A real-looking source for the wrong outcome
Fictional example: Mira asks a chatbot whether a particular learning method improves recall. The answer cites a scientific study. In this explicitly invented scenario, Mira finds the article, but it examines only how enjoyable people found the method. That is a different question. She removes the reference from her summary. She does not need to prove the method ineffective or dismiss the whole article; the paper simply does not establish the benefit claimed in the chatbot’s response.
Identify the publication first
Compare the title, authors, year and journal or publishing organisation. A DOI is a persistent identifier for a publication, not a quality seal for its conclusions. Crossref provides a tool for matching references to DOI records. PubMed can help locate biomedical literature using the title and other citation details. Still check the results you find: similar titles, preliminary versions and later publications can be confused. Not every legitimate source has a DOI, and not every relevant publication is indexed in PubMed.
Compare the content with the actual claim
Read beyond the headline. What question was investigated, and which people, materials or situations were involved? A laboratory result does not automatically become advice for people. An observed association does not by itself establish a cause. Notice whether a cautious “might” has become a confident “works” in the AI answer. For important statements, read the original results and limitations. If only an abstract is available, say so explicitly instead of giving the impression that you assessed the full methods and discussion.
Handle missing links and contradictions carefully
A broken link can result from a website move, a typing error or access restrictions. Search the title and authors on the publisher’s or organisation’s site as well. If matching details remain unavailable, the reference is unconfirmed; inaccessibility alone does not prove invention. NIST nevertheless identifies fabricated or misleading citations as a possible generative-AI problem. Further unchecked references from the same conversation do not resolve that problem. A library or knowledgeable professional may help when a publication is difficult to locate.
Report what you actually established
Keep three brief notes: the claim, the original passage you found and any remaining limitation. Link to the publication you checked, rather than to an invented summary attributed to it. If the evidence supports only part of the statement, narrow your wording. For health and other consequential decisions, a genuine study still does not establish that an intervention is suitable for you personally. This page teaches source checking and is not medical advice. Do not change treatment yourself because you located a reference.
Your checklist
The checkboxes are for your own reference. No entries are sent to us. To keep a copy, print the page or save it as a PDF using your browser menu.
Common questions
Does a working DOI prove quality?
No. It helps identify a publication. The reliability of its methods and conclusions, and its relevance to your question, need separate assessment.
Is a source invented if I cannot find it?
Not necessarily. Details may be incorrect, the title may differ or access may be restricted. Until clarified, do not use it as a confirmed reference.
Is the abstract enough?
It can provide an initial overview, but may omit important methods and limitations. Make clear when you have not checked the full publication.
Sources and further information
Primary sources from public authorities, consumer organisations and providers. A source may be in another language; follow the guidance that applies to your service and where you live.
- NIST: Generative Artificial Intelligence ProfileSource review date:
- Crossref: Simple Text QuerySource review date:
- National Library of Medicine: PubMed HelpSource review date:
