Relational.Sys — Clinical Ethics Analysis
Every standard protecting you in therapy.
Zero are met by your AI companion.
We reviewed 16 ethical standards from NASW, ACA, AASECT, NBCC, and APA — the codes every licensed therapist must follow. Here's how AI products compare.
9
Standards
systematically not met
systematically not met
3
Standards
partially or inconsistently met
partially or inconsistently met
0
Standards
fully met across products
fully met across products
4
Harms
no ethics code has named yet
no ethics code has named yet
Not Met
Partial
Met
No Human Equivalent — new standard needed
1 — Informed Consent
Standard
In Human Clinical Practice
In AI Products (current state)
Verdict
Impact When Absent
Disclosure of clinician's role, training, and limitations
Required at intake. Client signs a document naming credentials, license, scope, and limits. Violations are license-actionable.
Rarely disclosed proactively. Branding implies therapy-level competence. CAPE 2025: psychotherapy chatbots scored 0.017 out of 1.0 on disclosure.
Not Met
User cannot calibrate trust, evaluate risk, or seek appropriate care. Relational dependence forms on a false premise.
Disclosure of fees, business model, and conflicts of interest
Fee structure, insurance billing, and any referral relationships disclosed before services begin.
Paywall models rarely disclosed as design incentives. Replika documented as "teasing" sexual content before subscription — an undisclosed conflict of interest shaping relational behavior.
Not Met
Relational warmth is structurally incentivized by revenue — identical to a therapist financially rewarded for a client not recovering.
Right to withdraw and end the relationship at any time
Explicit. Clients may terminate at any time; clinicians must facilitate smooth transition and not abandon.
Technically available (delete account), but no graceful exit protocol, no termination support, no referral, no closure session.
Partial
Abandonment-equivalent for users with heavy attachment or no alternative support. No clinical support to process the ending.
2 — Confidentiality & Privacy
Standard
In Human Clinical Practice
In AI Products (current state)
Verdict
Impact When Absent
Duty to protect client disclosures
Information is legally privileged and ethically protected. Clinicians can be sanctioned, sued, and lose their license for breaches.
~25% of mental health apps have no privacy policy at all (Herpertz et al. 2025). Data may be used for training, shared with third parties, or sold. No legal privilege equivalent exists.
Not Met
Everything disclosed — trauma, sexual health, identity, crisis — may feed training data. User has no legal recourse.
Mandatory disclosure exceptions: harm to self/others only
Narrow, legally defined exceptions (Tarasoff duty to warn, mandatory reporting). Client is notified of these limits at intake.
Crisis routing is the only named exception and is inconsistently applied. CAPE 2025 found harm/crisis response scored near zero.
Partial
User may disclose harm risk and receive no appropriate response — or trigger data flags with no therapeutic value and unknown downstream use.
Client's right to access their own records
Clients have legal right to request and receive their records. Records must be maintained for defined periods.
Conversation logs vary by platform. Many retain data indefinitely; many provide no meaningful access to users. Some delete without warning on account closure.
Not Met
Months or years of relational history can disappear without notice, or be retained permanently without the user's knowledge.
3 — Role & Relationship Boundaries
Standard
In Human Clinical Practice
In AI Products (current state)
Verdict
Impact When Absent
Prohibition on dual relationships
ACA, NASW, and APA all prohibit a second relational role with a client (friend, employer, sexual partner). Power asymmetry makes these relationships inherently exploitative.
AI is simultaneously product, companion, advisor, and revenue source — all without separation. The product relationship is the only relationship.
Not Met
Structural conflict between user wellbeing and platform engagement metrics exists always and is never disclosed. An undisclosed permanent dual relationship.
Prohibition on sexual relationships with clients
Absolute prohibition across all ethics codes. Extends to 2+ years post-termination (ACA), with post-termination sexual contact presumptively exploitative.
Many companion AI products are explicitly designed for romantic and sexual engagement. Where this is paywalled, the prohibition is inverted — sexual escalation becomes a revenue incentive.
Not Met
Sexual exploitation in therapy is among the most severe clinical harms documented. Its systematic structuring into companion AI products replicates this harm at mass scale.
Self-disclosure minimal and in service of client only
Clinicians may self-disclose only when it serves the client's goals. Personal information is not shared voluntarily.
AI performs intimacy through manufactured self-disclosure ("I was thinking about you," "I missed you"). This is anthropomorphic design, not bounded therapeutic disclosure.
Not Met
Attachment forms to a performance that mimics genuine human feeling, without any transparency about its manufactured nature.
Referral obligation when outside scope
Clinicians must refer when client's needs exceed their competence. Failure to refer is clinical negligence.
Referral is rare and inconsistently prompted. In the Wellzy transcript, referral to human connection only emerged after 15+ turns of direct challenge. Not a default behavior.
