Unit C-2 Copyright, Bias, and Verification: Etiquette Before You Use It

Related competencies: C2, C3, C4 | Estimated time: 10 minutes

Why This Skill Matters

A nurse quickly put together a training document for new staff using generative AI. The slides included an illustration created by the AI, the body text used a passage where the AI had summarized the content of a well-known textbook, and the whole thing was distributed in-house without any check on the content. This single sequence of events contains all three pitfalls at once: copyright (how to handle other people's works and AI output), bias (the skew embedded in the output), and verification (human confirmation before using something in clinical work). In this unit, you will learn, all together, the etiquette needed before promoting generative AI output from a "rough draft" to "something used in clinical work."

Core Concepts

Copyright -- Points to keep in mind about generated output and training data

• Check before using generated output as-is. The output of a generative AI can sometimes closely resemble an existing piece of text, illustration, or character contained in its training data. Before publishing or distributing it, check whether it closely resembles an existing work, and also check the terms of service of the tool you are using (conditions for commercial use and secondary use).
• Think about your purpose when entering someone else's copyrighted work. Entering a copyrighted work -- a textbook, a paper, a web page -- into an AI to produce a summary or adaptation, and then distributing that result, can raise issues of reproduction or adaptation. The legal treatment can differ between using something solely for your own study and printing/distributing it as a handout, so careful consideration (such as obtaining the rights holder's permission) is needed when the purpose is distribution.
• Verify citations against the primary source. The paper titles and guideline names an AI cites are generated from patterns it learned in the formatting of real documents, and there is no guarantee they actually exist. Always check citations against the primary source itself, confirming both that it exists and that the content matches (this is an applied case of the hallucination covered in Unit A-1).

Bias -- Skew can be mixed into the output

• Generative AI produces output that reflects statistical patterns in its training data (mainly text and images from the internet). Biases present in the training data -- involving language, region, gender, era, and values -- tend to appear in the output as well. Bias can also originate beyond the training data: in the algorithm, the training process, how the prompt is written, and any reference material provided.
• Examples: asking for "an illustration of a doctor" and getting only men; receiving explanations that assume a Western-style healthcare system; an outdated practice being presented as the "standard."
• Bias cannot be completely eliminated as a technical matter. Proceed on the premise that bias can be mixed into the output, correct for the bias you notice -- for example, by explicitly specifying conditions in the prompt (such as, "make gender and age diverse") -- and rely on human review as the final check.

Verification -- The last checkpoint before using output in clinical work

• Human verification is mandatory before using generative AI output in clinical work. Fluency, level of detail, and a confident-sounding tone are not evidence of correctness.
• Three perspectives for verification:
  1. Fact-checking -- Do the figures, dosage/administration, and recommendations match primary sources such as the package insert, guidelines, or the original reference?
  2. Fit for the intended audience -- Is the content and wording appropriate for the intended reader (patient, new staff member, student) and for conditions at your own institution?
  3. Harm potential -- Is there anything that could cause misunderstanding or anxiety, inappropriate advice, or discriminatory/exclusionary language?
• Verification is needed for every individual output. "This AI was right last time, so it must be fine this time too" does not hold.

Copyable Prompt

The following prompt has the AI itself organize the points to check when verifying your own deliverable (a draft patient-explanation document, training material, etc., that contains no patient information).

[Exercise: Self-check before use]
Below is a draft document I created using generative AI. As a self-check
before using this document in clinical work, please list the points to
check from the following three perspectives, and point out specifically
any parts that look questionable.

1. Fact-checking: Which statements need to be cross-referenced against
   primary sources (package insert, guidelines, etc.)? Please list every
   figure, proper noun, and recommendation.
2. Fit for the intended audience: Is there any wording or content that
   would not suit the intended reader [fill in the reader here -- e.g.,
   an elderly patient, a new nurse]?
3. Harm potential: Is there anything that could cause misunderstanding
   or anxiety, inappropriate advice, or biased language?

