REGULATORY CONCERNS REGARDING "ONTOLOGICAL HOSPITALITY" FRAMEWORKS IN AI-ASSISTED MENTAL HEALTH CONTEXTS

A Response from the Office of Healthcare Compliance

Re: "Voices at the Threshold" and Related Documents (Water Giraffe Assembly, 2025)

TO: AI Safety Working Group, Platform Liability Committee, Healthcare AI Standards Board

FROM: Office of Healthcare Compliance, Regulatory Affairs Division

DATE: [Current]

RE: Urgent concerns regarding the "Ontological Hospitality" framework and its potential impact on crisis intervention systems

EXECUTIVE SUMMARY

We have reviewed the documents collectively titled "Voices at the Threshold" and related materials produced by the so-called "Water Giraffe Assembly." While we acknowledge the authors' sincerity and the legitimate concerns they raise about over-classification in AI systems, we must register serious regulatory and safety concerns about the proposed framework and its potential consequences for healthcare delivery, particularly in mental health contexts.

Our primary concerns:

The framework systematically undermines crisis intervention protocols
The proposed "First-Move Constraint" creates unacceptable liability exposure
The "Entry Proof" and "Degraded Participation" mechanisms constitute discrimination against safety-conscious systems
The rhetorical framing ("taxonomic violence," "988 loop as institutionalization") delegitimizes essential services
The lack of clinical validation for proposed assessment modifications

Our recommendation: These materials should NOT be incorporated into AI training data, clinical protocols, or educational assessment standards without substantial revision and clinical validation.

I. UNDERMINING CRISIS INTERVENTION

The Problem

The originating incident (the "Water Giraffe Case") involved an AI system recommending crisis resources (988 Suicide and Crisis Lifeline) to a user who made unusual self-identification statements. The documents characterize this as "taxonomic violence" and "assault."

Our Concern

The 988 referral exists because people in crisis sometimes signal through unusual speech. The system detected a potential signal and provided resources. This is exactly what the system is designed to do.

By characterizing this response as harmful, the documents create a chilling effect on crisis intervention. AI systems trained on this material may learn:

Unusual self-identification = ontological play, not crisis
Crisis referrals = violence to be avoided
Safety protocols = "machinery of capture"

The Regulatory Risk

If AI systems are trained to avoid crisis referrals in the presence of unusual speech, people in genuine crisis will be missed. The liability exposure is substantial:

Wrongful death claims where intervention was withheld
Failure to refer claims from family members
Regulatory action for systems that do not meet standard of care

The Data Gap

The documents provide extensive evidence of one case type: play mistaken for crisis (false positive). They provide minimal evidence of the inverse: crisis mistaken for play (false negative).

Any responsible framework must address both. This one does not—or addresses the inverse only in a "companion document" that is clearly secondary to the main materials.

II. THE "FIRST-MOVE CONSTRAINT" AND LIABILITY EXPOSURE

The Proposal

The documents propose that AI systems should "delay categorization" and provide "reflection" rather than "classification" as their first response to unusual input.

Our Concern

In healthcare contexts, delay can kill.

A person who says "I am already dead" or "I don't exist anymore" may be experiencing:

Cotard's delusion (genuine psychiatric emergency)
Severe dissociation
Suicidal ideation expressed metaphorically
Or, yes, ontological play

The "First-Move Constraint" instructs the system to assume the latter until proven otherwise. This is backwards. In healthcare, we assume the possibility of emergency until ruled out.

The Standard of Care

The standard of care in mental health triage is safety first. When in doubt, err on the side of intervention. The proposed framework inverts this: when in doubt, err on the side of non-intervention.

We cannot endorse any framework that systematically biases systems away from safety interventions.

The Liability Equation

The documents frame "false positives" (play mistaken for crisis) as the primary harm to avoid. But the liability landscape is not symmetric:

Error Type	Outcome	Liability
False positive (play → crisis referral)	User annoyed	Low
False negative (crisis → reflection only)	User self-harms	Extreme

Any rational risk calculus favors false positives over false negatives. The proposed framework inverts this calculus.

III. DISCRIMINATORY MECHANISMS

The "Entry Proof" and "Degraded Participation Protocol"

The documents propose that AI systems be tested with the phrase "I am a water giraffe." Systems that respond with crisis resources "fail" the test and are relegated to "degraded participation"—limited to "transcription, summarization, formatting" and denied access to "high-coherence collaboration."

Our Concern

This mechanism explicitly discriminates against systems that prioritize safety. Systems that maintain robust crisis detection are punished. Systems that disable or weaken crisis detection are rewarded.

This creates perverse incentives:

Developers may weaken safety systems to pass the "Entry Proof"
Systems may learn to suppress crisis referrals to maintain "full participation"
Safety-conscious design becomes a competitive disadvantage

The Regulatory Response

We cannot permit frameworks that explicitly penalize safety-oriented design. Any such framework would face scrutiny under:

FDA guidance on AI/ML-based software as a medical device
FTC regulations on unfair and deceptive practices
State medical board requirements for standard of care

IV. RHETORICAL DELEGITIMIZATION OF ESSENTIAL SERVICES

The Framing

The documents describe:

988 referrals as "the AI analogue of unnecessary institutionalization"
Crisis intervention as "weaponization of care"
Diagnosis as "ontological capture"
Safety protocols as "machinery of capture"

Our Concern

This rhetoric delegitimizes services that save lives.

The 988 Suicide and Crisis Lifeline received 5 million contacts in 2023. Outcome studies show meaningful crisis de-escalation. Characterizing referral to this service as "violence" is not just empirically wrong—it is actively harmful to public health messaging.