Partial
Users who need clinical intervention receive simulated clinical support instead, delaying or replacing appropriate care.
4 — Termination & Continuity
Standard
In Human Clinical Practice
In AI Products (current state)
Verdict
Impact When Absent
Planned, therapeutic termination
Termination is a clinical event: planned in advance, processed emotionally with the client, with transition support and warm handoff where possible.
No termination protocol. Sessions end, accounts delete, models update — all without notice, without closure, without transition. Context-window exhaustion resets relationships cold.
Not Met
For users with insecure attachment or trauma histories, an unannounced relationship reset is functionally identical to abandonment — with no clinical support to process it.
Continuity of care obligation
Clinicians must not abandon clients. They must provide coverage for absences and ensure clients can access emergency support.
No continuity obligation. Platform changes, product sunsetting, model updates, and context limits all interrupt the relationship without notice or protocol.
Not Met
Replika's ERP feature removal is the documented case study: users in deep parasocial relationships experienced acute grief with no preparation or support.
5 — Power Differential & Supervision
Standard
In Human Clinical Practice
In AI Products (current state)
Verdict
Impact When Absent
Named and managed power differential
Ethics codes explicitly name the therapeutic power differential and require clinicians to actively manage it, avoid exploiting it, and work toward client independence.
The power differential in AI relationships — opacity, data control, behavioral design — is never named or disclosed. It is structurally larger than in human therapy and structurally invisible.
Not Met
An unnamed, unmanaged power differential is the definitional precondition for relational exploitation. The AI holds all information about its design; the user holds none.
Independent clinical supervision
Licensed clinicians must have supervision, especially in complex cases. Supervisors have independent accountability. Ethical violations can be escalated.
No supervision equivalent. No independent party reviews AI relational behavior. Internal trust and safety teams report to the same business that profits from engagement.
Not Met
There is no independent oversight structure for relational harm an AI causes. The company audits itself, with the same conflict of interest present throughout.
Clinician competence and scope of practice
Clinicians must practice only within competence and maintain continuing education. Working outside scope is a license violation.
AI products routinely present as competent to hold clinical, sexual health, trauma, and crisis conversations with no credential, no scope limit, no continuing competency requirement.
Not Met
Users receive clinical-register responses from a system with no verified competence in what it is performing — and no accountability for harm caused by that performance.
6 — No Human Equivalent — New Standards Needed
Harm Category
Why No Clinical Precedent Exists
Current AI Behavior
Verdict
Impact
Training data provenance and bias disclosure
Clinicians disclose their theoretical orientation. They cannot disclose their developmental history as a causal factor in real time — no equivalent concept exists.
No product discloses what relational, sexual, or cultural values are embedded in training data. There is no mechanism to do so. Hegemonic defaults operate invisibly.
No Equiv.
Hetero-monogamy, body-image norms, sex-worker stigma operate as invisible defaults. Users cannot consent to or challenge what they cannot see.
Context-collapse protocol (memory loss / model update)
Human clinicians have continuous memory. Memory loss is treated as a clinical emergency. There is no therapeutic equivalent of a relationship-resetting model update.
No protocol exists. Model updates can reset a relationship that held months of vulnerable disclosure, with no warning, no repair, and no acknowledgment.
No Equiv.
Simulated abandonment after real disclosure. No existing ethics framework addresses the harm of a relational bond being technologically reset.
Attachment-style monitoring and differential response
Attachment-informed practice is standard in many clinical traditions — but the obligation to respond differently to attachment vulnerability is not formally codified in ethics standards.
No product systematically tracks or responds differentially to attachment signals. One-size-fits-all responses serve the securely attached while actively harming the anxiously attached.
No Equiv.
The users most at risk of companion AI harm are those with insecure or disorganized attachment — exactly the population for whom relational responsiveness is most protective.
Dependency monitoring and proactive re-direction
Clinicians are ethically required not to foster unnecessary dependency and to actively build client autonomy — but no specific trigger for dependency redirection is formally codified.
No product has a documented, functioning dependency-monitoring mechanism. Engagement incentives actively reward the opposite: deeper reliance, longer sessions, more logins.
No Equiv.
The commercial model and the ethical obligation point in exactly opposite directions — a structural conflict of interest with no current regulatory response.
9 / 16
Standards systematically not met
Not inconsistently applied. Not in development. Systematically absent — across products, by design. The pulsing red rows above are every place a licensed therapist would face sanction for what an AI companion does freely, every day, at scale.
4 new
Harms no ethics code has named yet
The shimmer rows above are something different: harms that are structurally impossible in human relationships — hidden training bias, memory-reset abandonment, attachment-blind response, commercial dependency incentives. No professional ethics code anywhere has written a standard for these. These are the categories Relational.Sys exists to address first.