The failure is concrete and increasingly common, so the analysis should be too. A UK finals candidate asks an AI study tool about first-line management of a common condition; the answer arrives fluent, referenced and correct, for the United States. The student learns it, reproduces it in a UKMLA-style paper, and loses the mark, or worse, carries it toward a ward. Who was responsible? The honest answer is a distribution, not a culprit, and mapping duties across the chain, student, platform, medical school, content author, model provider, is more useful than the single-villain version, because each party's duty is the thing that party can actually fix.
The chain, duty by duty
The student's duty is verification proportionate to stakes: for consequential claims, the jurisdiction stamp is one of the three checks the five-minute workflow makes routine, /blog/does-ai-save-medical-students-time-verification-burden, and a learner who never checks country has left the commonest failure mode ungoverned. That duty is real and it is not the whole story, because it scales with what the platform makes checkable. The platform's duty is architectural: jurisdiction selection or clear jurisdiction labelling; content versioning with visible source dates; correction routes that work and are seen to work; audit trails for what changed when; and honest intended-use statements about which examinations and countries the content serves. A platform that serves US management to a user in UKMLA mode has failed a duty no amount of student diligence should have to absorb; a platform without jurisdiction awareness at all has made every international user its unpaid localisation department. The school's duty is curricular: teaching the verification and appraisal skills, and, where it recommends or procures tools, evaluating jurisdiction fit as a first-class criterion, the rubric item at /blog/how-medical-schools-evaluate-ai-education-vendor. The content author's duty, where humans author or review, is the accuracy and dating of what they signed. And the model provider's duty, for general-purpose systems, is honest capability description, which is precisely why general tools without clinical intended use should not be the sole source for guideline-flavoured claims, a boundary their own documentation increasingly draws.
The worked case, run both directions
Run the failure in reverse, an American student taught UK management by a UK-grounded tool, and the symmetry teaches the design point: neither answer is wrong medicine, both are wrong for the examination in front of the learner, and the fix is identical, jurisdiction as an explicit, visible, selectable property of educational content rather than an ambient assumption. This is why blueprint-mapped platforms carry an inherent advantage on this specific failure: content organised against a named examination has declared its country, while general study tools answer from a training distribution whose centre of gravity the learner cannot see. It is also why the responsible-platform disclosure list is short and checkable: which jurisdictions and examinations the content targets; where sources are shown and dated; how users report suspected wrong-country or outdated content; what the correction turnaround is; and what changed, visibly, when guidance moved. Platforms meeting that list have made the student's residual duty small and performable; platforms meeting none of it have priced their answers below their verification cost, whatever the subscription says.
What each reader should do with this
Students: run the jurisdiction stamp on every guideline-flavoured claim from any general tool, prefer jurisdiction-explicit platforms for examination preparation, and report wrong-country answers through whatever route exists, because the report is the system's error signal. Platforms, ourselves included: treat the disclosure list as a floor, and treat wrong-jurisdiction reports as priority defects rather than edge cases, since this failure mode uniquely converts fluency into examination harm. Schools: teach the stamp, procure for it, and put one wrong-country case into assessment, a planted-error critique station, so the skill is examined rather than assumed. Responsibility distributed this way is not diluted; it is allocated to where each fix lives, which is the only version of accountability that actually reduces the failure's frequency.
Frequently asked questions
Does a disclaimer shift responsibility to the student?
A disclaimer without jurisdiction architecture is a liability costume: the platform duty is design, labelling, dating, correction, and boilerplate acknowledging errors possible does not perform it.
What about doctors rather than students?
The same chain with higher stakes and a settled top layer: clinical responsibility remains the clinician's, which is exactly why grounded, jurisdiction-correct tools matter more after graduation, not less.
Is this a reason to avoid AI study tools?
It is a reason to choose them by jurisdiction discipline: the failure is preventable by architecture plus a five-second check, and the tools built for your examination have already done the heavy half.
Could contract or consumer law reach these failures eventually?
Plausibly, as educational AI matures and intended-use statements harden; students should not wait for it, because the verification habit and jurisdiction-explicit tooling solve today what liability doctrine may address in years.