Clinicians could feel more capable when assistance makes difficult work easier, without improving their ability to detect an incorrect proposal. That is a hypothesis to test, not an established effect of Heidi II. Confidence in completing a task, confidence in the system and demonstrated independent competence should be measured separately.
Heidi's September 2026 launch announcement describes agents taking on work around consultations. It does not establish how users' self-assessment changes. As of the announcement, the new capabilities are not available in the UK or EU.
A smoother experience can mean several different things
A clinician might feel more capable because the system has genuinely improved access to relevant information. They might also feel more comfortable because a difficult case now arrives in a coherent format. Those experiences can feel similar even when the clinician's ability to challenge the output differs.
Comfort is not inherently misleading. A well-organised workflow can remove unnecessary friction, and appropriate confidence can help someone act decisively. The problem arises when the user attributes an assisted success to a capability they have not independently demonstrated.
An evaluation should therefore avoid treating increased confidence as either a benefit or a warning on its own. It needs to establish what the confidence refers to and how well it tracks performance.
Separate trust in the tool from confidence in yourself
"I expect this system to perform this task reliably" is a judgement about the tool. "I can recognise when its result is wrong" is a judgement about supervision. "I can perform the underlying reasoning without this assistance" is a third claim.
A clinician may reasonably hold different levels of confidence in each. They might rely on a system for document handling while remaining cautious about its interpretation of conflicting records. They might also understand the clinical issue well but be unfamiliar with the connected software's completion states.
The useful objective is calibrated reliance, not maximum trust. A user who limits delegation after recognising a weakness may be responding appropriately. A user who reports unwavering confidence despite repeated consequential mismatches should not automatically be counted as a successful adopter.
What educational research can and cannot tell us
In the 2025 PNAS mathematics experiment, students' perceptions of learning did not consistently match their independently measured performance. Some felt they had benefited more than their later unaided results supported. The study concerned school mathematics, not clinicians or Heidi II, so its relevance is a question about measurement rather than evidence of a clinical effect. See the original experiment.
The clinical implication to investigate is whether users distinguish successful assisted completion from retained understanding. A satisfaction survey or a question about feeling more effective cannot answer that on its own.
It would also be wrong to infer that lower confidence is always preferable. A clinician who becomes more accurate and appropriately more confident may have gained something valuable. The concern is mismatch, not confidence itself.
A fictional case of misplaced attribution
Imagine a doctor reviewing a general evidence question about a proposed service pathway. An assisted answer is clear and cites a relevant study. The doctor feels reassured and becomes more confident in the proposed direction.
During a later review, the doctor is asked whether the study population resembles the intended population. They cannot explain the distinction and had not considered whether it mattered. Their confidence increased because the answer was coherent, but the ability needed to judge applicability was not demonstrated.
This is a fictional teaching scenario, not a claim that Heidi II supplies unsuitable evidence. The same scenario could arise with a colleague's summary, a textbook passage or an iatroX answer. The educational opportunity is to make the basis of confidence explicit before accepting the conclusion.
Record confidence at meaningful points
A proposed assessment would ask participants for a judgement before assistance, after reviewing the output and after inspecting the originating source. Each question should specify what is being rated: correctness of the proposed action, adequacy of the available information or ability to detect an error.
The study could then compare those ratings with independently scored performance. Cases should include correct assistance, subtly incomplete assistance and situations in which the available information does not justify a firm answer. This avoids rewarding indiscriminate doubt or uncritical agreement.
Confidence should also be linked to the actual decision. A broad statement such as "I am comfortable using AI" is not equivalent to confidence that a particular action fits a particular case. The latter is the more useful unit for studying calibration.
Feedback should address the mismatch
When confidence exceeds justified certainty, feedback should identify the missing check rather than merely tell the clinician to be less confident. In the fictional pathway example, the learning point is the unexamined population difference and why it could matter.
When confidence is lower than performance warrants, feedback should identify what the clinician assessed correctly. Otherwise an evaluation could make users more hesitant without making them more capable. The aim is better alignment between evidence, judgement and confidence.
A useful follow-up case would change the setting while preserving the same reasoning issue. The clinician would need to decide which source features support transfer and which limit it. Improvement would be demonstrated by better interpretation and appropriately adjusted confidence, not by more assertive language.
Avoid confusing an attractive interface with competence evidence
A fluent explanation, reassuring status message or familiar layout may improve the experience of using a system. None independently demonstrates that the clinician can supervise the work. Equally, a difficult interface does not make the underlying reasoning more rigorous merely because it demands effort.
The evaluation should ask whether the presentation helps the user identify uncertainty and inspect the basis of the proposed action. It should also examine whether users can distinguish an explicit instruction from an inferred next step. These are practical ways to connect confidence to something inspectable.
No confidence-calibration study of Heidi II has been conducted for this article. Any results should come from an actual documented assessment, with the tasks, confidence questions and scoring criteria reported. General educational findings should not be repackaged as a product-specific safety result.
What a useful outcome would look like
For clinical leaders, a useful outcome is not simply that staff feel more empowered. It is that confidence tracks appropriate decisions and that users become more capable of recognising when help is needed. For educators, the priority is feedback on the reasoning behind confidence. For suppliers, it is an interface that makes uncertainty understandable rather than merely making completion feel effortless.
These objectives can support adoption rather than oppose it. A clinician who understands where the system helps and where scrutiny remains necessary can use it more deliberately. That is a stronger proposition than encouraging either blanket distrust or confidence unsupported by performance.
Frequently asked questions
Has Heidi II been shown to make doctors overconfident?
No such product-specific finding was identified in the public materials reviewed on 29 September 2026. This article proposes how confidence and performance could be evaluated together.
Is increased confidence a bad outcome?
Not when it reflects improved performance and an appropriate understanding of uncertainty. The concern is confidence that becomes disconnected from the ability to judge the particular task.
How can a clinician check whether confidence is justified?
Identify the evidence and assumptions supporting the decision, then consider what would change it. In educational settings, compare that judgement with feedback and later performance on a different case.