When an AI explains a clinical concept, the only question that matters long-term is: grounded in what? There are five honest answers in 2026, and knowing which one you are getting changes how much you should trust the output.
The five grounding models
Model knowledge: the answer comes from what the underlying model absorbed in training. Fluent, broad, uncited by default and occasionally confidently wrong. Base ChatGPT lives here.
Open-web retrieval: the model searches and cites the live web. Traceable, but the sources vary in quality per query.
Uploaded-document grounding: the answer is built from your own files. Neural Consult's model. Highly relevant to your curriculum, and exactly as reliable as your notes.
Proprietary library grounding: answers draw on an editorially controlled corpus. AMBOSS cites into its physician-edited library, UWorld's UAsk into its own explanations and reference material, Osmosis AI into Elsevier content with links out to videos and flashcards. Quality control is strong; the trail usually ends inside the product.
Guideline-grounded retrieval: answers are built from named national guidance with direct links to the source documents. This is the askiatroX model, grounded in NICE, CKS, SIGN and SmPC information, so a claim about first-line management resolves to the actual guideline paragraph rather than to an internal article about it.
Why the difference matters for learning
For exam preparation, proprietary libraries are often ideal, since the explanation is written to teach. For clinical practice, and for UK exams that test national guidance specifically, the guideline trail matters more: "because our editors say so" and "because NICE says so, here" are different epistemic claims, and examiners test the second. There is also a skill-building argument. Reading real guidelines through an AI that links to them trains the appraisal habit you will need for the rest of a career; reading only smooth summaries does not.
Run the citation test yourself
Take one question you actually care about, something like the first-line management of a common condition in your jurisdiction, and put it to each tool you are considering. Then audit the trail rather than the prose. Does the answer cite at all? Do the citations resolve to primary guidance, an internal article, your own uploaded lecture, or nothing? Do the cited passages actually support the claim made? How does the tool behave where evidence is genuinely uncertain: does it hedge honestly or manufacture confidence?
Ten minutes of this is more informative than any review, including this one.
An honest ranking, with limits
For source transparency on UK clinical questions, iatroX leads this group because the grounding is regional guidance with direct links. For depth of US-focused editorial explanation, AMBOSS and UWorld are stronger, and their libraries go places UK guidelines do not. Osmosis wins for learners who want every citation to open a video. Neural Consult is the right choice when the source that matters is your own lecturer, with the standing caveat that it inherits your materials' errors. ChatGPT remains the most flexible and the least anchored.
Pick the grounding that matches what your exam, or your patient, will actually hold you to.
A worked example of the citation test
Take a deliberately ordinary question: first-line drug treatment for newly diagnosed hypertension in a 52-year-old. Put it to each tool and follow the trail rather than the prose.
A guideline-grounded tool should answer with the age-and-context logic current UK guidance actually uses and cite its way to the specific NICE recommendation; on askiatroX the citations resolve to the guidance itself, which you can open and check in one click. A library-grounded tool, AMBOSS say, will answer well and cite its own hypertension article; the trail is real but ends inside the product, and a UK reader must notice where its editorial line follows US convention. An upload-grounded tool answers from whatever your pharmacology lecture said, which is a feature if your lecturer was current and a trap if the deck is three years old. A pure model answer arrives fluent and citation-free, and you have no way to distinguish its 2026 knowledge from its 2023 memories.
Now push one level deeper, because this is where tools separate: ask each one what the evidence is behind the recommendation, and whether any part is contested. The grounded tools should surface the guideline's own rationale; the honest ones will flag genuine uncertainty rather than smoothing it over. A tool that manufactures consensus where none exists has told you something important about every future answer it will give you.
Ten minutes, one question you already half-know, and the trails tell you more than any feature table. Run it before you subscribe to anything, including us.
One caution as you run it: do not let citation formatting stand in for citation quality. A wall of references can decorate an ungrounded answer, and a single link to the exact guideline paragraph outweighs ten ornamental ones. The test is always whether the cited passage, opened and read, actually supports the claim made, which is a habit that will serve you long after this comparison is out of date.