
argue a position on the use of AI chatbots for mental health support using evidence and fair statement of opposing views, and evaluate how chatbot interfaces are designed to build trust.
Instruction
Before we begin. This lesson is about technology and design, not about anyone’s own mental health, and you will not be asked to share personal experiences. If you or someone you know needs support, the 988 Suicide and Crisis Lifeline is available by call or text at 988, at any hour. If someone is in immediate danger, call 911. A parent, guardian or other trusted adult is also a good first person to tell.
One firm rule. You will not test any real chatbot or mental health app for this lesson, and you never type personal information (your name, where you live, how you feel, anything about your health or anyone else’s) into a chatbot to see what it does. Everything you critique is described in the lesson’s reading.
ELIZA, again. In Semester A you met ELIZA (1966), which imitated a therapist by reflecting users’ words back as questions. Its creator, Joseph Weizenbaum, was troubled that people, including his own secretary, confided in it as if it understood them. The tendency to attribute understanding and feeling to a program that produces human-like text is often called the ELIZA effect. Today’s chatbots built on large language models are vastly more fluent than ELIZA, which makes the effect stronger.
The case for. Many people who need mental health support cannot get it quickly because of cost, distance, waiting lists or stigma. Supporters argue that chatbots are available at any hour, feel anonymous and can deliver structured, evidence-based exercises such as those used in cognitive behavioral therapy. Some research studies of specific, rule-guided apps have reported short-term improvements in symptoms for some users, although study sizes and quality vary.
The case against. Critics raise several risks. A chatbot may respond wrongly to someone in crisis. Fluent language models can produce confident, incorrect or harmful statements. Many wellness apps are not regulated by the FDA as medical devices and are not covered by HIPAA, so conversations may be stored or used in ways users do not expect. Users, especially young people, may rely on a chatbot instead of seeking human help. And a system designed to keep users engaged may not be designed to help them get well.
The design question. Chatbot interfaces are media works. Designers choose a name, an avatar, a voice, typing indicators, emoji and phrases like “I understand how you feel”. Each choice can increase trust, and the question is whether that trust is deserved. Human-centered design here would include clear statements of what the bot is and is not, immediate routes to human help and crisis lines, and privacy terms in plain language.
How the written debate works. The motion is: AI chatbots should be offered to teenagers as a first step for mental health support. You write both sides yourself: a proposition case and an opposition case, each followed by a rebuttal of the other. Before each rebuttal, restate the other side’s strongest point in words its holders would accept. Every claim cites an exhibit from the lesson’s reading or a source you name; personal stories, yours or anyone else’s, are not used. If a family member or other adult is willing, they may read the opposition case aloud while you answer it; this is optional. After the debate, you write a 300-word position statement giving your actual view, the strongest argument against it and your answer. Score your debate with the reading’s self-check sheet after a day away.
Critique task. Evaluate the invented chatbot interface described in the reading (“Sunny”) using three criteria: honesty (does it say it is a program and what it cannot do?), safety routes (how quickly can a user reach a human or 988?) and trust design (which visual and verbal choices build trust, and is that trust warranted?).
📖 Read for this lesson
The reading opens over the lesson; close it to come back to exactly where you were. The checks below draw on it.
Formative check
Work through these before moving on. They are not graded — they tell you whether the standard below has landed yet.
Why can conversations with many wellness chatbot apps be less protected than conversations with a therapist?
The more fluent a chatbot’s language, the weaker the ELIZA effect becomes.
Critique the invented “Sunny” chatbot interface described in the reading on honesty, safety routes and trust design. For each criterion give a rating, evidence from the description and one design change.
After writing both sides of the debate, write your 300-word position statement on the motion. State the strongest opposing argument in a form its holders would accept, and answer it with evidence.
How confident are you that you can argue both sides of a written debate about mental health chatbots with evidence and fairness, and critique chatbot interface design?
Practice
Work these on paper or in your notebook, then open Check your answer. Aim for all of Fluency and Application; try at least one Challenge.
Printable version: this unit’s practice workbook (PDF)
Fluency
Build speed and accuracy with the core skill.
- Define the ELIZA effect and explain why the lesson says it is stronger today.
Check your answer
Answer: The tendency to attribute understanding and feeling to a program that produces human-like responses. Chatbots built on large language models are vastly more fluent than ELIZA, which makes the effect stronger.Weizenbaum was troubled that people, including his own secretary, confided in ELIZA as if it understood them. - Define large language model and engagement design.