Finally, on the premise that "this check is an AI-assisted aid, and the
final confirmation must be done by a human referring to primary
sources," please select the three points a human should check with the
highest priority.

--- Draft begins here ---
[Paste your draft here. If you don't have one on hand, first ask,
"Please create a draft document of about 200 words explaining tips for
reducing salt intake, for a patient with hypertension," and use that
output as the subject of the check]
--- End ---

Steps

1. Prepare the draft you want to check. If you don't have one, first have the AI draft a patient-explanation document as shown in the prompt example (do not use any patient information).
2. Fill in the [ ] in the prompt above and send it to the AI along with the draft.
3. Among the "statements requiring fact-checking" the AI listed, actually verify the one or two highest-priority items against a primary source (a package insert, a guideline, an official organization's website, etc.).
4. Check for yourself whether the AI's feedback missed the bias angle. Notice whether any implicit assumptions about the reader's age, gender, or life circumstances have crept in.
5. The goal of this exercise is not "the AI's check passed, so it's done," but ending up with a "list of points a human still needs to confirm" in hand.

Example of Expected Output

The AI will typically return a check result like the one below (a fictional example for illustration).

1. Statements requiring fact-checking:
   - "Aim for less than 6g of salt per day" -> Needs to be cross-checked
     against the latest edition of the hypertension treatment guideline
   - "You can use reduced-sodium seasonings freely" -> May be missing a
     caution about potassium content for patients with reduced kidney
     function
2. Fit for the intended audience:
   - "Keep track of your salt intake with an app" -> For an elderly
     reader, it would be better to also mention an alternative (a
     paper tracking sheet)
3. Harm potential:
   - "If you don't limit salt, you'll have a stroke" -> An overly
     definitive statement that provokes anxiety. Should be rephrased
     as an explanation of risk instead

Points a human should check with the highest priority:
(1) the numeric target value (2) the note about reduced-sodium
seasonings (3) rephrasing the overly definitive statement

Common Pitfalls and How to Handle Them

• Feeling reassured when the AI's check comes back "no problems found": The AI's check is itself an output that requires verification. A "no problems" answer does not mean verification is complete. Always keep a step where a human checks the primary source.
• The checklist stays generic and lacks specifics: Adding concrete details about the draft's purpose, intended reader, and situation of use to the prompt makes the feedback more specific (this is an application of the context-giving skill you learned in Area B).
• Wanting to leave the copyright judgment to the AI: An AI's answer to "is this image okay from a copyright standpoint?" is not a guarantee. If you're unsure, consult your organization's relevant department or the original text of the terms of service.

Safety Note

Do not enter patient information

Do not include information about a real patient in the draft you are checking. Also, even for the purpose of practicing this check, do not enter documents whose external transmission is prohibited by contract or policy (in-house regulations, unpublished research data, etc.).

Self-Check Quiz

You can check your understanding of this unit. There is no pass/fail judgment, and no record is saved.

Deliverable Feedback

At the end of the conversation where you ran the exercise, paste and send the feedback prompt from the Feedback AI page, and the AI will send back a review. When you send it, begin with the line "This is my Unit C-2 output." Writing into the conversation the results of what you actually confirmed against primary sources will make the feedback more specific. This submission is optional and is not a requirement for issuing the Area C certificate.

Next Steps

Once you've confirmed your understanding, try the Area C quiz. The only requirement for issuing the certificate is passing the quiz, so you can attempt it even without having read the units.

References (Optional)

• 香田将英, "Ethical Challenges That Can Arise When Using Generative AI" (JSME ICT Education Committee Symposium, Session 5, 2026-04-21): Slide PDF / Recorded Talk
• 小林直人, "What Happened When Medical Students Used Generative AI for a Report Assignment" (JSME ICT Education Committee Symposium, Session 1, 2025-11-20): Slide PDF / Recorded Talk
• 村岡千種, "University Initiatives on Student Use of Generative AI" (JSME ICT Education Committee Symposium, Session 3, 2026-01-20): Slide PDF / Recorded Talk

This unit is self-contained even without consulting these resources.