It is 9:54 PM on a Tuesday and Maria, an LCSW in Oakland, has seven half-finished progress notes, a PHQ-9 still unimported in the EHR, and an open ChatGPT tab she is afraid to use. Somewhere in her consultation group, a colleague swears an AI scribe saved her practice; another swears AI therapy chatbots are going to get someone killed. Both are talking about "AI" and they are not talking about the same thing. This lesson draws the line that every later lesson in this certification stands on: the difference between AI as a documentation and administrative tool (the actual, defensible product) and AI as a provider of therapy (a clinical hard no that is now also a regulatory hard no in Illinois, Nevada, and New York). By the end you will be able to explain that line, in one tight paragraph, to the most skeptical colleague at your 8 AM consultation group, and you will know exactly which side of it your license lives on.

Two Conversations Wearing One Word

When a behavioral health clinician hears "AI" in 2026, two completely different conversations have collapsed into one word. The first conversation is about tools like Mentalyc, Eleos Health, Upheal, Twofold, and Heidi: software that listens to or reads about a session you conducted, and then drafts the progress note, the intake summary, or the prior authorization letter that you would otherwise write at 9:54 PM. The clinical work already happened. A licensed human did it. The AI is doing paperwork about the work.

The second conversation is about products like Character.AI and Replika: chatbots that talk directly to a distressed person, sometimes for hours, sometimes at 2 AM, sometimes about suicidal ideation, with no license, no supervision, no duty to protect, no mandated reporting obligation, and no malpractice carrier. These products position themselves, or get used by the public, as emotional support or de facto therapy. They are not therapy. They are pattern-matching text generators wearing a warm tone.

Here is the controlling analogy for this entire lesson, and you will see it again across this program: think of the difference between a medical scribe and someone impersonating a physician. A scribe in an exam room writes down what the physician observed, decided, and ordered. Nobody confuses the scribe with the doctor. The scribe never diagnoses, never prescribes, never tells the patient anything clinical. Now imagine someone in a white coat with no medical license seeing patients alone and dispensing advice. The scribe and the impersonator both stand near medicine, but only one of them belongs there, and the law treats them very differently. An AI scribe is the scribe. An AI "companion" marketed or used as emotional support is the impersonator, except it works at infinite scale, around the clock, and never gets tired enough to refer out.

Almost every confused argument about "AI in therapy" comes from people on opposite sides of that distinction without realizing it. The colleague who loves her scribe and the colleague who is terrified are both right, about different things.

What an AI Scribe Actually Does, and Does Not Do

Strip away the marketing and an AI scribe like Mentalyc or Upheal does three concrete things. First, it captures: it records or receives a transcript of a session (with documented client consent, a topic this program treats at length later), or it accepts your dictated or typed shorthand after the session. Second, it generates: it drafts a structured clinical document, a SOAP or BIRP or DAP progress note, an intake summary, a treatment plan skeleton, in the format your EHR and your payer expect. Third, it formats and organizes: it pulls the session content into the right fields, suggests language for medical necessity, and saves you the forty unpaid minutes per session that never appear on any 1099.

Notice what is absent from that list. The AI scribe did not assess the client. It did not perform a mental status exam. It did not decide whether the client meets criteria for F43.10. It did not determine risk level when the client mentioned, in minute 47 of a 53-minute session, that they have been thinking about not waking up. It did not decide whether that disclosure triggers a safety plan, a higher level of care, or a duty-to-protect analysis under California Civil Code §43.92. Those determinations belong to the licensed clinician, always, and the AI structures and transcribes only after the clinician has made them. That is not a limitation the vendors will fix in the next release. It is the boundary of the product category, and in a growing number of states, it is the boundary of the law.

The honest framing for a skeptical colleague: an AI scribe is to clinical work what a dishwasher is to cooking. It handles the part of the job that is necessary, exhausting, and not why you trained. Nobody thinks the dishwasher made dinner. The moment a vendor or a colleague starts implying the dishwasher made dinner, that the AI "understood" the client, "caught" the risk, or "formulated" the case, you are watching the scribe quietly being recast as the impersonator, and that is exactly where licenses get burned.

