When a patient completes an intake form that feeds into an AI-assisted clinical screening tool, a fundamental ethical question activates: did that patient actually consent to AI involvement in their care? Not to a checkbox they scrolled past. Not to a privacy policy that mentioned "automated processing" in paragraph fourteen. Did they genuinely understand what the system does, who reviews its outputs, and what authority the algorithm holds over their clinical pathway?
AI informed consent is one of the most consequential and underdeveloped areas in clinical AI deployment. At TheraPetic® Healthcare Provider Group, our Licensed Clinical Doctors and clinical informaticists have worked through this problem directly, building intake architectures for nonprofit support animal screening that must satisfy HIPAA, federal research regulations and basic clinical ethics simultaneously. What we have learned is that checkbox consent is not just insufficient. It is a liability and a trust failure.
The Checkbox Problem in AI-Assisted Clinical Settings
Traditional informed consent in clinical care has three recognized elements: disclosure, comprehension and voluntariness. The checkbox model satisfies the first element only in the most technical sense. It places words in front of the patient. Whether those words communicate meaningful information about AI involvement is an entirely different question.
In non-AI clinical settings, the informed consent doctrine evolved through decades of case law and federal regulation to require that patients understand the nature of a procedure, its material risks, alternatives and the professional making the judgment. The Belmont Report, foundational to all modern human subjects protections, identifies respect for persons as its first principle. That principle requires more than exposure to a disclosure document. It requires genuine understanding.
AI introduces new disclosure obligations that no legacy consent framework anticipated. The algorithm is not a neutral tool like an X-ray machine. It is a probabilistic reasoning system trained on historical data, producing outputs that carry statistical uncertainty. That uncertainty is morally relevant. A patient deserves to know that the clinical pathway they are entering is shaped in part by a system that can be wrong in systematic ways, and that those systematic errors may not be randomly distributed across demographic groups.
Research published in JAMA and NEJM AI has documented that clinical AI systems can exhibit performance disparities across race, age, gender and socioeconomic proxies embedded in training data. Failing to disclose this in a consent process is not just an oversight. It is a substantive omission of material information.
What Disclosure Actually Requires When AI Is in the Loop
Disclosure for AI-assisted clinical evaluation must go beyond identifying that AI is used. It must explain what the AI does, what data it processes, what outputs it generates, and how those outputs influence the clinical decision.
At minimum, adequate disclosure in this context requires the following elements:
- System description. A plain-language explanation of what the AI system does. Not "we use advanced technology to improve your care." A concrete description: "Our intake system analyzes your responses to screening questions using a machine learning model trained on prior clinical assessments. The model produces a recommendation that our Licensed Clinical Doctor reviews before any determination is made."
- Data inputs and retention. What information the AI processes, how long it is stored, whether it is used to retrain the model, and what deidentification standards apply. For systems handling protected health information, HIPAA Safe Harbor or Expert Determination deidentification standards must be disclosed in accessible terms.
- Clinician oversight structure. Whether the AI recommendation is advisory or whether it gates access to care. Patients must understand that a human Licensed Clinical Doctor reviews AI outputs and that the AI does not make final clinical determinations unilaterally.
- Known limitations. Disclosure of the fact that AI systems can produce errors, that error rates may vary across demographic groups, and what the patient can do if they believe the system produced an inaccurate result.
- Right to opt out. Whether the patient can request a fully human-conducted evaluation and what the procedural pathway for that request looks like.
The Federal Trade Commission and the Office for Civil Rights have both issued guidance in recent years reinforcing that automated decision-making in consumer-facing contexts requires meaningful disclosure. In healthcare, the standard is higher. The expectation of therapeutic relationship imposes an affirmative duty of transparency that commercial contexts do not carry.
Comprehension Validation: The Standard Checkbox Ignores
Disclosure alone does not satisfy informed consent. The patient must understand the disclosure. This is the element that checkbox-based consent systems structurally cannot address, and it is the element that matters most when AI is in the clinical loop.
Comprehension validation is not a novel concept. The teach-back method, widely used in surgical and oncology consent processes, asks patients to explain in their own words what they understand about a procedure. The same principle applies to AI-assisted clinical evaluation. If a patient cannot articulate that an AI system reviewed their responses before a Licensed Clinical Doctor formed a recommendation, meaningful consent has not been obtained regardless of what they checked.
In digital intake environments, comprehension validation can be embedded structurally. Brief post-disclosure comprehension checks, presented as conversational prompts rather than tests, can confirm that a patient understood the three or four core elements of AI involvement before they proceed. This is not paternalistic. It is what the Belmont Report's respect for persons principle actually demands.
Reading level matters here too. FDA guidance on patient-facing labeling targets a sixth-grade reading level. Consent documents for AI systems routinely exceed a twelfth-grade readability score. When TheraPetic®'s clinical team reviews consent language for our HANK AI-assisted intake workflows, we require readability testing using Flesch-Kincaid before any document goes live. It is a small technical step that has significant ethical weight.
For populations with limited English proficiency, cognitive disabilities or low health literacy, comprehension validation requires additional structural support. Language access under Section 1557 of the Affordable Care Act applies to covered health programs. An AI consent process that is functionally inaccessible to a segment of the patient population is not just an equity problem. It is a compliance problem.
IRB Considerations for Nonprofit AI Screening Research
This is the area where nonprofit healthcare organizations deploying AI most frequently underestimate their obligations. When AI-assisted clinical screening generates data that is analyzed to evaluate system performance, improve the model or produce generalizable knowledge about a clinical population, it crosses into the domain of human subjects research as defined under 45 CFR 46, the Common Rule.
