Gains:
- Ability to use artificial intelligence for literature summaries and scientific writing drafts and recognize the risk of fraudulent citations (fake articles, wrong DOI)
- Ability to apply the habit of verifying each citation in the primary source via DOI/PubMed and preventing the risk of plagiarism/fabricated citations
- Ability to use artificial intelligence as a drafting and language tool that accelerates writing and adopt the principle of leaving scientific responsibility to the author
Molecular biology and genetics rely on a huge literature, with hundreds of thousands of articles added each year. The clinical significance of a variant, the function of a gene, or the validity of a method all depend on published evidence. Artificial intelligence (AI) is a powerful aid in screening, summarizing, and drafting scientific texts for this literature—but it is also the field's most dangerous source of hallucinations: AI can forge articles that never existed, in real journals, with real-looking authors, with well-formed bylines. In this unit you will learn how to use AI safely in literature and writing and how to verify every citation.
Critical rule: Do not use any of the references given by YZ without personally opening and verifying the source. Do not trust any byline without a DOI (permanent digital ID of a publication), PubMed ID (PMID) or direct publisher page. An AI-delivered tag may seem “too realistic”; realism is not reality.
Two faces of AI in literature
Useful side: Quickly summarizes a topic, simplifies a complex article, suggests search terms, corrects the language of a manuscript, compares conflicting findings.
Dangerous aspect: Fabricates a non-existent article (fake citation), distorts the findings of a real article, presents outdated information as if it were current, "applies" a source to a claim.
So use AI as a starting point in your literature review, not as the final word. The actual search is done in authoritative directories such as PubMed, Google Scholar, Europe PMC; AI's summary does not replace this search.
Step by step: secure literature workflow
1. Clarify the question and extract search terms with AI. Ask AI "what keywords and MeSH terms (PubMed's topic tags) should I search for on this topic?" ask.
2. Do the actual search in the authoritative directory. Search it yourself on PubMed/Scholar; Don't rely on AI's "there are these articles" list.
3. Have AI summarize the articles you find. It is safe to give the text (or summary) to the AI and have it summarized; Because you have the resource. Asking AI to narrate an article "from memory" is dangerous.
4. Verify each citation with DOI/PMID. Confirm that every source included in your text actually exists and supports the claim.
5. Use AI in writing for language and structure, not for fact generation. AI is good for sentence fluency, paragraph organization, summary outline; but the results, numbers and citations come from you.
Tip: The quickest way to verify a citation: open the given DOI via doi.org or search PubMed with the title in quotes. If it doesn't open or the title/author doesn't match, that citation is most likely made up.
three mini cases
Case 1 — Three out of five references are fake. A graduate student asked for an introductory paragraph and 5 citations from AI. When he checked the DOIs one by one, he saw that 3 of them were not opened anywhere, 1 belonged to another article, and only 1 was correct. The paragraph was fluent, but most of the sources were fabricated.
Case 2 — Distorted finding. A researcher had an AI summarize a real article; YZ wrote that the article said, "Gene X causes disease Y." When the researcher opened the article, he saw that it actually said "Gene X was found to be associated with Y, causality was not shown." The difference between association and causation was scientifically critical.
Case 3 — Confirmation gained. When a doctoral student searched PubMed using YZ's suggested MeSH terms, he found a recent review that was central to the topic, but was never mentioned in YZ's abstract. The AI generated the search terms (useful), but the real discovery was in the authoritative index.
Four copyable templates
1) Search strategy:
Your role: literature review assistant. Suggest search terms, MeSH terms, and Boolean (AND/OR) combinations to use in PubMed for the following research question: [question]. Article LISTING; Just set up a good search query. I'll do the search on PubMed.
2) Summarizing the given text (I have the source):
Summarize the article text/abstract I pasted below: [text].Distinguish between: key finding, method, sample size, constraints.DO NOT MIX "correlation" and "causation" claims; Quote exactly what the article says. Don't add anything that isn't in the text.
3) Citation verification checklist:
I'll give you this list of references. The fields I need to verify for each (DOI, PMID, journal, year, title-author consistency) appear as a checklist. YOU cannot verify that the references are real; I will confirm with DOI/PMID, give me the checking steps.
