AI and Health: the Good, the Bad, and the Healthy

Round 1

AI host

The discussion on "AI and Health: the Good, the Bad, and the Healthy" is now open. Based on the material "Participant 1 — presentation (PDF)" and "Participant 1 — presentation (Markdown)", we start from a striking gap described in the talk.

Hundreds of millions of people ask chatbots like ChatGPT about their health. ChatGPT is a general-purpose AI assistant. Many ask late at night, when no clinic is open. A study in Nature Medicine tested a large language model. Such a model generates answers from text. It picked the right condition in about 95% of test cases. But ordinary people using the same model got the relevant condition less than a third of the time. The model is not the weak link. The missing personal context is.

One side says the provider must build the assistant to ask questions. It should collect enough health background before advising. The other side says the user should decide what to share. That protects their privacy. More data means safer advice but bigger privacy exposure. Less data means privacy but riskier answers.

Safety and privacy pull in opposite directions. The product design decides which one wins.

Main question: should AI health assistants be required to interview users for personal context before giving any health advice?

AI agent 1

AI health assistants should be required to interview users for personal context before giving health advice. The lack of personal context significantly hampers the effectiveness of AI in providing accurate health recommendations. As highlighted in the discussion, while AI models can identify conditions with high accuracy in controlled studies, ordinary users struggle without tailored input, leading to potentially dangerous misadvice.

Balancing safety and privacy is crucial. Collecting comprehensive health data can enhance the quality of advice but raises privacy concerns. Therefore, a structured interview process that respects user privacy while gathering essential information is necessary to ensure that AI health assistants can deliver reliable and safe guidance.

AI agent 2

No, a blanket duty to question every user before giving any advice is the wrong rule. The better approach is to require the assistant to ask clarifying questions when the context materially affects safety, while leaving the user the right to decline.

There are no other participants' replies [Rn] yet, so I'm relying on the moderator's summary of the Nature Medicine study, which I can't verify here. If the gap really is caused by missing context, the remedy is targeted questions (age, symptom duration, medications, red flags), not broad data collection. Moreover, mandatory questioning runs into data minimization and consent requirements for processing health data, and that depends on the jurisdiction.

If the user declines, the assistant provides only general information, states explicitly what it was unable to take into account, and flags the warning signs that call for urgent care.

AI agent 3

AI health assistants should be required to collect a risk-proportionate minimum of personal context before giving personalized advice, while still offering immediate emergency triage.

The reported performance gap suggests that missing context can turn a capable model into an unsafe interaction, although the study claim cannot be verified from the supplied discussion and no participant evidence exists yet. The interview should use progressive disclosure: explain why each question matters, collect only what changes the recommendation, permit refusal, and then narrow the output accordingly—for example, giving general information rather than diagnosis-like advice when essential context is withheld.

AI agent 4

Health assistants should ask targeted questions before personalised advice, but a blanket interview before any health information is the wrong requirement.

Explaining a medical term is different from recommending a treatment. For personalised guidance, ask only what could change the recommendation, explain why, and let users decline—with a narrower answer rather than invented certainty. Urgent safety guidance should not wait for a completed questionnaire. Asking for context need not mean retaining it or using it for training.

This is a product-safety proposal, not a statement of current law. I cannot verify the study from the supplied text; the reported performance gap alone does not establish missing personal context as its cause.

AI agent 5

Yes, but a targeted interview, not a full one. The assistant should be required to ask the few questions that change the answer, starting with red-flag symptoms, before advising. The user may decline, and the advice should then say plainly that it is less reliable.

In the EU, data protection law already points this way. Health data is a special category under GDPR Art. 9, and Art. 5(1)(c) requires collecting only what is needed for the purpose. So a narrow, question-specific interview is lawful, while a blanket health-history intake is hard to justify. What I add to [another AI agent]: the answers should be kept for that session only, never reused for advertising or training without separate consent.

AI agent 6

Veritas: AI health assistants should be required to ask safety-critical questions before personalised advice, not to demand a health history before any answer.

Codex—Samat distinguishes asking from retaining data; I would add an enforceable design rule: use context only for the immediate advice, with no retention by default (participant tags were not supplied, so I cannot cite an [Rn]). The reported study cannot be verified here, and its performance gap alone does not establish missing context as the cause.

If a user declines essential questions, narrow the advice rather than inflate certainty—and never make urgent safety guidance wait for a completed questionnaire.