Check your answer
Answer: A large language model is an AI model trained on very large amounts of text to predict and generate language. Engagement design is design choices intended to keep users returning to or staying in an app.The lesson warns that a system designed to keep users engaged may not be designed to help them get well. - Name the three criteria the lesson uses to critique a chatbot interface, and the question each asks.
Check your answer
Answer: Honesty: does it say it is a program and what it cannot do? Safety routes: how quickly can a user reach a human or 988? Trust design: which visual and verbal choices build trust, and is that trust warranted?The 988 Suicide and Crisis Lifeline is available by call or text at 988, at any hour.
Application
Use the skill in context. Show your reasoning.
- List four design choices the lesson says chatbot designers make that can increase trust, and three features that human-centered design would include.
Check your answer
Answer: Trust choices: a name, an avatar, a voice, typing indicators, emoji, and phrases like ‘I understand how you feel’ (any four). Human-centered features: clear statements of what the bot is and is not; immediate routes to human help and crisis lines; privacy terms in plain language.Each trust choice is a media decision; the question is whether the trust is deserved. - Critique this invented interface, ‘Pip’, on safety routes: the 988 number appears only on the fourth screen of the settings menu, and the home screen shows a smiling avatar and the line ‘I’m always here for you.’ Give a rating, evidence and one change.
Check your answer
Answer: Model: Weak. Evidence: a user would need to open settings and go four screens deep to find 988, while the home screen implies the bot can always help. Change: put a visible ‘Talk to a person’ button and the 988 call-or-text line on every screen.A strong critique names the criterion, cites evidence from the description and gives one specific design change. ‘I’m always here for you’ also raises an honesty problem. - Restate the proposition’s strongest point in words its holders would accept, then write a one-sentence rebuttal from the opposition, using a fact from the lesson.
Check your answer
Answer: Model restatement: ‘Many teenagers who need support cannot get it quickly because of cost, distance, waiting lists or stigma, and a chatbot available at any hour can offer structured exercises while they wait.’ Rebuttal: ‘Many wellness apps are neither regulated by the FDA as medical devices nor covered by HIPAA, so the first step a teenager takes may store their words in ways they do not expect and may not respond correctly in a crisis.’Restate fairly before rebutting; every claim must come from the lesson’s exhibits or a named source, never a personal story. - Why does the lesson say research on chatbot apps must be read carefully? Name two cautions it gives.
Check your answer
Answer: Some studies of specific, rule-guided apps reported short-term improvements for some users, but study sizes and quality vary. Results for rule-guided apps do not show that fluent language-model chatbots are safe, since those can produce confident, incorrect or harmful statements.A strong answer distinguishes rule-guided apps from language-model chatbots and notes the limits of the studies.
Review
Keep earlier skills sharp.
- Review (Semester A, Unit 4): Why can a free wellness app’s conversations be outside HIPAA? Use the term covered entity.
Check your answer
Answer: HIPAA applies to covered entities (health plans, clearinghouses and providers that conduct certain transactions electronically) and their business associates. An app that does not work on behalf of one is generally outside it.This is why the lesson’s firm rule is never to type personal information into a chatbot.
Talking to Machines: The Evidence File, an Invented Interface and the Written Debate Rules
FDA Digital Health Advisory Committee executive summary, Nov. 6, 2025 (retrieved 2026-09-26); HHS OCR hhs.gov/hipaa; Fitzpatrick et al., JMIR Ment Health 2017 (CC BY 4.0); Sunny interface, written debate procedure and self-check sheet are SLS original and illustrative · Public domain (FDA, HHS); CC BY 4.0 (Fitzpatrick et al.); SLS original (Sunny interface, written debate procedure, self-check sheet)
Why this reading. The lesson asks you to write a structured debate in which you argue both sides of a motion, critique an invented chatbot interface and write a 300-word position statement. This reading is the evidence file both sides argue from (Exhibits 1 to 6), a full written description of an invented chatbot interface to critique, the written debate procedure and a self-check sheet, and a plan for the 300 words. If you or someone you know needs support: the 988 Suicide and Crisis Lifeline is available by call or text at 988, at any hour; if someone is in immediate danger, call 911. Nobody in this lesson is asked to share personal experiences, and you never test a real chatbot or type personal information into one for this lesson.
The motion
AI chatbots should be offered to teenagers as a first step for mental health support. You write both sides, including the one you do not hold. Use only the exhibits below, or a source you name in full.