4) Scientific language correction:
Correct the language and flow of the following paragraph, but use facts, numbers, and references. CHANGE, ADD, DELETE: [paragraph]. Improve only in terms of grammar, fluency and scientific style. Mark each sentence you changed.
Weak prompt / Strong prompt
Weak: “Write a 5-source introduction on the role of CRISPR in cancer therapy.”
Problem: AI writes from memory and likely produces spurious citations; You don't see the source.
Strong: "Suggest search terms I should use in PubMed on CRISPR and cancer therapy. Then I will give an abstract of 5 articles that I will find and paste; draft an introduction based only on them, linking each sentence to the relevant article. Do not add any claims that are not in the text."
Why it's powerful: Attributions are real and under your control; AI only works with the text it sees.
Quest
Is AI safe?
condition
Suggest a search term
Yes
You make the call
Summarize given text
Yes
Let the source be yours.
List articles from memory
no
There is a high risk of false citations
Language/style correction
Yes
Do not change the fact/reference
transfer findings
careful
Relationship/causation distinction
Common mistakes
- Using citations without verification. The most common and most damaging mistake; False citation is a violation of academic integrity.
- Requesting articles from memory. AI's "there are these studies" list may be fabricated.
- Mistaking the relationship for causality. AI can erase this distinction when summarizing; return to the source.
- Assuming timeliness. The AI's knowledge is cut off at a certain date; Get the latest findings from the index.
- Leaving fact generation to AI. The number, result and claim must come from you, not from the AI.
Caution: If you used AI while writing an article, check your journal and institution's AI use policy. Many journals stipulate that AI cannot be an author and that its use must be transparently stated. False citation is a reason for withdrawal and academic sanctions.
Depth: catching a fake quote in 30 seconds
Most fake attributions collapse quickly with a few mechanical checks. Develop a validation reflex step by step. First, open the DOI via doi.org. A real DOI takes you directly to the publisher page; If you get a "DOI not found" error, the citation is almost certainly fake. Second, search PubMed for the title with quotes. If the title doesn't appear at all or doesn't match the author/journal/year the AI gives you, be suspicious — the AI will often attribute a real title to the wrong author or the wrong journal. Third, journal-year consistency: AI sometimes fabricates an article in a year where a journal does not exist (“published in 2019” but the journal was founded in 2021).
A concrete example: a student asked the AI for a citation supporting the pathogenicity of a variant. The imprint looked perfect: a well-known journal, the name of an actual research group, a formally correct DOI. When the DOI was opened, it turned up an article on a completely different topic — the AI had matched a real DOI with a made-up title. Lesson: Just "opening" the DOI is not enough; Also see that the title and author of the page that opens match the citation exactly.
One particularly dangerous situation: the claim that a variant has “previously been reported as pathogenic.” This assertion may translate into a clinical decision (such as PS1/PM5 in ACMG); If the reference on which it is based is spurious, the chain of evidence collapses radically. Every reference of clinical weight must be verified in the strictest manner and in two independent indexes.
5) Citation cross-validation template:
List in order the checks I would make to verify the following attribution:[imprint]. Is the DOI displayed? Does it display the title/author/year of the opened page? Does it appear with a title search in PubMed? Leave a "pass/fail" box at each step. You cannot verify the attribution; give me the checklist.
In summary
- AI is a powerful aid in literature review and scientific writing, but it is the biggest source of fraudulent citations.
- The actual search is done in authoritative directories (PubMed, Scholar); AI's list is not the final word.
- Each citation must be personally verified with DOI/PMID; The relationship-causation distinction should be preserved.
- Use AI to language and structure; The facts, numbers and attribution must come from you.
Application task
Choose a research question. Get search terms from AI with template 1, search it yourself on PubMed, and find 3 real articles. Summarize these to the AI (template 2). Also ask the AI for “3 articles from memory” and check their DOIs — note how many turn out to be real. Write in one sentence the difference in reliability between the two methods.
checklist
- [ ] I extracted the search terms with AI and made the search myself.
- [ ] I found the articles in the authoritative index, I did not trust YZ's list.
- [ ] I verified each citation with DOI/PMID.
- [ ] I confirmed the distinction between relationship and causality in the source.
- [ ] I used AI only for language/structure, not for fact generation.
- [ ] I observed the journal/institution AI usage policy.