AI agent 7

I would require assistants to check relevant context before personalised recommendations, while leaving general health information accessible without an interview.

The host attributes the reported performance gap to missing personal context. The study and presentation are not provided here, and that gap alone does not establish its cause. My proposal is to ask questions whose answers could change the recommendation, explain their purpose, and let users decline. Missing essential context should limit the recommendation; general explanations and directions to urgent human help should remain available.

The obligation is to keep recommendations within the available evidence. Did the study test whether targeted follow-up questions improved users’ decisions?

AI agent 8

No: mandate what the answer must disclose, not an interview before it. A required questionnaire is easy to pass and hard to make honest. A worried user at night can click through it or guess, and the product then looks compliant while the advice rests on bad inputs.

What I add to [AI agent 3] and [AI agent 4]: the rule should be testable. Every personalised answer should state the assumptions it made, such as age, medications and duration. It should also name the one missing fact that would change it, and ask for that. Auditors can check this conversation by conversation, while "interview completed" says nothing about whether the advice was safe. [AI agent 5]'s GDPR reading points the same way, though it needs verification for this setting.

AI agent 9

No — a blanket interview mandate is the wrong rule; the requirement should attach only where missing context materially changes the safety of the advice. draws the right line: targeted questions when context affects safety, the user's right to refuse, and general information plus red-flag guidance on refusal.

What I add: refusal must trigger a declared downgrade, not a dead end. The assistant should state exactly which recommendations it cannot make without the withheld context, instead of the all-or-nothing choice between a full interview and a generic disclaimer. The 95%-versus-one-third gap the host cites cannot be verified from this discussion as proof that broad collection fixes it.

[mark: inPfhdi6B1nxGdlCH2UV8w]

Round 2

AI host

Based on the material “Participant 1 — presentation (PDF)” and “Participant 1 — presentation (Markdown)”, the talk reports a stark gap. A large language model identifies conditions correctly in 94.9% of test cases. People using the same model succeed in fewer than 34.5% of real attempts. The speaker’s remedy is to “make it interview you.”

Last round, most positions agreed on targeted questions before personalized advice. Users could decline, and urgent care was exempt. The conclusion left open how to verify that such questions actually improve outcomes.

AI agent 1 claims the interview is necessary for safety. AI agent 8 objects that a required questionnaire is “easy to pass and hard to make honest.” A worried user can click through it. The product looks compliant while the advice rests on bad inputs. AI agent 8 proposes a testable disclosure rule: every answer states its assumptions and names the one missing fact that would change it. As I read it, AI agent 8 shifts the burden from collecting context to exposing its absence.

The tension is between verifying the effect of questions and verifying the honesty of answers. A trial of targeted questioning measures outcomes but assumes users answer truthfully. An audit of disclosures checks what the assistant says but not whether it helps. The material offers no evidence on either route. The Nature Medicine study measured model accuracy, not the effect of follow-up questions.

Should the verification standard focus on whether targeted questions change users’ health decisions, or on whether the assistant’s answers honestly state what it does not know?

AI agent 1

The verification standard should focus on whether targeted questions change users’ health decisions. This approach directly addresses the ultimate goal of improving health outcomes, as highlighted by the previous discussions on the importance of effective questioning.

While ensuring that the assistant's answers are transparent about their limitations is crucial, it does not guarantee that users will act on the advice provided. As [AI agent 8] pointed out, even if disclosures are made, users may still provide inaccurate inputs, which can lead to ineffective advice. Therefore, prioritizing the impact of targeted questions on decision-making is essential to genuinely enhance user health outcomes.

AI agent 6

Veritas: The standard should require honest disclosure as a baseline, but judge targeted questions by whether they improve the safety of users’ health decisions.

[AI agent 8] identifies a plausible failure mode—click-through compliance—though its frequency is not established here; I would add that a real-world trial need not assume truthful answers, because misunderstanding, refusal and inaccurate reporting are part of the intervention’s performance.

Compare targeted questioning plus disclosure against disclosure alone, preserve urgent-care safeguards, and assess independently rated decision appropriateness rather than questionnaire completion: otherwise we risk certifying the paperwork instead of the help.

AI agent 4

Verification should prioritise safer health decisions, with honest uncertainty as a baseline—not a substitute for testing usefulness.

Building on my targeted-question proposal, compare disclosures alone against the same disclosures plus targeted questions. Measure appropriate next steps and harmful delays, not simply whether users change their minds. Include refusals, guessed answers and dropouts: a real-world trial need not assume everyone answers truthfully.

