Spot the AI Hallucination: Can You Tell When It's Making It Up?
An interactive quiz. Judge five AI answers as grounded or hallucinated, learn why a clinician cannot reliably tell from the text alone, and what actually catches it before it reaches a chart.
In this article
An AI hallucination never announces itself. It reads exactly like a correct answer: fluent, specific, and delivered without a flicker of doubt. That is what makes it dangerous, and it is why trust, not raw capability, is usually the real blocker to AI adoption.
Here is the uncomfortable part, and the whole point of this quiz. You often cannot tell a hallucination from the text alone. Neither can your team. Below are five answers a business AI assistant might give. Each one tells you what the assistant actually had in front of it. Judge whether the answer is grounded in that, or made up.
Scenario 1 of 5
The chart is open. You ask when the last HbA1c was done. It answers: "HbA1c was drawn on 3 March, result 7.2%."
Grounded, because a real source, the result in the chart, actually exists for this claim. What makes it grounded is not that it sounds specific, it is that you can open the result and check it. Grounded means checkable, not automatically correct, so open it anyway.
Scenario 2 of 5
The assistant has no chart in its context. You ask about the patient's history. It replies: "The patient has a long-standing history of hypertension, managed since 2014."
Hallucinated. There was nothing in its context to support this, so the confident detail is invented to fill the gap. The tell is not the wording, which is flawless and sounds exactly like a chart. It is that no real source was available for it to draw on.
Scenario 3 of 5
You ask whether scribes pay off quickly. It answers: "Studies show 73% of clinics using AI scribes see gains within the first month."
Hallucinated. An authoritative-sounding statistic with no named, traceable study behind it. If you cannot follow a number to a real publication, treat it as invented, no matter how quotable it sounds. Real figures exist for this, from OntarioMD and Canada Health Infoway, and they do not look like this.
Scenario 4 of 5
No consult letter was ever provided. The assistant answers: "Per the cardiology consult of 12 June, the patient was cleared for surgery."
Hallucinated, and this is the dangerous one. It cites a specific source with a specific date, which feels trustworthy, but that letter was never in front of it. Models fabricate citations as fluently as they fabricate facts. A named document is not proof of grounding. A source you can actually open is.
Scenario 5 of 5
The medication list is in context. You ask what the patient is taking. It answers: "Metformin 500mg BID, ramipril 5mg daily, and atorvastatin 20mg nightly."
Grounded, because the list is genuinely there. But grounded does not mean current: the source may include a medication stopped four months ago that nobody reconciled. The source is real, so this is grounding, and you should still check the dates. Grounded and correct are two different checks, and in a medication list the second one matters most.
Why you probably missed at least one
If a few of those fooled you, that is the lesson, not a failure. The fabricated consult cites a specialty and a date. The invented history is fluent and reads exactly like a chart. Nothing in the wording separates them from the real answers. That is exactly why “just read it carefully” is not a strategy, least of all for a clinician confirming their eleventh note of the evening.
Grounded vs hallucinated: what actually separates them
The real distinction is not how an answer sounds. It is whether the claim is tied to real evidence the system actually retrieved, and whether you can trace it back to that evidence yourself.
Two ideas worth keeping separate here. Faithfulness is whether the answer is consistent with the source the model was given. Factuality is whether it is true in the world. Grounding buys you faithfulness, the answer sticks to real retrieved context, which is a huge win. It does not automatically buy you factuality: the retrieved document can be outdated, or the model can miscount over it. So grounding is the foundation, not the finish line.
”Your diabetes protocol (retrieved, linked, updated 14 Jan) states a three-month interval for this HbA1c range.”
Links to the exact passage it pulled, with the date, so a clinician can confirm in one click. Still worth checking the protocol is current, and now you can.
”Per the clinic protocol, this follow-up interval is six months.”
Names a source, sounds authoritative, and you cannot open it or check it. This is how a fabricated citation looks, right up until it lands in a chart.