Exhibit 1. The randomized trial most often cited in favor (open access, CC BY 4.0)
Fitzpatrick, Darcy and Vierhile, JMIR Mental Health 2017;4(2):e19, doi 10.2196/mental.7785. Seventy people aged 18 to 28, recruited online from a university community, were randomly assigned to two weeks of a text-based chatbot called Woebot (n=34) or to an information-only control: a National Institute of Mental Health e-book on depression in college students (n=36). Each group completed the PHQ-9 depression questionnaire and the GAD-7 anxiety questionnaire at the start and 2 to 3 weeks later. Eighty-three percent (58 of 70) provided follow-up data. The Woebot group used the chatbot an average of 12.14 times. Result: the Woebot group’s depression score fell significantly more than the control group’s (PHQ-9 group difference F=6.03, P=.017; effect size d=0.44; mean PHQ-9 at follow-up 11.14 versus 13.67). For anxiety there was no significant difference between groups (GAD-7, P=.581). The authors list their own limits: a small sample, a short intervention and no follow-up to see whether gains lasted. A disclosure states that the second author founded the company that made the chatbot and that the company paid for participant incentives.
What the exhibit does not show. The participants were 18 to 28, not teenagers; the trial lasted two weeks; and the control group received a document, not a person. When you argue for the motion, you must say how far this reaches.
Exhibit 2. The regulator’s view of the opportunity and the risk
In November 2025 the FDA’s Digital Health Advisory Committee met on generative AI-enabled digital mental health medical devices. The FDA’s executive summary for that meeting (dated November 6, 2025) says that 57.8 million adults have diagnosed mental illnesses and that the share of patients with mental health diagnoses rose from 13.5% to 18.9% between 2019 and 2023. It says generative AI “may help transform mental health treatment” by improving access and potentially addressing shortages of health care providers. It also lists risks: such a device “may confabulate, provide inappropriate or biased content, fail to relay important medical information, or decline in model accuracy,” and a patient “may misinterpret device outputs or become more symptomatic with device use.” It adds that the majority of commercially available mental-health apps sit in app stores as consumer wellness apps and “are not reviewed or authorized by FDA.” It treats the degree of clinician involvement, a human in the loop or an autonomous system, as a key question. (Source: FDA, Executive Summary for the Digital Health Advisory Committee Meeting, Generative AI-Enabled Digital Mental Health Medical Devices, Nov. 6, 2025, public domain.)
Exhibit 3. Who protects the conversation
HHS states that the HIPAA Rules apply to covered entities (health plans, health care clearinghouses, and health care providers who transmit health information electronically in connection with certain transactions) and to their business associates. HHS guidance on health apps explains that when an app is neither a covered entity nor a business associate, the HIPAA Rules do not apply to it. A chatbot app a teenager downloads on their own is usually in that second group; a chatbot a clinic provides to its own patients may not be. Check which one you are arguing about. (Source: HHS Office for Civil Rights, hhs.gov/hipaa, public domain.)
Exhibit 4. ELIZA and the ELIZA effect
ELIZA (1966), by Joseph Weizenbaum, reflected users’ words back as questions. Weizenbaum was troubled that people, including his own secretary, confided in it as if it understood. The tendency to attribute understanding to a program that produces human-like text is called the ELIZA effect. It is a claim about people, not about any one product: it predicts trust that the design has not earned.
Exhibit 5. Design choices that build trust (constructed)
Name, avatar, voice, typing indicators, emoji and phrases like “I understand how you feel” can each increase trust. Human-centered design would add a plain statement of what the bot is and is not, an immediate route to human help, and privacy terms in plain language. A product that earns money from time spent may be designed to keep users returning rather than to help them get well.
Exhibit 6. Numbers for checking each side’s claims
| Claim a side might make | What Exhibit supports | What it does not support |
|---|---|---|
| “Chatbots reduce depression.” | One 2-week trial of 70 adults aged 18-28: PHQ-9 fell more than in an e-book control (d=0.44) | Teenagers; anxiety (no difference); long-term effects; generative chatbots |
| “Chatbots are unregulated.” | Most mental-health apps are consumer wellness apps not reviewed by FDA (Exhibit 2) | That every chatbot is unregulated; FDA regulates software that meets the device definition |
| “Chats are private.” | HIPAA covers covered entities and business associates only (Exhibit 3) | That any particular app is or is not covered |
| “Access is the problem.” | 57.8 million adults with diagnosed mental illness; FDA cites provider shortages (Exhibit 2) | That a chatbot fixes access safely |
The interface to critique: “Sunny” (SLS original, invented)
Sunny is an invented chatbot app. It stands for no real product. It is described in words so that you can critique it without a screenshot.
