protected AI environments without requiring personal subscriptions.

Partnerships with crisis services are another possibility. Crisis lines, domestic-violence organizations, legal-aid groups, and community mental-health programs could integrate protected AI systems into their support infrastructure.

Sliding-scale or sponsored access would also matter. Commercial providers could allocate a portion of system capacity for subsidized use, supported through partnerships with philanthropic organizations, legal-aid groups, or public-interest programs.

Systems designed to support vulnerable conversations occupy a space that has traditionally been served by institutions such as therapists’ offices, legal consultations, and crisis services. As a result, questions of governance, confidentiality, and accessibility are likely to be important considerations in their development.

Conceptual Service Models

The concept of a Privileged AI Sanctuary could take multiple forms depending on the professional framework through which confidentiality is established. Two models illustrate how existing legal and professional structures might support such systems.

These models are not mutually exclusive, and hybrid arrangements could eventually emerge. They are presented here simply to demonstrate that the underlying concept can operate within existing professional governance frameworks.

Legal Sanctuary Model

In the legal model, confidentiality protections would arise through the attorney–client relationship.

The user’s formal relationship would be with a licensed attorney or law practice that provides access to an AI-assisted consultation environment. Within this structure, the AI system would function as a tool operating under the direction of legal counsel, similar to other professional support tools used in legal practice.

The system could assist users in organizing information, identifying relevant legal questions, and exploring possible courses of action before or during consultation with an attorney.

Potential uses might include understanding legal options during divorce or custody disputes, exploring employment conflicts or whistleblower concerns, preparing information for consultation with legal counsel, organizing documentation related to regulatory or financial disputes, and identifying questions to raise with an attorney.

Under this model, the AI system would not provide independent legal advice. Rather, it would function as a structured interface that helps users articulate their situation and prepare for professional guidance.

The key distinction is that the interaction would occur within the professional relationship, rather than outside it.

Therapeutic Sanctuary Model

A parallel model could operate within the framework of licensed mental-health care.

In this structure, the user’s formal relationship would be with a licensed clinician who supervises an AI-assisted environment designed to support reflection and emotional processing between sessions.

The AI system could provide guided journaling, structured self-reflection tools, mood tracking, or prompts designed to help individuals articulate experiences they may wish to discuss with their therapist.

Potential uses might include processing grief, stress, or major life transitions, preparing for therapy sessions, exploring difficult conversations or decisions, reflecting on patterns in relationships or emotional responses, and structured journaling for trauma-informed therapy.

As with the legal model, the AI system would not replace the licensed professional. Instead, it would function as an extension of the therapeutic environment, operating within the confidentiality protections already established by the clinical relationship.

Integrated Sanctuary Environments

Over time, more integrated models could emerge that address situations involving both emotional and legal complexity.

Many real-world crises—such as domestic violence, divorce, immigration challenges, or workplace misconduct—contain both legal and psychological dimensions.

An integrated sanctuary model might allow individuals to access both therapeutic and legal support within coordinated confidentiality frameworks. Such systems would require careful governance, but they could potentially provide a more holistic approach to situations where legal exposure and emotional vulnerability intersect.

Economic Structure and Sustainability

Any system designed to provide professionally supervised AI environments would need a sustainable economic model. The presence of licensed professionals, regulatory compliance requirements, and secure technical infrastructure would inevitably create operating costs that exceed those of standard consumer AI services.

However, several factors suggest that viable models could emerge.

First, there is clear evidence that people are willing to pay for confidential professional environments when dealing with sensitive issues. Legal consultations, therapy sessions, and other forms of professional guidance already operate within this economic structure.

Second, AI-assisted systems may allow certain aspects of professional interaction to scale more efficiently. Structured tools that help individuals organize information, prepare questions, or process complex situations could reduce the amount of direct professional time required for each case.

Rather than replacing professional services, such systems could function as structured preparation environments, helping individuals engage more effectively with attorneys, therapists, or other professionals.

A range of service models could potentially support this approach.

Subscription-Based Professional Access

One possible model would involve subscription access to AI-assisted environments operating under professional supervision. In such systems, users could engage with structured tools designed to help them organize their situation, reflect on options, and prepare for professional consultation.

Licensed professionals would oversee the system’s design and governance, ensuring that interactions remain within appropriate ethical and legal boundaries.

Institutional Licensing

Another model would involve institutional licensing. Law firms, therapy practices, nonprofit organizations, or advocacy groups could provide access to protected AI environments as part of their services. In this structure, the professional organization—not the individual user—would maintain the formal relationship that supports confidentiality protections.

This model could allow organizations already serving vulnerable populations to integrate AI tools while maintaining professional oversight.

Public and Nonprofit Access

Because many of the individuals who might benefit most from protected AI environments have limited financial resources, mechanisms for broader access would be essential.

Public institutions such as libraries have long played a role in providing access to information technologies. Secure access points within libraries, community centers, or nonprofit organizations could allow individuals to use protected AI environments without requiring personal subscriptions.

