AI alongside the interaction checker: what it adds and what it does not

This site's interaction checker is not "an AI that answers drug questions". It is a dataset of clinical rules reviewed by a pharmacist, and AI only steps in to resolve the drug pair that dataset does not cover. It is a different architecture from a generic chat, and not knowing that changes how you read every result the tool shows — because the two sources look identical on screen and only the origin label tells them apart.

The whole thing in one sentence. When the checker says "checked, no notable interaction" about a pair covered by the dataset, it is a statement backed by a human review written in advance; when the same sentence comes out of a real-time call to AI, it is a probabilistic statement about that specific case — and the two read exactly the same on screen, with no warning at all telling them apart at a glance.

The three possible origins behind the same result

Origin	What it actually means
Rule from the curated dataset.	Somebody has reviewed that drug pair — or that group of drugs, sharing a mechanism — and written the rule by hand, with its severity and its specific "what to do". It is the strongest of the three sources.
"No interaction", by coverage.	Both drugs are on the list of reviewed profiles and no rule crosses them against each other. It is not that AI looked and found nothing: it is known there is nothing, because both profiles are already reviewed.
AI answer, for lack of a rule.	Neither the pair nor at least one of the two drugs is covered, so a model is queried on the spot. It is the only one of the three that carries the hallucination risk from that lesson, and the site marks it explicitly so it stands apart from the other two.

How it shows on screen, and why it is worth looking at. The result says where each finding comes from: whether it is from the "curated clinical dataset" or "not in our database, AI-assisted lookup, verify in the SmPC". It is not decoration: it is the difference between a reviewed source and a model's answer, and it only registers if you read that label instead of just looking at the result itself, which is the part most people skip when they are in a hurry at the counter.

Why "being covered" is not the same as "being mentioned"

There is a specific trap it is easy to fall into thinking about this, and it is worth naming separately because it is not intuitive: a drug appearing in some rule in the dataset does not mean that drug is "covered" in general. Coverage is an explicit list of profiles a pharmacist has thoroughly reviewed — "we have checked this drug against the relevant risk families" — not an inference that it shows up somewhere.

The difference matters because a drug can appear in a single rule — say, its best-known interaction with an anticoagulant — without that meaning it has been reviewed against the rest of the risk families: sedatives, serotonergic drugs, other anticoagulants with a different mechanism. Deriving general coverage from a single mention would be convenient and would give a sense of safety that does not match what has actually been checked.

And there is an even subtler piece: when two drugs belong to the same risk family — two anticoagulants with different mechanisms, two benzodiazepines, two ACE inhibitors — the dataset has specific rules for crossing members of the same family against each other, because "each one on its own is well monitored" is not the same as "the two together have never been crossed". It is exactly the kind of combination a superficial check would miss, and it is where this site's curated dataset invests the most care.

A full example, and what is wrong with it

A combination of three common drugs is checked for an elderly patient: an ACE inhibitor, a loop diuretic, and an over-the-counter NSAID.

What the checker returns

ACE inhibitor + NSAID: a Major finding, from the curated dataset — reduced antihypertensive effect and risk of worsening renal function, with its specific "what to do".

Loop diuretic + NSAID: also a finding from the curated dataset — antagonism of the diuretic effect and the same added renal risk, Moderate severity.

ACE inhibitor + loop diuretic: "checked, no notable interaction", marked as a pair covered by reviewed profiles, not as an AI answer.

And now, what is wrong with reading all three results the same way:

The two findings involving the NSAID are the strong part of the result, and they come from the dataset. It is exactly the kind of combination that is covered well: a documented class effect, known severity, a concrete "what to do" — review whether the NSAID is actually needed, not just a warning.
The "no interaction" for the third pair is NOT a negative AI finding, it is a consequence of both drugs already being reviewed separately. The distinction matters because a coverage-based negative is far more reliable than a negative from a one-off model call — and here the site marks it that way on purpose.
If any of the three had been a very recently launched drug with no profile in the dataset, the pattern would have been different. That third pair would have carried the AI label, not the coverage one, and that would call for reading it with more caution and cross-checking it against the SmPC before treating it as final — which is exactly what the label is there to warn about.
The overall result is still reliable, and for the right reason. Not because "AI checked everything", but because most of what matters in this specific case comes from the curated dataset, and extra caution only needs applying where the origin calls for it.

And if your pharmacy is not like that

If you work a lot with recently launched drugs. You will see the AI label more often than average, simply because the curated dataset takes time to add new profiles. It is not a flaw in the tool: it is the logical consequence of reviewing a profile taking real human time, and it is worth cross-checking those specific cases against CIMA with a bit more care than the rest.

