A trainee does not necessarily learn the reasoning that produced an excellent first draft merely by reviewing it. They may learn through comparison, explanation and feedback, or they may approve a polished account without constructing their own understanding. With Heidi II, the educational question is which parts of the work remain opportunities to develop judgement.
Heidi's September 2026 launch announcement describes a broader role in preparing and carrying out work around consultations. It does not establish how the new product affects trainees' learning. The capabilities discussed here are not available in the UK or EU as of the announcement.
Which part of drafting is actually educational?
Formatting a heading and deciding which patient detail changes the interpretation are not equivalent tasks. Removing the former may free time without sacrificing much learning. Removing the latter changes the opportunity to practise synthesis.
A clinical first draft can contain several forms of thinking. The writer selects relevant information, compresses a long account, distinguishes observation from interpretation and explains why the proposed next step follows. These activities can be educational when accompanied by appropriate supervision and feedback. They are not made educational simply by being manual or time-consuming.
The practical response is therefore selective. Educators should identify the intellectual work they want learners to practise, rather than preserving every administrative burden because it happened to accompany that work historically.
A polished record can hide a trainee's uncertainty
Consider a fictional trainee reviewing a patient with persistent fatigue and several interacting practical concerns. An assisted draft organises the history elegantly and presents a coherent account of the unresolved problem. The supervisor reads it and assumes the trainee recognised which details were important.
That assumption may be correct, but the note cannot establish it. The trainee might have understood the relationships independently, recognised them only after reading the draft or accepted them without being able to explain why they matter. Those situations require different teaching responses.
A short discussion could reveal the distinction. The supervisor might ask which detail most changed the trainee's interpretation, which alternative remains open and what new information would alter the plan. This makes understanding visible without requiring the trainee to reproduce the entire record by hand.
What the mathematics experiment contributes
A large experiment reported in PNAS in 2025 examined generative AI support in high-school mathematics. Better performance while assistance was available did not reliably translate into better independent performance. Learning-oriented safeguards substantially mitigated the negative effect seen with unrestricted assistance, but did not establish a significant independent performance advantage for the guarded tutor over the control condition. See the original study.
This was not research on doctors, Heidi II or clinical documentation. Its relevance is a design warning: successful assisted completion and retained learning need separate assessment. It does not justify assuming that every AI first draft damages learning or that a tutoring label guarantees improvement.
For medical educators, the useful question is whether the chosen activity requires the learner to process the reasoning, rather than merely encounter an acceptable answer.
Preserve an independent attempt where it serves the learner
One proposed approach is a brief interpretation before comparison with an assisted draft. The learner identifies the main uncertainty, a relevant alternative and the information that would change their view. After seeing the draft, they explain where it improved their account and where they disagree.
This should be adapted to experience. A novice may need an expert worked example before they can make a useful attempt. An experienced trainee may benefit from formulating a position first. Demanding unsupported reasoning from someone without the necessary foundations is not automatically better teaching than showing a well-designed example.
Nor does every clinical task require this exercise. A protected selection of appropriate cases may be more practical than turning all live documentation into an examination. Patient care and necessary supervision should not be delayed to create an artificial educational challenge.
Feedback should reveal a change in understanding
Feedback that only corrects the note can leave the underlying misconception untouched. A learner might fix the sentence without understanding why it was wrong. We would therefore ask the learner to identify the assumption that changed and explain how it would affect a different case.
In the fictional fatigue scenario, the useful learning outcome might be recognising that chronology matters more than the order in which information was entered. A later case could present the same reasoning problem with different symptoms and a different social context. Repeating the identical case would be a weaker test of transfer.
The supervisor should also make uncertainty legitimate. A polished assisted note can create pressure to sound more certain than the learner feels. An educational workflow should reward appropriately identifying what remains unresolved rather than merely matching the fluency of the generated text.
Assess learning after the polished output disappears
An evaluation could compare immediate documentation quality with delayed performance on unfamiliar simulated cases. It should include explanation of alternatives, recognition of changed circumstances and the ability to identify an unsupported inference.
The comparator matters. Learners should retain the usual teaching and reference support appropriate to the assessment. A test that unexpectedly removes all ordinary resources may measure unfamiliar examination conditions rather than the effect of drafting assistance.
Assessors should record the actual learning activity as well. Reading an assisted draft, editing it, discussing it with a supervisor and explaining its reasoning are different exposures. Reporting them all as "AI use" would obscure the mechanism being studied.
These are proposed educational assessments, not results from a Heidi II trial. No evidence in the launch materials reviewed on 29 September 2026 establishes that the product improves or impairs these outcomes.
Separate the service benefit from the training responsibility
An assisted draft may be valuable because it reduces unfinished documentation. That service benefit does not automatically demonstrate that trainees are receiving the reasoning practice they need. Conversely, preserving educational work does not require denying learners every useful tool.
A training programme can acknowledge both responsibilities. It can use assistance where appropriate while protecting selected opportunities for independent interpretation, expert modelling, feedback and later application. The important question is whether those opportunities occur reliably, not whether every note begins with a blank page.
For readers choosing learning support, this article is published by iatroX and includes its educational approach in the discussion. Per iatroX's September 2026 product description, the Socratic Tutor starts from an attempted question and uses targeted follow-ups to explore the learner's misconception. That is a learning design, not proof that it prevents deskilling or certification to supervise Heidi II. A workplace drafting system, a human supervisor and a separate learning tool address different needs. For a trainee whose uncertainty is specific to a real encounter, workplace supervision is the appropriate starting point. For an eligible clinician seeking documentation support, assess the actual drafting workflow. For a recurring examination or knowledge gap, question-based learning can provide a separate opportunity to practise, without replacing supervision.
Frequently asked questions
Does an AI-written first draft stop trainees learning?
Not necessarily, because comparison and explanation may themselves support learning. The important question is whether the trainee processes the reasoning and can apply it later without simply reproducing the draft.
Did the PNAS study involve medical trainees?
No, the 2025 experiment concerned high-school mathematics. It offers an educational design warning, not direct evidence about Heidi II or doctors.
Should every trainee write every note without assistance?
No, that would confuse preserving learning with preserving all manual work. Educators should protect selected reasoning activities and assess understanding while allowing appropriate support for patient care.
Develop reasoning from attempted questions with the Socratic Tutor →