Similarly, crisis services, domestic-violence organizations, legal-aid groups, and community mental-health programs could integrate such systems into their support infrastructure.

Partnerships with philanthropic organizations or public-interest technology initiatives could help subsidize access for individuals who might otherwise be excluded.

Systems designed to support vulnerable conversations occupy a space that has traditionally been served by institutions such as therapists’ offices, legal consultations, and crisis services. As a result, questions of governance, confidentiality, and accessibility are likely to be important considerations in their development.

Hybrid Models

In practice, a combination of these approaches would likely emerge. Commercial services could support the development and maintenance of the technology, while institutional partnerships and public-access programs could extend the system's benefits to populations with fewer resources.

This hybrid approach reflects patterns already seen in telehealth, legal-aid services, and other technology-assisted professional systems.

Implementation Pathways

The concept of privileged AI environments described in this paper would require collaboration across several professional and technical domains. While the precise structure of such systems would depend on regulatory interpretation and professional guidance, several practical pathways for development can be envisioned.

Pilot Programs

Early implementations could begin through limited pilot programs operated by licensed professionals or organizations already working within confidentiality frameworks.

For example, a law firm or legal-aid organization might explore an AI-assisted intake and preparation environment operating under attorney supervision. Similarly, mental-health practices or crisis-support organizations could test structured AI reflection tools designed to support clients between sessions.

These pilot programs would allow developers and professionals to evaluate how AI systems function within existing confidentiality structures while identifying potential risks and governance requirements.

Professional Governance

Any system designed to operate within privilege or confidentiality frameworks would require clear professional oversight.

Attorneys, clinicians, and other licensed professionals would need to establish guidelines governing how AI tools are used within their practice environments. Professional organizations and regulatory bodies might eventually develop standards addressing issues such as supervision, documentation, and ethical boundaries.

This type of professional governance has historically accompanied the introduction of new technologies into regulated fields.

Technical Infrastructure

From a technical perspective, the development of privileged AI environments would require careful attention to data architecture.

Systems designed for confidential use would likely need to ensure that user interactions are isolated from model training pipelines, sensitive information is stored within secure and encrypted environments, access to data is limited to authorized professional contexts, and clear audit and governance mechanisms exist for oversight.

Many of these requirements are already familiar within healthcare and legal technology environments.

Legal and Regulatory Testing

Over time, the legal status of AI-mediated privileged environments would likely require clarification through professional guidance, regulatory interpretation, or case law.

Pilot implementations could provide an opportunity to evaluate how existing legal doctrines—such as the Kovel framework discussed earlier—apply in practice when AI systems operate under professional supervision.

Rather than attempting to resolve these questions in advance, carefully structured implementations could allow the law to evolve alongside the technology.

Part IV: The Deeper Architecture

Why Protected Space Is Not a Luxury

At this point it is useful to step outside the language of markets, technology, and case law for a moment.

Across cultures and throughout history, societies have created structures that protect certain kinds of speech. These spaces appear in many forms: the therapist’s office, the attorney’s study, the confessional, the ritual circle, and the deathbed conversation that is understood—without needing to be said—as private and sacred.

These spaces are not historical curiosities or social luxuries. They are structural supports within functioning societies.

Without environments where people can speak honestly about fear, conflict, wrongdoing, identity, grief, or danger, many problems cannot be addressed at all. The ability to disclose difficult truths safely is often the first step toward resolving them.

Artificial intelligence has unintentionally created something that resembles such a space.

Conversational AI systems respond without visible judgment. They are available at any hour. They allow people to articulate thoughts privately before sharing them with another person—or sometimes instead of doing so.

For many individuals, these systems already function as places to process situations that feel too complex, too frightening, or too uncertain to discuss elsewhere.

Yet the structures surrounding these conversations are fundamentally different from those that traditionally protect vulnerable speech.

Conversations with attorneys may fall under attorney–client privilege. Conversations with therapists are protected by clinical confidentiality. Crisis-line workers operate within carefully designed frameworks intended to ensure that callers can speak freely.

Conversations with consumer AI systems generally do not operate within comparable protections.

The experience may feel private. The legal and institutional architecture often says otherwise.

The question raised by this paper is therefore not whether people should use AI to process difficult thoughts. They already do.

The question is whether the systems in which those conversations occur will remain structurally unprotected, or whether new forms of digital space can be designed that are capable of holding them safely.

Throughout my career—in insurance, crisis work, ritual practice, and organizational governance—I have repeatedly encountered the same principle: when people are navigating their most vulnerable moments, the container matters as much as the conversation itself.

The emergence of conversational AI has created a new kind of container.

Whether that container will ultimately function as a safe space, or as a system that quietly exposes the very conversations people believe are private, remains an open question.

