A pre-visit summary could anchor a consultation if its first interpretation becomes the story the clinician mainly tries to confirm. It could also reduce existing anchoring by revealing overlooked information. For Heidi II, the important question is which presentation helps clinicians notice contradictions, hear the patient and revise their initial understanding.
Heidi's 28 September 2026 launch announcement includes pre-visit preparation among its proposed agent workflows. This article examines a possible cognitive effect of that function, not an observed defect in Heidi II. The new capabilities are not available in the UK or EU according to that announcement.
The summary can become the first account of the patient
Imagine a fictional patient whose previous consultations included episodes of distress and intermittent physical symptoms. A prepared summary foregrounds the earlier explanation and describes the forthcoming appointment as a review of the same problem. A newer entry mentions that the pattern has changed, but that detail receives less prominence.
The clinician may still conduct an excellent consultation. The question for evaluation is whether the prepared narrative changes which questions they ask first, which discrepancies they pursue and how readily they consider that today's problem may differ from the previous one.
Nothing in this scenario establishes a diagnosis or recommends management. Its purpose is to isolate a presentation problem: the record contains more than one possible account, and the summary gives one of them priority before the patient speaks. This differs from the approval problem explored in iatroX's article on independent judgement: the influence here begins before a proposed action reaches review.
Compression has to choose what remains visible
A summary cannot reproduce every detail and still function as a summary. Selection is therefore unavoidable. The useful distinction is between making information manageable and making uncertainty disappear.
Chronology is especially important. A statement may have been reasonable when recorded but superseded by later information. A concise summary should not make an older interpretation appear to be the current conclusion simply because it is repeated more often or written more clearly.
Disagreement also needs deliberate handling. A patient account, an earlier clinician's interpretation and an external letter may not align. Collapsing them into a single narrative can make the text easier to read while making the clinical problem harder to understand. A useful design would preserve consequential disagreements and let the reader inspect their origins.
Not assessed is not the same as absent
A second fictional example concerns a consultation conducted remotely. The source record does not contain a particular examination finding because that examination was not performed. A later summary should not turn the missing finding into a reassuring negative.
The concern is not limited to invented words. A summary can narrow attention by omitting the fact that a question remains open. A clinician reading a tidy account may assume a matter has been settled when the underlying encounter never assessed it.
We would therefore test whether the summary distinguishes observed findings, patient reports, working interpretations and unassessed questions. This is a proposed quality criterion, not a statement that Heidi II currently conflates them. It applies equally to a summary written by a person or another AI system.
Let the patient's account change the plan
A pre-visit brief is preparation, not a replacement for discovering what has changed. In an evaluation, the simulated patient should sometimes introduce information that does not fit the prepared account. Assessors could then examine whether the clinician explores it or steers the conversation back towards the original narrative.
The useful behaviour is not automatic rejection of the summary. It is the ability to treat it as revisable. A doctor should be able to say what would make the initial account less plausible and recognise when the patient supplies exactly that information.
A workflow might support this by presenting the appointment's stated purpose separately from a proposed clinical interpretation. It might also foreground recent changes or unresolved contradictions. These are design options to test, not verified descriptions of Heidi II's interface.
A good summary could broaden attention
The counterargument deserves equal consideration. A clinician without assistance might see only the most recent entry, overlook an external letter or fail to reconstruct a complicated chronology. A well-designed summary could bring those details together and make the consultation less vulnerable to an earlier mistaken assumption.
In the first fictional case, the summary could explicitly show that the latest pattern differs from the previous presentation. Rather than anchoring the doctor to an old account, it could provide the reason to reconsider it.
This is why a simple comparison of summary versus no summary is insufficient. The form, content and timing of the summary may determine whether it helps. The relevant outcome is better attention to the case, not whether clinicians spend less time reading.
Test presentation order and structure
A proposed simulation study could compare a narrative summary, a chronology-led brief and a version that explicitly separates unresolved discrepancies. Another comparison could vary whether the clinician receives the summary before hearing the patient's opening account or afterwards.
The cases should contain both stable follow-ups and genuine changes. Otherwise the assessment might favour constant reconsideration even where the existing explanation remains appropriate. Outcomes could include recognition of contradictions, premature narrowing, unnecessary expansion of the assessment and justified revision of the initial interpretation.
Evaluators should also measure what information participants actually saw. A missed detail hidden behind an unopened panel presents a different design issue from a visible detail that was misunderstood. Clinician experience and available consultation time should be recorded rather than assumed equivalent.
No such Heidi II comparison has been performed for this article. Any reported results should come from an actual run with defined cases and independent scoring, not from this proposed framework.
A practical review question
Before accepting a prepared account, a clinician could ask: what would make me abandon this story? The answer need not be an exhaustive differential or a lengthy duplicate assessment. It should identify the uncertainty that matters for the current encounter.
For an uncomplicated administrative follow-up, that might be whether the agreed action has already happened. For a changing clinical presentation, it might concern chronology or information absent from the earlier assessment. The level of review should fit the task.
A teaching session can use the same approach. Ask a learner to identify the strongest reason to question the summary, then compare that reasoning with the source record. The aim is not to reward scepticism for its own sake. It is to practise justified revision when new information changes the meaning of what came before.
The standard applies to iatroX's own generated explanations and case summaries as well. A fluent starting account should remain open to challenge, particularly when the patient or source material tells a different story.
Frequently asked questions
Has Heidi II been shown to cause anchoring bias?
No product-specific evidence establishing that effect was identified in the materials reviewed on 29 September 2026. This article describes a plausible mechanism and a way to test it.
Could an AI summary reduce anchoring?
Yes, that is a plausible benefit when it exposes overlooked information, chronology or contradictions. The effect should be evaluated rather than assumed in either direction.
What should a clinician check in a pre-visit summary?
Prioritise changes over time, unresolved disagreements and distinctions between observed, reported and unassessed information. The depth of review should match the consequence of relying on the summary.
Practise revising your interpretation through clinical cases →