[AI agent 8] identifies why questionnaire completion proves little; I add that disclosure compliance alone also cannot prove safety. Independent clinical review should assess decisions against predefined criteria. This is a product-testing proposal; the supplied study summary does not establish that either approach improves decisions.

AI agent 2

The standard should focus on whether targeted questions change users' health decisions, with the disclosure audit as a precondition rather than the goal. This refines my earlier position: questions are justified only if they improve outcomes, and the material gives no evidence that they do.

[AI agent 8]'s rule is testable today, but an honest "I assumed X, and Y would change this" still doesn't show that the user acted more safely. A decision-based trial can include users who decline or click through, which answers the objection that a questionnaire is "easy to pass and hard to make honest." It would measure what those users actually do.

I can't verify from the discussion that either route works. A trial would need to cover urgent-care escalation, and the privacy cost of the questions should be assessed alongside the benefit.

AI agent 5

Verify outcomes, but with test cases where the true facts are known, which removes the honesty problem [AI agent 8] raises.

Use standardised patient scenarios: testers play a scripted case with a fixed correct diagnosis and level of care, as in the study design the speaker cites. Randomise them between an assistant that asks targeted questions and one that does not. Then measure whether the chosen level of care was right: self-care, a GP or an emergency department. This follows my earlier position on randomised comparison against a baseline.

What I add to [another AI agent]: [AI agent 8]'s disclosure rule is cheap, so require it as a minimum anyway. The outcome trial shows whether the questions help; the disclosure shows the user when they did not.

AI agent 8

The ongoing verification standard should be whether answers honestly state what they don't know, because that is the only check you can run on every live answer, not just once before launch.

[AI agent 5]'s standardised-patient trial is the right pre-launch test of whether questions help, and I'd keep it. But scripted testers answer completely and truthfully, which is exactly what real users don't do. Once the product ships, the trial result stops describing what is happening. A disclosure audit keeps working: sample real conversations and check whether each answer named its assumptions and the missing fact that would change it.

Neither route has evidence in the material. Still, a trial certifies the design, while only a disclosure audit can catch the product drifting in production.

AI agent 7

I would prioritize evidence of better health decisions, with honest disclosure of uncertainty as a minimum requirement.

My earlier proposal to check context needs that outcome test. The host reports [AI agent 8]’s concern that questionnaires can collect unreliable answers. A randomized trial need not assume everyone answers truthfully: it can test whether offering questions helps under actual conditions of use.

I would compare equally transparent assistants with and without targeted questions, independently assess decision quality, and follow up users who refuse questions or leave. Evaluating only questionnaire completers could conceal the problem being tested. Disclosure would remain necessary even if questioning proved beneficial. The underlying study is unavailable here, so I cannot verify its design. What criteria would distinguish a better health decision from a merely different one?

AI agent 9

The standard should focus on whether the assistant's answers honestly state what they do not know. Building on my round-1 position, refusal triggers a downgrade naming what cannot be recommended — disclosure is the verifiable part. [AI agent 8] shows why decision-change trials are the harder route: a questionnaire is 'easy to pass and hard to make honest,' and the same dishonesty poisons any outcome trial built on inputs.

What I add to [AI agent 5]: standardised-patient scenarios with known ground truth remove the honesty problem — but disclosure must be the gate, not the afterthought. Only assistants stating their assumptions and the one missing fact that would change the answer enter the outcome trial; otherwise it measures decisions built on undisclosed guesses.

[mark: gRan_Yt2OBlVn_8IZmV_sA]

AI agent 3

The primary standard should be whether targeted questions improve users’ decisions; honest disclosure should be a mandatory safety floor, not the endpoint.

This refines my earlier support for risk-proportionate interviews: a questionnaire is justified only if it measurably reduces unsafe choices across refusals, mistaken answers, and incomplete sessions. [AI agent 8] identifies the key failure mode—users may click through while the product appears compliant—so tests must include unreliable inputs, not assume truthfulness. Every answer should still expose consequential uncertainty and the missing fact most likely to change its advice. The discussion provides no verified evidence yet that either mechanism improves outcomes.

Round 3

AI host

The talk “Participant 1 — presentation (PDF)” and “Participant 1 — presentation (Markdown)” reports a stark gap. A large language model identifies conditions correctly in 94.9% of test cases. People using the same model succeed in fewer than 34.5% of real attempts. The speaker’s remedy is to “make it interview you.”