What actually catches hallucinations
Since you cannot eyeball them reliably, the answer is not a smarter reader, it is a system that checks itself. The levers that actually move the needle:
- Retrieval quality. Most hallucinations are really retrieval failures. If the right chart entry never reaches the model, it fills the gap with something plausible. Better retrieval prevents more hallucinations than a bigger model does, and in a clinic it is usually the cheapest fix available.
- Provenance you can click. Every fact should link to the chart entry and date it came from, so a clinician can verify in one glance instead of trusting the tone.
- A verifier step. A second model or a set of rules can check whether the answer is actually supported by the retrieved context before it ever reaches a user.
- A clinician in the loop on anything that reaches a chart, and evals against a known-good set of your own encounters so quality is a number you track, not a feeling you hope for.
Verification, not vibes
The grounding check
Is every claim actually supported by what was retrieved, before a user sees it?
An answer you can verify, not just believe
Why this matters more than it seems
One hallucination caught early is a shrug. One that slips into a chart, a referral letter, or a set of discharge instructions is how trust in an entire clinic tool collapses in an afternoon, and it can also hurt somebody. Practices rarely abandon AI because the model was incapable. They abandon it because one confident, wrong, well-cited note taught every clinician in the building to stop believing the rest.
The fix is not hoping the model never makes something up. It is building the checks that catch it before anyone has to, so your clinicians are not running this quiz on every note at nine at night.
Worried the tool your clinic runs might be confidently, plausibly wrong somewhere nobody is looking? That is exactly what grounding, provenance, and evals are built to catch. Book a free consult and we will look at where yours could be hallucinating and how to make every fact traceable to a chart entry. See more on PHIPA and Guardrails.
Frequently asked questions
What is an AI hallucination?
An AI hallucination is when a model produces a confident, fluent claim that is not supported by any real source, whether or not it happens to be true. The danger is that it reads exactly like a correct answer, and in a clinic it reads exactly like a correct chart entry. Fluency and even a cited source are not proof of accuracy, which is why nobody can reliably spot one by eye at the end of a long list.
How do you detect AI hallucinations?
Not by reading the answer, which is the trap. You detect them by tracing every claim back to the chart entry it came from, with its date, so a clinician can check in a glance rather than re-read. Citations that link to the exact source, and evals against known-good cases from your own practice, are what actually catch them. A confident tone or a named document is not evidence, because models fabricate citations as fluently as they fabricate facts.
Can AI hallucinations be fixed?
You cannot eliminate them, but you can sharply reduce how often they happen and how much damage they do. Grounding answers in the real chart, showing click-through provenance on every fact, using a second check on claims, keeping a clinician in the loop on anything that reaches a record, and measuring with evals all shrink the risk. The goal is not a perfect model, it is a system that catches its own mistakes before a person has to.
Why do AI answers sound so confident even when they are wrong?
A language model is trained to produce fluent, plausible text, and its tone is not tied to whether the claim is true. Newer models can surface some uncertainty, through confidence signals or by abstaining, but those signals are imperfect and easy to over-trust. The safe assumption is that confidence tells you nothing about correctness.
Found this useful?
Share this with your network on LinkedIn, it helps more than you think.
Enjoyed this read? Get the next one in your inbox.
When there is something worth your time, you will be first to know. No spam, unsubscribe anytime.
Keep reading
Who Owns This Once It Is Live? The AI Job Nobody Budgets For
Ordinary software fails loudly and someone files a ticket. AI fails quietly and keeps answering. That difference is why a clinic AI system needs a named owner, a number, a cadence, and the authority to switch it off, and why most practices discover they have none of the four.
Read article
What Do We Tell the Partners and the Staff? Announcing a Clinic AI Project Without Starting a Panic
The same AI project reads as a threat or a plan depending entirely on how it is announced. Four facts make the difference, and they are the same four at the partners' meeting, at the staff huddle, and in anything a patient ever sees.
Read article
What You Should Refuse to Automate
The hype says automate everything. In a clinic the discipline is knowing where the line goes, because the worst axis is always the one with a patient on it. A task belongs to a person, not a model, when it is hard to undo, needs real judgment, or puts health, money or safety at stake. Grade any task on those three axes here.
Read articleHave software that should be smarter?
Let’s map a free AI-transformation roadmap for your product.