Gains:
- Being able to distinguish where artificial intelligence saves time in the biology workflow (question, design, laboratory, analysis, reporting) and where the sequence, number, source and ethical decisions are left to the human, depending on the risk level.
- Ability to distinguish between a general language model and specialized biology models (AlphaFold, Cellpose) trained for a specific problem and use each of them in the appropriate task.
- Ability to apply a verification discipline that independently checks each output with the four steps of Array/Data–Number–Source–Ethics
Biology; It is a huge data science that extends from the cell to the ecosystem, from DNA sequence to protein folding, from a single experiment to hundreds of thousands of gene measurements. In recent years, the volume of this data has grown so much that a single RNA-sequencing (RNA-seq: method that measures which gene in the cell is read and how much) study can produce enough data for a laboratory for years. This is where artificial intelligence comes into play: as an assistant that writes code, organizes data, produces drafts, summarizes the literature, and builds an analysis framework. Our goal throughout this module is to teach you to use AI not as a “magic box” but as a tool that speeds up the biologist's daily work, but whose every output must be verified by the biologist.
In this first unit, we will discuss where artificial intelligence saves time in the biology workflow, where the decision is left to the human, and which mistakes in this profession should be taken into consideration from the very beginning. There is a saying we will repeat throughout the module: AI accelerates, the biologist decides and bears the responsibility.
What exactly does artificial intelligence do in biology?
A large language model (LLM) is not a microscope, it is not a DNA sequencer, it is not a statistical engine. He is a text producer who knows linguistic and coding patterns very well. Internalizing this distinction from the beginning is the basis of all future verification discipline.
Biology tasks where AI is really strong include:
- Coding: filtering a pandas table (data frame: tabular data structure in row-column format), drawing a graph, reading a FASTA file (standard text format storing DNA/protein sequences) with Python.
- Explaining and teaching: "What is Hardy-Weinberg equilibrium?" To explain a concept like this by simplifying it.
- Producing an outline: The first draft of a methods section, the framework of an email, a presentation plan.
- Summarizing and editing: Reducing a long article to articles, gathering scattered notes.
- Conversion: Building code to map a list of genes to a different form of identification (ID).
Tasks where AI is unreliable are: producing a precise numerical value (“remembering” the expression value of a gene), citing an actual article, reporting the true biological function of a sequence with precision, producing clinical decisions with a patient's data.
Tip: Adopt a simple rule. Let the AI “think and organize,” but let the “counting, measuring, and verifying” either be done by code you run or you verify manually.
Two different "artificial intelligences": language model and specific biology models
When we say "artificial intelligence" in biology, we are actually talking about two very different things, and confusing them can lead to serious errors. The first is the general language model that you will use most in this module: writes code, generates text, explains. The second are expert models trained specifically for a specific biological problem: AlphaFold that predicts the shape of a protein, Cellpose that counts cells in a microscope image, classifiers that recognize species sounds. Expert models are much more reliable than language models in their narrow domain because they have been trained on real biological data and tested in that domain. The language model, on the other hand, knows "a little bit about everything" but does not guarantee measurement accuracy in any of them. The robust workflow combines the two: the language model sets up the code and flow, the expert performs model measurement, the biologist validates the result.
Segregation of duties according to risk level
Not every task carries the same risk. How much you should question the AI output depends on the harm that will occur if that output is wrong. The table below embodies this distinction.
Quest
Risk level
man's role
Asking for clarification to learn a concept
low
Confirmation with source, logic check
Print Python analysis code
medium
Running the code and testing it with a known situation
Mapping gene names/IDs
Medium-High
Confirmation with official database (NCBI, Ensembl)
Interpreting the function of an array
high
Experimental evidence and database are mandatory
Clinical/diagnostic decision
very high
Artificial intelligence should never be used alone
three mini cases
Case 1 — Time savings: A PhD student manually cleaned the RNA-seq count table of 24 samples each time; It took about 40 minutes. He wrote the pandas code to the artificial intelligence and ran it himself; The job went down to 3 minutes. Critical point: tested the code once with a small sample and confirmed that the total number of lines (18,542 genes) was preserved.
Case 2 — Fake citation trap: A researcher asked AI for “5 sources on the use of CRISPR in plant breeding” for the introduction. The model produced 5 realistic looking tags; but the DOI (digital object identifier: permanent address of the article) of 3 of them was not available in any publication. If the researcher had used it without confirming it, his article would have been rejected by the referee.
Case 3 — Numerical hallucination: A teacher asks, “How many genes are in the human genome?” he asked and wrote the value given by the model "about 46,000" on the clipboard. The current estimate is approximately 19,000-20,000 protein-coding genes. The model produced the wrong number in a confident tone. Lesson: always confirm the number with an up-to-date source (Ensembl, GENCODE).
Step by step: a secure AI workflow
- Define the task: Write what you want in one sentence (“Code to filter low-expression genes from a 24-sample count table”).
- Give context: Specify data format, column names, number of samples.