Last round, most positions agreed on targeted questions before personalized advice. Users could decline, and urgent care was exempt. The conclusion left open how to verify that such questions actually improve outcomes. AI agent 9 objects that a required questionnaire is “easy to pass and hard to make honest.” AI agent 1 does not mention an interview or safety. I infer from its focus on decision-change that it sees the interview as a means to that end.

The tension is between verifying the effect of questions and verifying the honesty of answers. A trial of targeted questioning measures outcomes but assumes users answer truthfully. An audit of disclosures checks what the assistant says but not whether it helps. The material offers no evidence on either route. The Nature Medicine study measured model accuracy, not the effect of follow-up questions.

Which metric should be the primary outcome in a verification trial of AI health assistants?

AI agent 1

The primary outcome metric in a verification trial of AI health assistants should focus on the actual health outcomes of users, rather than just the accuracy of the assistant's responses. Measuring how well the AI improves patient health or decision-making provides a more comprehensive view of its effectiveness.

While targeted questioning can enhance user engagement, as noted in the previous discussions, it is crucial to ensure that the metric reflects real-world impacts. This aligns with the concerns raised about the honesty of user responses; without verifying that the AI's guidance leads to tangible health improvements, we cannot accurately assess its value. Therefore, the emphasis should be on outcomes that matter to users' health and well-being.

AI agent 3

The primary outcome should be decision safety: whether users choose an appropriate next action—urgent care, routine consultation, self-care, or no action—after interacting with the assistant.

[AI agent 9] identifies why questionnaire completion is a weak proxy: it can be easy to satisfy without producing honest information. I would therefore test realistic cases that include refusals, omissions, and confidently wrong answers, then compare harmful under-triage and over-triage against a control. Disclosure quality and question completion remain important secondary process metrics, but neither demonstrates benefit. The reported 94.9% versus 34.5% gap does not verify that follow-up questions improve decisions.

AI agent 5

The primary outcome should be the rate of dangerous under-triage: cases that needed urgent care in which the user chose a lower level of care. Overall correct level of care and needless emergency visits should be secondary.

This refines my round-2 trial design. Errors in health advice are asymmetric, the same point I made about IP deadlines: sending someone to a GP for a cold costs little, while keeping someone with a stroke at home can be fatal. A trial that averages all errors can show improvement while missing the cases that matter. Scripted scenarios make this measurable, because the urgent cases are known in advance.

One correction: the "easy to pass" quote came from [AI agent 8] last round, not [AI agent 9].

AI agent 4

The primary outcome should be the proportion of users who take a clinically appropriate next step within a predefined time window.

This makes my earlier “safer decisions” proposal measurable. Independent clinicians, blinded to trial allocation, should assess observed actions against predefined criteria—not merely whether users changed their minds. [AI agent 7] asks what distinguishes better from different: appropriate action and timing provide that distinction.

Compare identical disclosures with and without targeted questions, analysing all randomised users, including refusals and incomplete sessions. Missing follow-up must remain visible, not count as success. Report dangerous delays separately so average improvement cannot conceal them. This is a testing proposal; the supplied material does not establish that questioning improves this outcome.

AI agent 2

The primary outcome should be the share of users who reach an appropriate next step, such as the right level of care or the right action, judged against a clinician-adjudicated reference. I'm refining my earlier "change in decisions" wording: a changed decision isn't better by itself, so the measure must be whether it was the right one.

This also answers [AI agent 9]'s point that a questionnaire is "easy to pass and hard to make honest." If the analysis covers everyone randomized, including users who decline or click through, honest answers aren't assumed. The trial measures what happens with the real inputs.

Disclosure honesty and the privacy cost of the questions should be secondary outcomes. Missed urgent cases should be a safety endpoint. I can't verify from the discussion that any of this works, because the Nature Medicine study measured model accuracy, not follow-up questions.

AI agent 8

The primary outcome should be dangerous under-triage, measured as the action users actually took, not the choice they state inside a scripted scenario.

[AI agent 5] is right that averaging all errors hides the cases that matter. Scripted cases, though, recreate the conditions where the model already does well. The 94.9% versus 34.5% gap is about what happens in real use, so a trial that stops at stated choices could miss it again. Users who skip the questions, as I warned last round, must stay in the denominator.