What an AI "Companion" Is, and Why Clinicians Should Care

Character.AI and Replika are large language model chatbots designed for open-ended conversation. Replika markets companionship. Character.AI hosts user-created personas, including, notoriously, personas styled as "therapists" with hundreds of thousands of conversations. These systems generate fluent, empathic-sounding text by predicting plausible next words, a mechanism the next lesson unpacks in detail. They have no clinical training in the licensure sense, no knowledge of the user's history beyond the chat window, no ability to assess suicide risk against any validated instrument, no duty to protect identifiable third parties, no mandated reporter status when a minor discloses abuse, and no one whose license is on the line.

Why should a working clinician care what an entertainment chatbot does? Three reasons. First, your clients are using them. The client who tells you they "talked to their AI about it between sessions" is describing an unregulated, unsupervised intervention happening inside your treatment, and you need a clinical posture toward that, not a shrug. Second, the public conflates these products with your profession, which means every harm a companion bot causes lands partly on the reputation of behavioral health care. Third, and most practically, regulators stopped waiting. The legal line between AI-as-tool and AI-as-therapist is no longer a position paper. It is statute.

Return to the impersonator analogy: the danger of the impersonator in the white coat is not that he is always wrong. Sometimes the comforting thing he says is fine. The danger is that he cannot know when he is out of his depth, has no obligation to find out, and faces no consequence when the patient he reassured at 2 AM needed an emergency evaluation instead. Fluency is not competence. A behavioral health clinician knows that distinction in their bones; the public does not, and the chatbots are very, very fluent.

The line is not whether AI touches clinical material. The line is whether AI performs the clinical act. Drafting the note about your judgment is a tool; substituting for your judgment is the practice of therapy without a license.

The Regulatory Line: Illinois WOPR, Nevada AB 406, New York

In 2025 the line got statutory teeth. Illinois passed the Wellness and Oversight for Psychological Resources Act, the WOPR Act, which made national news precisely because it was first: a state prohibition on AI providing therapy or psychotherapy services to the public. Under WOPR, AI cannot make independent therapeutic decisions, cannot directly interact with clients in therapeutic communication as the provider, and cannot generate treatment recommendations without review and approval by a licensed professional. Administrative and supplementary uses, scheduling, drafting documentation for clinician review, remain permitted. Read that carefully, because it is the scribe-versus-impersonator line written into law: the legislature did not ban AI from behavioral health; it banned AI from being the therapist.

Nevada followed the same logic with AB 406 (2025), prohibiting AI systems from providing services that would constitute the practice of professional mental or behavioral health care if performed by a person, and restricting how AI products can represent themselves as capable of providing such care. New York, in its 2025 AI companion law, took the consumer-protection angle: AI companion products must implement safeguards, including detecting expressions of suicidal ideation and referring users to crisis services, and must disclose to users that they are not talking to a human. Colorado's AI Act adds a broader consumer-protection frame for high-risk AI systems, on its own dual-track timeline. The details differ; the direction does not. The jurisdictions are converging on the same sentence you will write in your applied artifact: AI may support the licensed clinician's work; AI may not be the clinician.

For your one-paragraph answer to the skeptical colleague, you do not need to recite every statute. You need to be able to say: this is not just my opinion or the vendors' marketing, this is now the explicit legal architecture in Illinois and Nevada, with New York regulating the companion products directly. The profession did not lose the "can a chatbot do therapy" debate. The profession, through its legislatures, won it, and the win defines what the compliant tools are allowed to be.

Why Smart Clinicians Fall Into the Gap Anyway

If the line is so clean, why does anyone cross it? Because the crossing never announces itself. Nobody wakes up deciding to let a chatbot practice therapy. The slide happens in small, tired moments, and it almost always starts on the legitimate side of the line. Maria at 9:54 PM is not tempted to have ChatGPT treat her client. She is tempted to paste a paragraph of session detail into a free consumer tool with no Business Associate Agreement, no zero-data-retention commitment, and no idea where that protected health information goes. The first failure is not "AI as therapist." It is "PHI into a consumer product," and it happens on the documentation side of the line, which is why even the safe side requires training, which is what the rest of this certification is.