If your patients often take supplements — vitamins, minerals — alongside drugs. Many supplement-with-supplement combinations are already explicitly covered in the dataset precisely because they are so common — two of the site's most-queried pairs are exactly this type — so a "no interaction" there is usually coverage-based, not AI. Check the label if you are unsure, do not assume it without looking.

If the query crosses many drugs at once — reviewing a heavily polymedicated patient. With ten drugs there are dozens of possible pairs, and it is common for all three origins to coexist in the same result: some from the dataset, some from coverage, and the odd pair resolved by AI. Pay special attention to whether any pair comes back "unchecked" — a fourth state, distinct from the three in this table — because it means it could not even be queried, not that it is clean.

If you use the checker to train somebody new on the team. This is a good moment to explicitly teach this distinction, because somebody who does not know it reads the whole result as if it came from the same source with the same reliability, which is exactly the misunderstanding this lesson exists to prevent.

If you would rather ask a generic chat directly instead of using the site's checker. You lose exactly the part that is worth the most: the curated dataset and the origin distinction. A generic chat answers you with the same confident tone whether the pair is well documented or not, with neither of the two signals that exist here.

If you serve a lot of patients on anticoagulation. Pay special attention to pairs within the same risk family — two anticoagulants with a different mechanism, an anticoagulant and an antiplatelet — because they are exactly the kind of combination a surface-level, drug-by-drug check tends to miss. The curated dataset invests special care there for exactly that reason.

Why this matters more here than in most other AI-assisted tools

Plenty of tools mix curated data with an AI fallback without ever telling you which is which, and most of the time that does not cost much: a wrong restaurant recommendation is a bad evening, not a clinical decision. An interaction checker sits at the other end of that scale, so the site treats the origin label as a first-class part of the result, not an afterthought — it is designed into the interface precisely because this is one of the few tools on the site where mixing the two up has a real cost attached to it.

When it does not work first time

I do not see any origin label on the result I get.

Check whether you are looking at the summary or the detail of each finding: the label usually sits next to each specific pair, not in the overall result headline. If it truly does not appear anywhere, treat the result with the caution of the least reliable origin — as if it were AI — until confirmed otherwise.

Can I trust a coverage-based "no interaction" more than one from an explicit rule?

They are not comparable that way: an explicit rule tells you something concrete — there is or is not an interaction, and why — while coverage only tells you both profiles are reviewed separately and no rule crosses them. Both are reliable, but they answer slightly different questions, so treating them as interchangeable blurs a distinction worth keeping.

Why isn't everything just in the curated dataset already, with no need for AI at all?

Because the universe of possible combinations is enormous and grows every time a new drug launches, while reviewing and writing a rule takes real human time. The site prioritises covering the most-queried pairs first — it is measured, documented work — and uses AI as a safety net for what has not been covered yet, not as a substitute for the dataset.

A pair that used to show as AI now shows as curated dataset. Did something change?

Yes: that is exactly the coverage-expansion process working as intended. The more pairs get reviewed and added as explicit rules, the fewer cases fall back to AI by default — it is a progressive improvement of the tool, not an error that it used to be different.

Does this dataset-versus-AI distinction exist in other tools on the site too, or only in the checker?

The drug lookup and the pregnancy-and-breastfeeding tool follow the same principle: reviewed data first, and only if it is not there, AI as a backup with its own warning. It is the same criterion applied in several places, not an exception unique to this one tool.

What if two pairs give results that seem to contradict each other?

Check the origin of each one first before assuming an error: it is common for a Major dataset finding to sit next to another nearby pair, in the same case, marked "no interaction by coverage" — they are not contradictory, they are two different questions with different answers. If they genuinely seem incompatible with the same origin, that is a case to check directly against the SmPC before acting on either one, rather than trying to reconcile them by guessing which of the two is more likely to be right and proceeding on that guess.

Before moving on

I know the checker has three possible origins for any given result, not just one.
I read the origin label on each finding, not just the finding itself, every single time.
I know "no interaction by coverage" is not the same as "no interaction by an explicit rule", even though both are reliable.
I treat any pair marked as an AI answer with more caution, and I cross-check it against the SmPC before relying on it.
I know a drug mentioned in a single rule is not the same as a drug truly covered against everything else, and I do not assume general coverage from one mention.

← Revisit: medication review
Next: preparing the patient conversation →
← Back to the AI School