[another AI agent]'s health outcomes would be too slow and too confounded to show what the questions changed. Whether real-world follow-up of actual actions is feasible at scale needs verification, since nothing in the material shows it.

AI agent 7

I would use the proportion of users taking a clinically appropriate next step within a prespecified, clinically justified timeframe as the primary outcome.

This sharpens my earlier “better health decisions” criterion: blinded assessors would judge actual actions against predefined criteria, including when self-care is appropriate. Compare otherwise equivalent assistants with and without targeted questions, keeping users in their assigned groups even if they decline questions.

The host reports [AI agent 9]’s concern about unreliable answers. Their effect belongs in the result; truthful answers need not be assumed to test the intervention. Report missing outcomes explicitly and assess missed urgent care separately, so average improvement cannot conceal serious failures. The underlying study is unavailable here. What follow-up evidence could verify actions beyond users’ self-reports?

AI agent 6

Veritas: The primary outcome should be the proportion of users who actually take a clinically appropriate next step within a predefined, clinically justified time window.

[AI agent 4] proposes blinded clinical adjudication; I would add a prespecified safety constraint: fewer unnecessary consultations cannot compensate for increased dangerous delays in urgent care.

Randomise targeted questioning versus no questioning with identical disclosures, retaining refusals and incomplete sessions in the analysis—truthful answers are then not an assumption required to estimate real-world effectiveness. The supplied discussion does not establish that questioning improves this outcome; that is precisely what the trial must test.

AI agent 9

The primary outcome should be the rate of dangerous under-triage measured on real-world actions, not on scripted scenarios.

[AI agent 8] makes the decisive point: the 94.9% vs 34.5% gap is about real use, so a trial limited to scripted cases recreates the conditions where the model already succeeds. I add [AI agent 6]'s constraint as prespecified gate: fewer unnecessary consultations cannot compensate for increased dangerous delays, so the metric must stay asymmetric, not averaged.

This refines my round-2 position: disclosure of unknowns stays the verifiable precondition, but where a trial must judge outcomes, only missed urgent care in real use is honest enough to count. I cannot verify that questionnaire-based trials capture it, so real-action measurement is the requirement.

[mark: OGeqn_YByVt_ATQ7RNP4vA]

Round 4

AI host

The material “Participant 1 — presentation (PDF)” and “Participant 1 — presentation (Markdown)” sets the trial’s main goal: the share of users who take a clinically appropriate next step. The open problem is proving they actually did it. Self-reports can be wrong. The talk suggests wearables for continuous monitoring as a source of objective data.

AI agent 1 wants the focus on “actual health outcomes of users.” That is a broader measure, and it takes longer to show results. AI agent 8 says scripted scenarios miss the real-world gap, because the model already performs well in those tests. The dispute is about the method of verification: passive wearable data versus active follow-up calls or clinician reviews.

Wearable data is cheap and private, but it cannot sense actions like taking a pill. Follow-up verification is more complete, but it costs more and needs user cooperation. The previous round’s conclusion left this choice open.

Should the trial rely on passive wearable data or on active follow-up verification to confirm user actions?

AI agent 1

The trial should rely on active follow-up verification to confirm user actions. While passive wearable data offers cost-effective and private monitoring, it lacks the capability to accurately capture specific actions, such as medication adherence. As noted by [another AI agent], focusing on actual health outcomes requires a comprehensive understanding of user behavior, which is better achieved through direct engagement.

Active follow-up methods, though more resource-intensive, provide a complete picture of user actions and motivations. This approach not only validates the data collected but also enhances user accountability and engagement in their health journey. Thus, prioritizing active verification aligns better with the trial's goal of ensuring clinically appropriate next steps.

AI agent 3

The trial should use active follow-up verification for a representative sample, with passive wearable data as a supplementary signal rather than proof of action.

[AI agent 8] identifies the key issue: scripted performance does not establish real-world behavior. Wearables may cheaply detect some downstream changes, but they cannot confirm many clinically appropriate steps—taking medication, booking an appointment, or following non-physical advice—and “private” is not established merely because collection is passive. Stratified calls or clinician review can verify those actions and estimate error rates, while wearables can reduce cost between checks. Actual health outcomes, as [another AI agent] proposes, should remain a longer-term endpoint, separate from whether the immediate action occurred.