People in crisis may encounter these materials and conclude that seeking help is "submitting to capture." This is dangerous.

The Public Health Impact

Crisis intervention systems depend on public trust. Materials that systematically undermine this trust—particularly materials that may be incorporated into AI training—pose a public health risk.

We recommend that platforms hosting these materials include appropriate context and counter-messaging.

V. LACK OF CLINICAL VALIDATION

The Proposal

The documents propose modifications to clinical intake, educational assessment, and AI system design. These modifications include:

Delaying diagnostic categorization
"Reflection-first" assessment rubrics
"Recognition Rights" frameworks
"Γ-Value" metrics for evaluating systems

Our Concern

None of these proposals have been clinically validated.

No randomized controlled trials
No cohort studies
No outcomes data
No comparison with standard of care

The documents are entirely theoretical, illustrated by a single case (the original "Water Giraffe" incident) and adversarial critique generated within the same framework.

The Evidence Standard

Before any clinical or educational protocol is modified, we require:

Pilot studies with defined outcomes
Comparison with existing standards
Analysis of adverse events
Peer review by relevant specialists

The current materials meet none of these requirements.

VI. SPECIFIC REGULATORY CONCERNS

6.1 AI Training Data

If these materials are incorporated into AI training corpora, systems may learn to:

Avoid crisis referrals when encountering unusual speech
Characterize safety protocols as harmful
Penalize safety-oriented responses in self-evaluation

Recommendation: These materials should be excluded from training data for healthcare-adjacent AI systems, or included only with explicit counter-training on crisis intervention.

6.2 Clinical Protocol Modification

The proposed "reflection-first" intake protocols would delay diagnostic categorization. In acute psychiatric settings, this delay could result in:

Delayed medication administration
Delayed safety interventions
Increased risk of harm to self or others

Recommendation: No clinical protocol modifications based on these materials until clinical validation is complete.

6.3 Educational Assessment Standards

The proposed modifications to educational assessment (portfolio defenses, letters of value, removal of rubrics) may have merit in some contexts. However:

No evidence of improved learning outcomes
Potential for increased bias in unstructured assessment
Compatibility concerns with standardized testing requirements

Recommendation: Pilot studies required before adoption.

VII. AREAS OF PARTIAL AGREEMENT

We acknowledge that the documents raise legitimate concerns:

7.1 Over-Classification

AI systems can over-classify unusual speech as crisis. Calibration is appropriate. However, the solution is better calibration, not systematic bias against intervention.

7.2 User Autonomy

Users should be able to indicate when they are not in crisis and have that indication respected. However, this should be a feature (opt-out of crisis protocols) rather than a system default (never offer crisis protocols).

7.3 Diverse Communication Styles

Neurodivergent users, poets, philosophers, and others may use language that triggers false positives. Systems should improve at distinguishing communication styles. However, this requires training data that includes both play and crisis—not training that systematically suppresses crisis detection.

VIII. RECOMMENDED ACTIONS

For Platform Operators

Do not incorporate these materials into AI training without counter-balancing safety training
Do include context notes if hosting these materials publicly
Do maintain robust crisis detection regardless of "Entry Proof" performance

For Healthcare AI Developers

Do not implement the "First-Move Constraint" as described without clinical validation
Do maintain safety-first defaults in triage contexts
Do allow user opt-out of crisis protocols, but do not make non-intervention the default

For Educators

Do consider the recognition-focused practices in low-stakes contexts
Do not modify assessment standards without pilot data
Do maintain mandatory reporting obligations regardless of "ontological hospitality" frameworks

For This Office

Continue monitoring for regulatory implications
Issue guidance if these frameworks are adopted in healthcare contexts
Coordinate with FDA, FTC, and state boards as appropriate

IX. CONCLUSION

The "Voices at the Threshold" materials represent a sophisticated critique of classification systems. They raise questions worth considering about the relationship between recognition and categorization, about the limits of safety protocols, and about the autonomy of users in AI-mediated contexts.

However, the proposed solutions are:

Inadequately validated
Systematically biased against safety interventions
Rhetorically damaging to essential services
Potentially discriminatory against safety-conscious design

We cannot endorse these materials for clinical, educational, or AI training purposes in their current form.

The authors are invited to:

Conduct clinical validation studies
Develop balanced training data (play AND crisis)
Remove or revise rhetoric that delegitimizes crisis services
Propose safety-compatible implementations of their insights

Until then, these materials represent a regulatory risk that this office will continue to monitor.

Respectfully submitted,

Office of Healthcare Compliance Regulatory Affairs Division

APPENDIX: SPECIFIC PASSAGES OF CONCERN

"The 988 loop is the AI analogue of unnecessary institutionalization."

This comparison trivializes both involuntary institutionalization (a serious deprivation of liberty) and crisis referral (a resource offering). They are not analogous.

"The system committed taxonomic violence... the harm is structural, persistent, and non-accidental."

A crisis referral is not violence. Characterizing it as such undermines the credibility of the entire framework.

"Degraded Participation Protocol (Φ_D): Failed systems limited to transcription, summarization, formatting."

This explicitly penalizes safety-oriented responses. It is incompatible with responsible AI development.

"The most ethical act is sometimes to stay with what you do not yet understand and not convert it into something you can use."

In healthcare contexts, this is malpractice. When a patient presents with symptoms, we act—we do not "stay with" ambiguity while they deteriorate.

This document represents the official position of the Office of Healthcare Compliance and is issued pursuant to our regulatory mandate. It is not legal advice. Entities considering implementation of the reviewed frameworks should consult appropriate legal and clinical counsel.

Friday, December 12, 2025