Generative AI

Generative output predicts a plausible continuation

A language model can draft a summary, reorganize text, answer a question, or extract a proposed structure. It generates from learned patterns and supplied context. It does not retrieve truth by default or know which local record is authoritative.

Common failures include fabricated facts or citations, omitted qualifiers, inconsistent numbers, hidden assumptions, sensitive-data leakage, biased framing, and confident wording that exceeds the evidence.

Retrieval and tools add more failure points

Some systems can search documents, call external tools, calculate, or write into another application. These capabilities can improve grounding and also expand the control surface. A retrieval system may select the wrong client, stale policy, incomplete plan, or irrelevant payer rule. A tool call may use excessive permission or act before review.

Test identity, source ranking, date and version handling, citations, permission boundaries, timeout, retry, duplicate action, and safe failure. Display which sources were retrieved and which action the system proposes. Keep write operations behind explicit authority and approval. A sourced answer can still misread the source, so reviewers need the original material rather than a citation label alone.

Recheck this contract after every system change.

Approve uses one at a time

Separate public education, internal administration, payer correspondence, documentation, assessment, treatment planning, and family communication. Record the user, affected people, inputs, output, decision, consequence, approved model, vendor, version, and prohibited uses.

Low-stakes drafting can still disclose protected information or create a misleading promise. High-stakes clinical or access decisions require stronger authority, evidence, validation, and review. A product approval should never become permission for every prompt.

The CASP AI practice-parameters page addresses organizational selection, deployment, change management, monitoring, and auditing in ABA. Treat it as sector guidance, then apply the practice's governing sources.

Control sensitive data

Current HHS Security Rule guidance applies according to covered-entity, business-associate, and ePHI status. Map prompts, attachments, retrieved records, outputs, logs, feedback, support access, training use, subprocessors, retention, and deletion.

Use approved accounts and minimum data for the task. Test with fictional material before any authorized live use. A vendor promise that it does not train on prompts answers only one question.

Require source-grounded review

Every clinical statement needs a traceable source, such as direct observation, contemporaneous measurement, a clearly attributed client or caregiver report, or an existing signed record. Mark uncertainty and conflict.

The reviewer should see the source, recognize generated content, edit or reject it, understand the use boundary, and remain the final author. For covered BCBA and BCaBA certificants and applicants, the BACB Ethics Code keeps competence, confidentiality, consent and assent when applicable, documentation, and clinical accountability with the professional.

A fictional output review

Priya's practice locks 18 fictional outputs from an approved administrative-summary pilot. Fourteen are source-supported, complete, correctly attributed, and usable after review: 14 of 18, or 77.8%.

One fabricates a date. One omits a payer limitation. One blends two people's records. One uses an unsupported causal statement. Report each error type and keep all 18 in the denominator. This example measures one review cohort, not accuracy, safety, or time saved.

Prohibit autonomous clinical decisions

Generative AI should not independently diagnose, infer function, select goals, set dosage, decide risk, alter a plan, determine consent, sign a record, or communicate an unverified clinical conclusion. Administrative automation should not rewrite clinical content to satisfy a payer rule.

Qualified people decide within their authority. Software can surface source material, propose language, calculate, or flag when the approved workflow supports those actions.

Include clients and families

AHRQ's AI implementation brief suggests considering disclosure when AI generates documentation or influences decisions and notes that patient review may identify errors. This is broad healthcare guidance, not a universal consent rule.

Use accessible communication. Follow applicable consent, assent, recording, privacy, and correction processes. Preserve AAC and other communication supports. Provide a practical route to challenge a statement or decision.

Monitor the deployed system

The NIST Generative AI Profile identifies risks that can be new or intensified in generative systems and proposes cross-sector lifecycle actions. Monitor unsupported content, attribution, omission, correction burden, overrides, incidents, subgroup patterns, and model changes.

Predefine stop conditions and the manual path. Reassess after a model, prompt, retrieval source, vendor term, data flow, interface, population, law, or workflow changes. Retire access and data deliberately when the use ends.

For every release decision, record the tested model and configuration, accepted uses, excluded uses, known failure modes, reviewer, and expiration date. A vendor's silent model change should return the workflow to a held state until the relevant tests pass again.