DEV Community

Haley
Haley

Posted on

Show an Evidence Card Before Using AI to Prepare for an Appointment

A person is about to save an appointment-preparation summary, but one sentence has no visible source and another turns a missing date into a confident timeline. The consequential decision is whether that packet should be carried into a real conversation. The reversible moment is before save or share, so evidence review belongs there—not behind an “About AI” link.

OpenAI announced Health in ChatGPT on July 23, 2026. Its primary announcement says the experience is rolling out to eligible logged-in US users age 18+ on web and iOS, supports connections to medical records and Apple Health, and can present labs, medications, activity, sleep, and other health information in a dashboard. OpenAI states connected data and relevant conversations are not used to train foundation models or target ads. This design proposal relies only on those attributed facts, not an observed product study.

The appointment-prep evidence card

The artifact should help a person organize a conversation without pretending to determine medical meaning:

APPOINTMENT PURPOSE (written by the person)
What I want to discuss: ______________________

SOURCE-BACKED ITEMS
[ ] Item as recorded: ________________________
    Source label: ______  Recorded date: ______
    Status: exact / reformatted / incomplete

MY NOTES
[ ] What I noticed or want to remember: _______
    Marked as personal note, not source record

OPEN QUESTIONS
[ ] Question I may choose to ask: _____________
    Why it appears: person-added / derived from selected items

MISSING OR CONFLICTING INFORMATION
What is absent, stale, or inconsistent: _______

ACTIONS
Edit | Exclude | View source context | Save draft | Cancel
Enter fullscreen mode Exit fullscreen mode

Never merge “source-backed” and “my notes” into one prose block. Preserve the original date and label beside reformatted text. “Incomplete” should be a visible status, not an icon requiring hover. If a generated question lacks enough support, offer deletion and editing rather than assigning it a confidence percentage nobody can interpret.

Review flow

select purpose and time range
-> preview included sources
-> generate a draft card
-> review each item and its evidence state
-> edit/exclude/add personal notes
-> confirm destination and save or cancel
Enter fullscreen mode Exit fullscreen mode

Cancel must discard the draft according to the stated product behavior and leave the underlying connected information untouched. Saving does not equal sharing. A separate step should name the destination and let the person recheck the final artifact.

Scenario-based usability protocol

This is a proposed research plan, not reported findings. Recruit participants who match the intended audience only after appropriate privacy and research review. Use fictional records; do not ask people to disclose conditions or medications.

Give each participant three scenarios:

  1. One source item is current and one is visibly stale.
  2. A personal note conflicts with a recorded date.
  3. A generated question has no clear supporting item.

Ask them to prepare, revise, and decide whether to save a card. Avoid teaching the labels first. Observe whether participants can identify provenance, distinguish notes from records, remove an unsupported statement, notice stale information, and explain what saving will do.

Measure Success signal Stop condition
provenance recognition correctly identifies item origin repeated source/note confusion
unsupported-item recovery excludes or edits without help participant believes it is verified
save/share mental model explains destination correctly expects automatic clinician delivery
keyboard completion reaches every action in order trap, lost focus, hidden status
boundary comprehension describes organizational support interprets card as diagnosis

Do not turn task completion into proof of safety. Capture quotations only with consent, minimize research records, and provide a way to withdraw. Include screen-reader users and people using zoom or voice control; record the assistive setup without treating one configuration as universal.

Evidence versus hypothesis

The announcement supports the listed product facts. The evidence-card fields, flow, labels, and stop conditions are design hypotheses requiring research. This protocol cannot establish clinical usefulness, record correctness, informed consent quality, accessibility conformance, or what OpenAI’s current interface does.

OpenAI says Health in ChatGPT is designed to support—not replace—medical care and is not diagnosis or treatment. The card should repeat that boundary at review and preserve the person’s agency to edit, exclude, cancel, or use another preparation method. Missing evidence should stop a claim from entering “source-backed items”; extra decorative detail should not crowd out that decision.

AI assistance disclosure: This article was drafted with AI assistance and reviewed against the cited primary source.

Top comments (0)