Checking AI content: is it true, and is it good evidence?

About 30 minutes · plain language

An AI answer is only as good as the data and sources behind it. If the data are wrong, old or one-sided, the answer will be too, however fluent it sounds. This lesson gives you a simple, evidence-based routine for checking anything AI tells you, or anything you read online, before you act on it or pass it on.

Why AI answers can be wrong

A language model learned from a huge amount of text. Some of that text was wrong, out of date or biased, and the model cannot tell which. It also does not look facts up unless it is connected to a search tool. It writes the most likely answer. So an answer can be:

• Made up: invented facts, numbers or references (hallucinations).
• Out of date: based on information from before its knowledge cut-off.
• Biased or one-sided: repeating what is common in its data, not what is best supported.
• Right in general but wrong for your case: a correct adult dose given for a child.

Good data in does not guarantee a good answer, but poor data in almost guarantees a poor one.

Step 1: Stop and name the claim

Before you use or share anything, write the claim in one sentence. “Chlorhexidine bathing reduces bloodstream infections in ICU patients.” A clear claim is easier to check than a whole paragraph. Notice how the answer makes you feel: a confident tone or an answer you were hoping for are reasons to check more, not less.

Step 2: Ask where it came from

Does the AI show a source? Is it a primary source (the study, guideline or law) or a secondary one (a summary)? Is the tool searching current sources, or answering from memory with a cut-off date? An answer with no source is a starting point for your own search, not a fact.

Step 3: Find and read the original

Use lateral reading: open new tabs and check the claim in trusted places, such as PubMed, a national or WHO guideline, the drug formulary or your institution’s policy. Confirm that each reference exists, that it says what the AI claims, and that it is current. Do not check one AI against another AI; that only checks that they agree, not that they are right.

Step 4: Judge the evidence

Not all sources are equal. Ask:

• What type of evidence is it? Guidelines and systematic reviews are usually stronger than single studies, which are stronger than expert opinion or anecdotes.
• Is it good quality? Use the appraisal tools you learned in research methods, and the AI-specific guidelines in this library for studies of AI tools.
• Is it current, relevant and from a reliable author? Simple checklists such as CRAAP (Currency, Relevance, Authority, Accuracy, Purpose) help with web pages.
• How sure are the authors? Good guidelines and reviews rate the certainty of their evidence, often with GRADE, from high to very low. Guidelines themselves vary in quality; AGREE II is a tool for judging how well a guideline was made.

An evidence ladder for any health content

One way to put Step 4 into practice is to give each source a place on a ladder. A published tool for rating the content of health apps (EBCRT-mHealth) scores sources from 5 down to 1. The ladder below is our simplified version of it:

1. Clinical guidelines (international, national, professional bodies): around 4.5 to 5
2. Systematic reviews and meta-analyses: around 4
3. Randomised trials, then other studies (cohort, case-control, cross-sectional, case reports): around 3.2 to 3.8
4. Websites of trusted institutions, specialist-reviewed sites, laboratory research, textbooks and expert opinion: around 2 to 3.1
5. News sites, Wikipedia and blogs: below 2. A source that is not about the topic at all scores 1

The ladder rates the source, not the claim. A strong source can still be misquoted, and an old guideline can be out of date, so always combine the ladder with Step 3 (does the source really say this?) and the date.

A note on language

AI tools learn from what has been written. Much less peer-reviewed health research is published in Arabic or Persian than in English, so answers in these languages may lean more on news sites and general web pages. This is a reason to check more, not to stop using your own language: ask the tool for its sources, and check high-stakes answers against the original guideline, which may be in English.

Step 5: Apply it to your patient and setting

Evidence-based practice has three parts: the best evidence, your clinical expertise, and the patient’s values and preferences. Even correct evidence may not fit this patient, this setting or local policy. The final decision, and the responsibility, stay with you.

When to check more, and when less

Match the effort to the stakes. A suggestion for a study timetable needs little checking. Anything about a patient, a dose, a diagnosis, a law, a statistic or a reference needs full checking. If you cannot verify it, do not use it, or say clearly that it is unverified.