Literature Review and Scientific Writing — Artificial Intelligence in Biology

Gains:

Ability to use artificial intelligence not in finding sources, but in summarizing the real article you provide, establishing a search strategy and positioning it as a language tool
Understand the mechanism of fabricated citations and eliminate each citation by independently verifying it with PubMed/DOI
Ability to provide scientific claims and numbers and transparently declare the use of artificial intelligence

Much of a biologist's time is spent reading and writing: keeping up with the field, seeing how much of a topic has been studied, turning his findings into a paper, thesis, or grant application. Artificial intelligence provides great acceleration in both of these areas; But this is where it contains the most dangerous trap: hallucinated citations. The model can produce article bylines that don't exist at all but look extremely realistic. In this unit, we will learn to use artificial intelligence safely in literature review and scientific writing.

The basic principle: let AI speed up your typing, not your thinking and proofing.

The right role of artificial intelligence in literature review

Artificial intelligence is not a literature database. A generic LLM does not reliably know which paper exists and cannot generate actual citations. The right tools for the job:

PubMed, Google Scholar, Semantic Scholar: Real article search.
Citation chaining tools (e.g. Connected Papers, Litmaps): Find related articles as a visual map.
Document-based AI: Tools that summarize the actual article (PDF/text) you provide and answer your questions based on that text. Here the model does not make up the source, it processes what you give it.

Safe use of LLM: summarizing real articles you find, comparing them, explaining a topic to you, suggesting search terms.

Tip: Don't tell the AI "give me 10 sources on this topic"; instead “what search terms should I use in PubMed on this topic?” ask. Then you make the real call. The model is reliable not in finding resources but in establishing a search strategy.

The problem of fabricated attribution: why and how to avoid it

LLM produces text; It works with the logic of "next possible word". When asked for a citation, instead of recalling an actual article, it "generates" a realistic-looking byline: plausible author names, realistic journal, appropriate year, even a fake DOI. Most of them never exist. Protection:

Independently verify each citation: search DOI on doi.org, citation on PubMed.
Only cite articles you have read yourself.
Never place the citation produced by the model directly in the bibliography.
Even if you're using a document-based tool, check that the quote is in the actual text.

Artificial intelligence in scientific writing

Here artificial intelligence is really powerful, because it no longer works as a "source of information" but as a "language tool":

Draft generation: The first version of a method paragraph, a summary.
Language correction: Fluency and grammar, especially for non-native English speakers.
Reorganization: Bringing scattered notes into logical flow.
Different audience: Rewriting the same finding for expert, student or public.
Response letter: Draft response to referee comments.

But scientific claims, numbers and references must always come from you and be verified by you.

Step by step: processing an article summary

Find the actual article (PubMed/Scholar).
Give the text to the AI (summary or full text).
Ask for a structured summary: question, method, findings, limitations.
Compare the claims with the text: Has the model made any additions?
Transfer it to your own note, record the source with the real citation.

Copiable prompt templates

Summarize the following article summary under 5 headings: (1) research question, (2) method and sample, (3) main finding (with numbers), (4) limitations, (5) relationship to my study. Do not add any information that is NOT in the text. Text: [abstract]

Suggest effective search strings to search for the following topic in PubMed: key terms, synonyms, MeSH terms, and boolean operators. Subject: [subject].Don't give me a list of sources, just give me a search strategy.

This method streamlines my paragraph and fixes the grammar, but don't change any numeric values, method names, or assertions. Mark the changes. Text: [paragraph]

Rewrite my finding for three different audiences: (1) field expert, (2) undergraduate student, (3) general public. Maintain scientific accuracy, do not exaggerate. Finding: [text]

Weak prompt / Strong prompt

Weak: "Write an introduction about CRISPR and add a source."

Strong: "Below are 4 articles that I have read and verified by myself (citations attached). Based on these articles, draft an introductory paragraph on CRISPR in plant breeding; indicate in parentheses which of these 4 sources supports each claim; do not add any claims that are not in these sources."

Difference: In the powerful prompt, you provide the resources and the model only uses them. The risk of fabricated attributions is eliminated.

three mini cases

Case 1 — Fake DOI: A graduate student included 8 sources provided by artificial intelligence in his thesis introduction. When the consultant checked, it turned out that 3 of them never existed and the authors of 2 of them were wrong. The student had to overhaul the entire bibliography. Lesson: never trust the attribution of the model.