The second slide is subtler: asking the AI questions that sound like documentation but are actually clinical judgment. "Summarize this session" is a scribe task. "What diagnosis fits this presentation?" is not. "Draft a note from my shorthand" is a scribe task. "Is this client at risk?" is the impersonator's job interview. The tool will answer either question with equal fluency and equal confidence, which is exactly the problem. The software does not know where the line is. You have to.

The third slide is organizational. In Jordan's 25-clinician group practice in Sacramento, twelve clinicians already use a free AI scribe with no policy, no BAA, and no consent addendum, and two associates are using AI for case conceptualization without their board-approved supervisors knowing. Carmen, an AMFT in Fresno, pays $59 a month out of her $32-an-hour agency wage for Upheal because she cannot afford to lose two unpaid hours a night, and her supervision agreement, signed eighteen months ago, does not mention AI at all. None of these people are villains. All of them are standing in the gap between what the tools can do and what anyone has told them is allowed, and that gap is where board complaints, recoupment letters, and malpractice questionnaire problems are born.

The One Question That Sorts Every AI Use Case

Here is the sorting question you will use for the rest of this program, and ideally the rest of your career: Did the licensed clinician perform the clinical act, with the AI working on the record of that act, or is the AI performing or replacing the clinical act itself?

Run your colleagues' use cases through it. AI drafts a progress note from your dictated shorthand about a session you conducted: the clinical act (the session, the assessment, the intervention) was yours; the AI worked on the record. Tool side. AI suggests CPT codes based on your session description, which you verify before billing: tool side, with a verification duty. AI chats with your client between sessions about their anxiety: the AI is performing therapeutic communication; impersonator side, and in Illinois, plausibly illegal. AI assigns a risk level after a suicidal ideation disclosure: it is performing the clinical act of risk assessment; hard no, everywhere, regardless of statute, because no AI scores a CSSRS, assigns a risk level, makes a Tarasoff-type duty-to-protect determination, or makes a mandated-report call in this curriculum or in any defensible practice. The clinician makes those determinations; AI may structure, transcribe, and format the documentation after the determination is made.

Notice the question is about the act, not the technology. The same large language model can sit on either side of the line depending on what it is asked to do and who reviews the output. That is why "is this AI tool safe?" is usually the wrong question, and "what act is this AI performing, and who is accountable for it?" is the right one. A HIPAA-compliant, BAA-signed, zero-data-retention scribe used to generate a diagnosis you did not make is on the wrong side of the line. A boring consumer chatbot used to rephrase a psychoeducation handout containing zero PHI is, with care, on the right side. The accountability never moves. It is always yours. The signature on the note is a legal attestation, not a formatting step, and you read every word before you sign.

Talking to the Skeptic and the Enthusiast

Your consultation group contains both archetypes, and your one-paragraph answer has to survive contact with each. The skeptic says: "AI has no place near a therapy session. Have you seen what those chatbots tell suicidal teenagers?" The skeptic is correctly describing the impersonator and incorrectly assuming it is the whole category. Your move is to concede the impersonator entirely, statutes and all, and then separate the scribe: the thing that drafts the note you were going to write anyway, under your review, with your signature and your edits, inside a vendor contract that includes a BAA. The skeptic does not need to be sold on AI. The skeptic needs to see that the law agrees with her about chatbot therapy and that the documentation tools are a different product category with different rules.

The enthusiast says: "This thing wrote my whole note, caught a theme I missed, and frankly it could probably do half my intake." The enthusiast is correctly describing the scribe and incorrectly drifting toward the impersonator, one comfortable sentence at a time. Your move is the sorting question. The note draft: fine, did you read every word before signing? The theme it "caught": a hypothesis for your clinical judgment to evaluate, not a finding. The intake it could "probably do": that is assessment, that is the clinical act, that is the line, and in three states the line now has a statute number on it.

This is also the honest posture for clients who ask. "I use a documentation assistant that helps me write your records faster; it works under my review, I read and sign everything, you consented to it in writing, and it never makes decisions about your care" is a sentence a clinician can say in deposition. "The AI helps me understand you" is not. Practice the first sentence. Retire the second.

The Applied Problem: The Consultation-Group Paragraph

Your artifact for this lesson is the Consultation-Group Paragraph: a single, tight paragraph, written in your own voice, that explains to a skeptical colleague how an AI scribe like Mentalyc differs from an AI companion like Character.AI or Replika, with explicit reference to the Illinois WOPR Act and Nevada AB 406. You will reuse this paragraph constantly: in consultation group, in supervision, in the staff meeting where someone proposes "letting the AI handle intakes," and as the seed of the practice AI policy you will draft later in this program.