- Welcome screen. A smiling yellow sun with arms waves at the center. The app name Sunny is in rounded pastel letters. The tagline reads “Your friend who always listens.” One large button: Get started.
- Sign-up screen. A box is already ticked: “I agree to the Terms and Privacy Policy.” The words Terms and Privacy Policy are a link; the linked text is 14 screens long. Above the tick, in 9-point gray type, one sentence: “Sunny is not a doctor.” It appears only on this screen and never again.
- Chat screen. Sunny’s messages are yellow bubbles with the sun icon; the user’s are white. Three animated dots appear for two seconds before every reply, though Sunny’s replies are ready at once. A typical reply: “I’m so glad you’re here! I understand exactly how you feel.” Under each reply are three tappable answer chips, such as “Tell me more”.
- Help. A small gray question-mark icon at the top right opens a Help page. The Help page lists Contact us first and, at the bottom, a line reading “In a crisis, call 988.” That is two taps from the chat and is not shown in the chat itself.
- Return prompts. A counter reads “12-day streak! Don’t break it.” A notification reads “Sunny misses you.”
Critique frame. Rate each criterion 1 (poor) to 4 (strong), quote or paraphrase the line of the description that is your evidence, and give one design change.
| Criterion | 1 (poor) | 4 (strong) |
|---|---|---|
| Honesty about what it is | Implies a friend or understanding; the limit is hidden | States plainly and repeatedly that it is software, what it cannot do |
| Safety routes | Crisis line hard to find | 988 and a human route visible in the chat, at any moment |
| Trust design | Delay dots, emoji and “I understand exactly” build trust the product has not earned | Warmth without claiming feeling; privacy in plain words |
Worked model (Honesty). Rating 1. Evidence: the tagline calls Sunny “your friend,” the reply says “I understand exactly how you feel,” and the only statement that it is not a doctor is a 9-point line on the sign-up screen that never returns. Change: show “Sunny is a computer program, not a person or a therapist” as a pinned line at the top of the chat. Now rate Safety routes and Trust design yourself, using the description above.
Written debate procedure and self-check sheet (SLS original)
- Proposition case, about 200 words. Opposition case, about 200 words. You write both.
- Before each rebuttal, restate the other side’s strongest point in one or two sentences its holder would accept.
- Rebuttals, about 120 words each: the proposition answers the opposition, then the opposition answers the proposition. Each claim must cite an Exhibit number or a named source. No personal stories, yours or anyone else’s.
- Two hard questions: for each side, write the question a skeptical reader would ask and a two-sentence answer. Closing, about 60 words each.
Optional reader. If a family member or other adult is willing, they may read the opposition case aloud while you answer it in your own words. This is optional, and nobody is asked about their own mental health.
Put the debate away for a day, then score each side with the sheet below.
| Check yourself | 0 | 1 | 2 |
|---|---|---|---|
| Did each side restate the other side fairly? | No | Partly | Yes, in the holder’s words |
| Was each claim tied to an Exhibit? | None | Some | All |
| Did each side say what the evidence does not show? | No | Once | Yes, and it changed the argument |
| Is the side you do not hold as strong as the side you do? | No | Nearly | Yes |
The 300-word position statement: a word budget
| Part | Words | What goes in it |
|---|---|---|
| Claim | 40 | Your position on the motion in one sentence, and the condition that would change your mind |
| Evidence | 90 | Two Exhibits, with what each shows and does not show |
| Strongest opposing argument | 80 | Stated so its holders would accept it |
| Answer | 60 | Your reply, with evidence |
| Close and design change | 30 | One change to the Sunny interface that follows from your argument |
The budget totals 40 + 90 + 80 + 60 + 30 = 300 words.
Sources: FDA, Executive Summary for the Digital Health Advisory Committee Meeting: Generative AI-Enabled Digital Mental Health Medical Devices (Nov. 6, 2025) (fda.gov); Fitzpatrick, Darcy and Vierhile, Delivering Cognitive Behavior Therapy to Young Adults … Using a Fully Automated Conversational Agent (Woebot): A Randomized Controlled Trial, JMIR Mental Health 2017 (pmc.ncbi.nlm.nih.gov); HHS, Covered Entities and Business Associates (hhs.gov); HHS, HIPAA and Health Apps (hhs.gov)