Case 2 — Added finding: A researcher had a paper summarized; the model added a “statistical significance” phrase that was not in the summary. The researcher noticed the difference when he compared it with the text. Lesson: always compare the abstract with the source.

Case 3 — Correct usage: A researcher, a non-native English speaker, had the AI streamline the method section he wrote; He stipulated that he should not change the numbers and claims. The result is a much more readable text without distorting the meaning. Lesson: as a language tool, AI is excellent.

comparison chart

Quest

Is artificial intelligence reliable?

Correct tool/method

sourcing

PubMed/Scholar

Generating a citation

No (makes up)

Verify manually

Summarizing the given article

Yes (confirm by text)

LLM + comparison

Language correction

Yes

LLM

Create a draft

Yes (claim from you)

LLM

Search strategy

Yes

LLM

Common mistakes

Using the attribution of the model without verifying it: The most common and most harmful mistake.
Mistaking LLM for a search engine: It doesn't find real articles.
Relying on the summary and not reading the source: May cause model addition/distortion.
Making artificial intelligence produce claims: The scientific claim must come from you.
Plagiarism and originality: Presenting the text produced by the model as if it were my own sentence; may violate journal policies.

Caution: Many journals and institutions require disclosure of AI use and do not acknowledge AI as an author. Transparently indicate the tool and scope you use in your AI-supported writing; You are responsible for the accuracy of the final text.

Systematic screening and the frontier of artificial intelligence

Artificial intelligence seems attractive in systematic reviews that want to comprehensively and unbiasedly scan an entire topic, but it should be used with caution. The essence of a systematic search is a reproducible and transparent search strategy: documenting which databases were searched, with what terms, with what inclusion/exclusion criteria. AI helps you set up this strategy, expand the terms, and write the screening protocol. But it is the actual database search that actually finds the articles; Just because the model says "here are relevant studies" is not a screening substitute and may miss important studies.

Another powerful and safe use is to help screen large numbers of articles: screening hundreds of abstracts to see if they meet your inclusion criteria takes time. The model can pre-classify each summary as “possibly relevant/irrelevant” based on your criteria; but the final decision to include and uncertain situations remains up to the human. Here again, the model is not a decision maker, but a pre-screening aid.

Evaluate the following article abstract against the following inclusion criteria:[criteria]. Classify as "relevant/irrelevant/uncertain" and give justification in one sentence. The final decision is mine; You just pre-qualify.Summary: [text]

Hidden traces of artificial intelligence in my writing

AI-generated texts are sometimes overly general, repetitive, or “cool but empty.” Concreteness is essential in scientific writing: every sentence should convey some information. Instead of leaving the model outline as is, start each paragraph with the question "does this sentence really say anything, or is it filler?" Eliminate. Additionally, the text produced by the model sometimes contains subtle mis-terms or common but erroneous expressions; Check each technical term with the eyes of a domain expert. AI is a good first draft generator; He is not a good proofreader. The final reading is always you.

Why attributions are made up: understanding the mechanism.

Seeing the model's attribution fabrication as a natural consequence of the way it works, rather than a "mistake", makes it easier to protect. During training, the model learns the pattern of millions of real bylines: what author names look like, the format of journal names, the structure of DOIs. When a citation is requested, it does not "look up" an actual record; It "generates" a statistically plausible tag that fits the pattern it has learned. So the result looks very realistic but may not be real; he may even incorporate the name of a real author into an article he never wrote. Once you understand this mechanism, the rule becomes clear: any identifier generated by the model is not considered to exist until you see it in the database. Document-dependent tools (where you provide the actual text) reduce this risk because the model relies on the text you provide rather than fitting; but still check that the quote is actually in that text.

In summary

AI doesn't find (make up) sources in a literature search, but it summarizes the actual articles you provide, builds a search strategy, and streamlines your writing. The biggest danger is made-up attributions; independently verify each byline and only use sources you have read. Scientific claims, numbers and citations come from you; AI only strengthens language and order. Declare usage transparently.

Application task

Give an article you actually read (abstract or full text) to artificial intelligence and create a structured summary (question, method, finding, limitation). Compare the summary with the source text to see if the model has added or distorted anything. Separately, ask the model for a list of resources on the same topic; Search every byline he gives on PubMed and count how many are real.

checklist

[ ] I found the sources myself in PubMed/Scholar, I did not make them fit the model.
[ ] I have independently verified each citation.
[ ] I compared the abstract with the source text.
[ ] I provided the scientific claims and numbers myself.
[ ] I only cited sources I have read.
[ ] I have transparently declared the use of artificial intelligence.