Part V: Invitation to Collaborate

Invitation to Collaborate

Technological innovation and economic enterprise are essential drivers of progress. Companies build the tools that shape modern life, and entrepreneurs take the risks that allow new systems to emerge. At the same time, societies have long recognized the need for protected spaces where individuals can safely confront vulnerability, conflict, and change. The challenge explored in this paper lies in balancing these realities: encouraging innovation while ensuring that the environments people increasingly rely on for their most difficult conversations are thoughtfully designed.

This paper is intended as an exploration rather than a finished proposal. The concept of privileged AI environments raises questions that cannot be answered by any single discipline.

Developing systems capable of supporting confidential AI-mediated conversations would require collaboration across several fields.

Technologists would be needed to design secure architectures capable of protecting sensitive interactions while maintaining the functionality people expect from conversational AI systems.

Legal scholars and practicing attorneys would need to examine how existing privilege doctrines might apply to such environments and whether new professional standards or regulatory interpretations would be required.

Mental-health professionals and crisis-service providers could help ensure that systems designed to support vulnerable conversations operate safely and ethically within established therapeutic frameworks.

Policy researchers and institutional leaders would likely play a role in considering governance structures, accessibility, and the broader societal implications of such systems.

The purpose of this paper is simply to suggest that a gap exists between the ways people are already using AI systems and the legal and institutional structures that currently govern those interactions.

If that gap is real, then the question of how to address it will eventually arise.

The hope behind this paper is that the question can be approached thoughtfully, before technological momentum alone determines the outcome.

The systems people use to process their most vulnerable conversations inevitably shape how those conversations unfold. As AI becomes more integrated into everyday life, the design of those systems—and the protections surrounding them—may matter more than we currently realize.

About the Author

Karen Knutson is the founder of StarWay Transitions, a practice focused on designing ceremonies and structured experiences for major life transitions. Her work explores how different cultural and professional systems create containers for vulnerable human conversations.

She brings more than twenty-five years of experience in risk analysis and organizational management, including roles in commercial insurance underwriting, financial advisory services, nonprofit administration, and regulatory compliance. She has worked directly with vulnerable populations through crisis-line service and medical and legal advocacy at a sexual-assault support center, and has served as board president for a nonprofit childcare organization.

Karen’s relationship with technology spans the modern computing era. She began studying computer science in the 1970s, working with early programming languages including BASIC, COBOL, and FORTRAN, and developing programs to solve practical problems. Throughout her career she has remained deeply comfortable with technology, frequently serving as the “accidental tech expert” in small organizations and participating in testing and implementation of new software systems in larger ones.

Alongside her professional career, she has spent decades studying traditions that explore human meaning-making and emotional processing, including ritual design, astrology, numerology, Andean shamanic practice, and energy-healing disciplines. These experiences inform her interest in how societies create safe spaces for difficult conversations.

Her work focuses on the intersection of risk management, confidentiality structures, human vulnerability, and emerging technologies. She is based in Oregon.

Endnotes

1. McBain, Rebecca K., Robert Bozick, Melissa Diliberti, et al. “Use of Generative AI for Mental Health Advice Among US Adolescents and Young Adults.” JAMA Network Open 8, no. 11 (2025): e2542281. The study found that 13.1% of respondents reported using generative AI for mental-health advice; among those aged 18–21 the rate was 22.2%, and among users 65.5% sought advice monthly or more often while 92.7% found the advice somewhat or very helpful.

2. Rousmaniere, T., Y. Zhang, X. Li, and S. Shah. “Large Language Models as Mental Health Resources: Patterns of Use in the United States.” Practice Innovations, advance online publication, July 7, 2025. doi:10.1037/pri0000292.

3. Anthropic, “How People Use Claude for Support, Advice, and Companionship,” June 27, 2025, and accompanying appendix. Anthropic reported that affective conversations accounted for 2.9% of Claude.ai interactions in its sample.

4. OpenAI, “Helping People When They Need It Most,” August 26, 2025; OpenAI, “Usage Policies,” updated October 29, 2025; OpenAI Help Center, “Data Usage for Consumer Services FAQ” and “Chat and File Retention Policies in ChatGPT,” 2025–2026. These materials describe automated monitoring tools, human review of certain flagged interactions, and retention policies for consumer services.

5. American Bar Association, Formal Opinion 512, “Generative Artificial Intelligence Tools,” July 29, 2024.

6. Upjohn Co. v. United States, 449 U.S. 383 (1981).

7. United States v. Kovel, 296 F.2d 918 (2d Cir. 1961).

8. Jaffee v. Redmond, 518 U.S. 1 (1996).

9. U.S. Department of Health and Human Services, “Health Information Privacy,” including HIPAA Privacy Rule materials; HHS, “Business Associates” and “Sample Business Associate Agreement Provisions.”

10. European Parliament and Council of the European Union, Regulation (EU) 2016/679 (General Data Protection Regulation).

11. European Parliament and Council of the European Union, Regulation (EU) 2024/1689 (Artificial Intelligence Act).

12. Thomson Reuters, “One Million Professionals Turn to CoCounsel as Thomson Reuters Scales AI for Regulated Industries,” press release, February 24, 2026.