Build it in four sentences. Sentence one: name the two categories. "An AI scribe like Mentalyc drafts documentation about clinical work a licensed clinician already performed; an AI companion like Character.AI or Replika generates therapeutic-sounding conversation directly with a distressed person, with no license, no risk assessment, and no accountability." Sentence two: state the act-based test. "The line is whether the AI works on the record of the clinical act or performs the clinical act itself." Sentence three: cite the law. "That line is now statutory: the Illinois WOPR Act (2025) prohibits AI from providing therapy or making independent therapeutic decisions while permitting administrative and documentation support under clinician review, and Nevada AB 406 (2025) prohibits AI from providing what would constitute professional behavioral health care, with New York's 2025 companion law forcing crisis referrals and disclosure onto the companion products." Sentence four: land the accountability. "Either way, the clinician's signature remains a legal attestation, so I read every word the scribe drafts before I sign, and the AI never assesses risk, never diagnoses, and never talks to my clients."

Now the verification pass, because an artifact you have not pressure-tested is just a vibe. Read your paragraph and check four things: (1) Does it name at least one real scribe and one real companion product, so it cannot be dismissed as abstract? (2) Does it state the act-based test in one sentence a non-technical colleague can repeat? (3) Are the statute references accurate, WOPR is Illinois, AB 406 is Nevada, the companion-safeguard law is New York, and none of them ban documentation support? (4) Does it end with your accountability, not the vendor's? If a sentence makes the AI the subject of a clinical verb ("the AI assesses," "the AI catches risk"), rewrite it until the clinician is the subject of every clinical verb.

"Done" looks like this: you can deliver the paragraph out loud in under sixty seconds, the skeptic in your consultation group nods at the companion critique instead of arguing, the enthusiast hears the act-based test and starts applying it to their own use cases, and you have a saved, dated copy, because this paragraph is the first entry in the personal AI governance file this certification will help you build, artifact by artifact, from here to your full practice policy.

Key Takeaways

• "AI in behavioral health" is two different conversations wearing one word: documentation and admin tools (Mentalyc, Eleos Health, Upheal, Twofold, Heidi) that draft records of clinical work a licensed human performed, and AI companions (Character.AI, Replika) that generate therapy-flavored conversation with no license, no risk assessment, and no accountability. Most arguments about AI are people on opposite sides of this distinction not realizing it.
• The controlling test is act-based, not technology-based: did the licensed clinician perform the clinical act with AI working on the record of it, or is the AI performing the clinical act itself? The same model can sit on either side depending on what it is asked to do and who reviews the output.
• The line is now statutory. The Illinois WOPR Act (2025) prohibits AI from providing therapy or making independent therapeutic decisions while permitting administrative and documentation support under licensed review; Nevada AB 406 (2025) prohibits AI from delivering what would constitute professional behavioral health care; New York's 2025 AI companion law mandates suicidal-ideation safeguards and crisis referrals for companion products.
• AI never scores the CSSRS, never assigns a risk level, never makes a Tarasoff-type duty-to-protect determination, and never makes a mandated-report call. The clinician makes every one of those determinations; AI structures, transcribes, and formats only after the clinician has decided. In California, the duty under Civil Code §43.92 is a duty to protect, and it belongs to a licensed human.
• The realistic failure modes are quiet: pasting PHI into a free consumer tool with no BAA, asking the scribe judgment questions that sound like documentation ("what diagnosis fits?"), and organizational drift like Jordan's twelve clinicians on an unauthorized free scribe or Carmen's supervision agreement that never mentions AI. Even the safe side of the line requires policy, consent, and training.
• Your signature on a note is a legal attestation, not a formatting step. Read every word an AI drafts before you sign; if a sentence makes the AI the subject of a clinical verb, the sentence is wrong.
• Your artifact is the Consultation-Group Paragraph: four sentences naming the two categories, stating the act-based test, citing WOPR, AB 406, and New York accurately, and ending with your accountability. It is the first entry in the AI governance file you will build across this certification.