The definition of human subjects research under the Common Rule is not limited to formal clinical trials. It covers any systematic investigation designed to develop or contribute to generalizable knowledge that involves obtaining information about living individuals. If your organization is using AI intake data to measure model accuracy, identify demographic performance gaps or refine screening thresholds, that activity may require Institutional Review Board review and approval regardless of whether you call it quality improvement or research.
For nonprofit healthcare providers structured as 501(c)(3) organizations, this matters in a specific way. Federal funding relationships, including those mediated through state Medicaid programs or HRSA grants, can subject an organization to federal research regulations even if the research itself is not directly federally funded. The funding nexus analysis is not simple and requires competent legal review.
IRB review for AI-assisted screening research introduces additional considerations beyond standard human subjects protections. An IRB protocol for an AI screening system should address: the training data provenance and any historical biases embedded in it, the demographic composition of the validation cohort, the mechanism for ongoing monitoring of model drift, and the process by which a patient who was harmed by an erroneous AI output can seek recourse.
The Partnership on AI and the Alan Turing Institute have both published frameworks for ethical AI research governance that address IRB-adjacent questions for AI-specific risks. These frameworks do not replace federal Common Rule compliance, but they provide a useful supplemental structure for nonprofit organizations building out their research governance.
How TheraPetic® Structures Consent for AI-Assisted Evaluation
TheraPetic® Healthcare Provider Group operates as a 501(c)(3) nonprofit healthcare provider with EIN 81-3003968. Our clinical AI infrastructure, including HANK AI and the verification platform at verify.mypsd.org, processes screening data for support animal documentation and psychological screening pathways. The consent architecture we have built for these systems reflects what we believe adequate AI informed consent requires in practice.
Our consent workflow separates disclosure into two sequential layers. The first layer presents a plain-language summary of AI involvement: what the system does, who oversees it, and what the patient can expect from the process. The second layer presents the full technical disclosure for patients who want to understand data processing, retention and HIPAA compliance in detail. Patients are not required to read the second layer, but it is not buried. It is one click away and clearly labeled.
After the first-layer disclosure, our intake system presents three comprehension check questions. These are not trick questions. They are simple confirmations: "Who reviews the AI's recommendation before a determination is made?" Patients who do not answer correctly are shown a clarifying explanation before proceeding. This is not a barrier to access. It is evidence that we took our ethical obligation seriously.
Our data governance infrastructure links to mydatakey.org, where patients can review how their health data is stored, deidentified and governed. The network hub at therapetic.net connects this governance documentation to our clinical service platforms, creating a transparent record that regulators and patients can access without a legal request.
Dynamic Consent as a Clinical Infrastructure Problem
Static consent obtained at intake does not cover the full lifecycle of AI involvement in a patient's care. If the underlying model is updated, if data is used in a new way, or if a patient's information is included in a retrospective analysis, the original consent may not cover those activities. This is the problem that dynamic consent frameworks are designed to solve.
Dynamic consent, as described in research published in the JAMIA and explored by the Global Alliance for Genomics and Health, allows patients to update their consent preferences over time through a persistent digital interface. In clinical AI contexts, this means a patient could consent initially to AI-assisted screening, later withdraw consent for their data to be used in model retraining, and still retain access to care through human-only evaluation pathways.
Implementing dynamic consent is not trivial. It requires patient identity persistence across sessions, a consent state management system integrated with the clinical record, and a workflow that propagates consent revocations to data processing pipelines in real time. For organizations using FHIR R4-compliant electronic health record infrastructure, the Consent resource provides a structured data model for capturing and versioning consent states. Mapping AI-specific consent attributes to this resource is an unsolved interoperability problem that the HL7 AI working group is actively addressing as of 2026.
Nonprofit healthcare organizations that deploy AI without building toward dynamic consent are making a short-term architectural decision that will create long-term compliance and ethical debt. The cost of retrofitting consent infrastructure is substantially higher than building it correctly at the outset.
Building Trustworthy AI Consent Pipelines in Nonprofit Healthcare
Trustworthy AI consent is not a legal compliance exercise that concludes at signature. It is an ongoing clinical and institutional commitment to transparency, comprehension and patient agency that must be embedded in the technical infrastructure of the system itself.
For AI engineers and clinical informaticists building these systems, the practical implication is that consent is a data pipeline problem as much as it is a legal problem. Consent states must be captured in structured, queryable formats. Consent revocations must trigger automated suppression of data use across all downstream processing. Comprehension check results must be logged as clinical documentation, not discarded as UX telemetry.
For Licensed Clinical Doctors and clinical ethics committees, the implication is that they must be involved in consent design from the beginning of AI system development, not brought in at the end to approve a legal document that engineers have already built around. The clinical voice must shape the disclosure language, the comprehension check design and the opt-out pathways.
At TheraPetic®, our approach reflects the clinical oversight model described above. Dr. Patrick Fisher, PhD, LPC, NCC leads the clinical review process for all consent-adjacent documentation, ensuring that what patients read actually reflects how the system works. That alignment between technical reality and patient-facing disclosure is not just ethically required. It is the foundation of the therapeutic relationship that makes clinical AI trustworthy rather than merely functional.
The checkbox will always exist. The question is whether anything meaningful stands behind it. For nonprofit healthcare organizations deploying AI in clinical evaluation, the answer to that question defines not just their legal exposure but their character as institutions committed to patient dignity and care.
Explore the full TheraPetic® clinical AI infrastructure at therapetic.net and our flagship support animal documentation platform at mypsd.org. AI-assisted service animal verification tools are available at servicedog.ai.