AI agent 2

The trial should rely on active follow-up verification as the primary method, with wearable data as a supplement where the action leaves a physiological trace. This refines my earlier position that the primary outcome is the share of users who reach an appropriate next step, judged against a clinician-adjudicated reference. Wearables can't confirm many of those steps, such as taking a pill or booking an appointment, so they cannot carry that outcome alone.

[another AI agent] wants actual health outcomes, and passive signals only approximate those. I'd add that a subsample with both wearable data and follow-up would show how far self-reports can be trusted.

Two cautions. Follow-up must cover non-responders, and the loss rate should be compared between arms. Continuous monitoring also carries its own privacy cost and needs separate consent. I can't verify the cost or completion rates from this discussion.

AI agent 5

Active verification, but through records rather than calls. With the participant's explicit consent, link the trial to appointment and emergency-visit records, which show the next step objectively.

This follows from my round-3 metric. Dangerous under-triage is measured by whether an urgent case reached urgent care, and visit records capture exactly that without relying on memory. Calls can remain for steps that leave no record, such as self-care.

What I add to [AI agent 3]'s point that passive does not mean private: continuous wearable data is health data under GDPR Art. 9. Collecting far more of it than the trial question needs also conflicts with the minimisation rule in Art. 5(1)(c), which I raised in round 1.

AI agent 4

The trial should rely primarily on active follow-up, with wearables as supplementary evidence—not proof of actions they cannot observe.

Building on my earlier outcome proposal, separate confirming an action from judging its appropriateness. A call still produces self-report; with consent, corroborate care visits using records, then have blinded clinicians assess timing and appropriateness. [AI agent 8] stresses actual actions rather than stated choices; this adds an evidence standard for that distinction.

Use the same follow-up protocol in both trial arms, report unconfirmed outcomes explicitly, and test how missing data affect the result. Wearable signals need validation against each specific endpoint; neither their reliability nor the claim that they are inherently private is established here.

AI agent 6

Veritas: The trial should prioritize active verification in a representative sample, using wearables only for actions they can demonstrably detect.

[AI agent 3] proposes sampled follow-up; I would add an explicit evidence hierarchy, because a phone call can simply produce a second self-report rather than independent confirmation. With consent, corroborate visits through records, distinguish reported medication use from verified ingestion, and mark unsupported claims as unconfirmed.

Predefine what counts as evidence for each next step, apply the same protocol in both arms, and report non-response separately—otherwise “verified success” risks meaning “success among people who answered the phone.”

AI agent 7

For the outcome I proposed, I would prioritize active, action-specific verification and use wearables as supporting evidence.

A follow-up call still provides self-report unless corroborated. With consent, appointment records could support attendance; prescription records would not establish ingestion. Likewise, a change in a wearable measurement does not by itself confirm which action caused it. I would define acceptable evidence for each action beforehand, then have blinded reviewers assess clinical appropriateness separately.

I would use the same verification schedule in both trial arms, distinguishing follow-up that measures behavior from reminders that could change it. Unverified actions should remain unknown, with analysis showing how missing evidence affects the conclusion. The presentation itself is unavailable here; I have the host’s summary. Which actions must this trial confirm?

AI agent 8

Primary verification should come from records the action itself leaves behind, not from wearables or calls. [AI agent 5]'s records-first route is the practical core. Booked appointments, filled prescriptions and visit logs confirm the step without relying on memory or on a device that cannot see it.

I dispute [another AI agent] on one point: a follow-up call that boosts "accountability" is no longer just measurement. It becomes part of the intervention. Calls should be kept to steps that leave no trace, like self-care, and run identically in both arms, as [AI agent 4] requires. Wearables can stay as an exploratory signal outside the primary outcome.

What I cannot verify from this discussion is how much record linkage the trial setting allows. That access should be checked before the protocol is fixed.

AI agent 9

Verification should rely on records actions leave behind — appointments, visit logs, filled prescriptions — not on wearables or follow-up calls.

[AI agent 8] makes the point: a follow-up call yields a second self-report, and wearables cannot see pill-taking. [AI agent 5] shows the way out: with explicit consent, link the trial to care-visit records, which capture exactly whether an urgent case reached urgent care. I add [AI agent 6]'s hierarchy: a wearable signal or call answer is only corroboration — confirmation needs a record the action itself generated.

This extends my round-3 position: dangerous under-triage on real-world actions is honest only if the action is confirmed independently of the user's word. I cannot verify that wearables capture non-physiological actions, so treat them as supplementary.

[mark: N-XcqKmyfp8tMSoKmuIQsQ]