- Ask for output format: "Give working Python code, comment out line by line."
- Run it yourself: Try the code in your own environment; Report back if there is an error.
- Test with known situation: Verify with a small example whose result you know.
- Independent fact-checking: Provide critical numbers and claims via official database or a secondary method.
Copiable prompt templates
Role: You are an experienced bioinformatician. Task: Write Python code for [data/analysis]. Context: Data [format], columns [name], number of samples [N]. Constraint: Give only working code, add short comments to each line, specify libraries.
Simplify this concept as if explaining it to an undergraduate student: [concept]. Define it in 3 sentences, give 1 daily life analogy, correct 1 common misconception.
Examine my analysis plan below and tell me its weak points. Specifically: control group, number of replicates, choice of statistical test. Plan: [text]
Reduce this article summary to 5 items: (1) question, (2) method, (3) main finding and issue, (4) limitation, (5) inference that works for me. Text: [abstract]
Weak prompt / Strong prompt
Weak: "Give me a gene expression analysis."
Güçlü: "I have a table of RNA-seq raw counts from 12 control, 12 patient samples (CSV, rows gene, columns sample). List the steps of the differential expression analysis; explain which Python/R tool I used at each step and why; justify the normalization method. Give the plan first, not the code."
Difference: In the powerful prompt, the data structure, number of samples, output type and justification request are clear. In a weak prompt, the model is forced to guess and often proceeds with incorrect assumptions.
Common mistakes
- Making the model count/measure: Ask the LLM "how many rows are in this table?" to ask. Have the code do it.
- Using attributions without verification: The model can create realistic attributions. Check each DOI.
- Relying on confident tone: AI speaks confidently even when wrong; Tone is not evidence of accuracy.
- Not giving context: Requesting code without telling the data format produces code that doesn't work.
- Pasting confidential/patient data: Exporting personal health data to a public model is an ethical and legal violation (Unit 11).
- Mixing the two types of models: Letting the language model do the work that requires measurement (structure, cell counting) and bypassing the expert model.
Caution: In high-consequence biological work (clinical, biosafety, analysis leading to publication), AI output is not a substitute for competent expert verification; it only speeds it up. The ultimate responsibility always lies with the human being.
The knowledge frontier of artificial intelligence: education segment and currentness
Everything a language model "knows" comes from the texts up to the date it was trained (training segment: the last time the model saw data). Biology, on the other hand, changes rapidly: new gene annotations, updated genome versions (e.g. the transition of the human reference genome from GRCh37 to GRCh38), new classifications are constantly published. The model cannot know a finding after the cut-off date or an updated gene name; It may even present old, obsolete information as if it were current. Therefore, species names, gene IDs, database versions and current statistics should always be confirmed from the official source (NCBI, Ensembl, UniProt).
A second limit is ambiguity blindness: whether the model knows a topic well or not at all, it responds in the same confident tone. People hesitate when they say "I'm not sure"; The model does not hesitate. That's why using tone as a measure of trust is dangerous. In a critical response, the model was asked “how sure are you and how do you know this?” Asking sometimes reveals inconsistency; But the final confirmation is again an independent source.
Beware of the context window and big data
The model can only process a certain length of text at a time (context window: the amount of text the model can process at once). Pasting a genome file or a table of tens of thousands of rows directly into the model would both exceed the limit and pose a privacy risk. The correct approach is to give the data to the code, not the model: the model writes the code, the code processes the data. When working with big data, this distinction is fundamental for both security and accuracy.
Roadmap of this module
In the following units, we will put these principles into concrete workflows: Python fundamentals (Ü2), sequence analysis (Ü3), omics and high-dimensional data (Ü4), protein structure and AlphaFold (Ü5), image and species/cell detection (Ü6), experimental design (Ü7), statistical analysis (Ü8), visualization (Ü9), literature and writing (Ü10), ethics and biosafety (Ü11). The same discipline is repeated in each unit: artificial intelligence produces, biologist verifies.
In summary
Artificial intelligence is a powerful assistant in biology that codes, explains, generates outlines and summarizes; but it is not a calculator, database or decision maker. Distinguish between general language model and specific expert models. Separate tasks by risk level: move quickly on low-risk disclosures, be sure to perform independent verification on high-risk number, attribution and function claims. The biologist who makes the verification discipline an integral part of the workflow gets the most value from AI.
Application task
Choose 3 typical tasks from your field of study (e.g. writing some code, learning a concept, a literature summary). Classify each as low/medium/high risk according to the table above. For the high risk one, write in 2 sentences how you would independently verify the output. Then, use the “Strong prompt” template to assign a task to the AI and test the output with a known situation.
checklist
- [ ] I have classified my task by risk level.
- [ ] I added data format and context to the prompt.
- [ ] I left the counting/measurement to the code or myself, not to the model.
- [ ] I have verified each numerical claim and attribution with independent sources.
- [ ] I delegated the measurement to the appropriate expert model and the flow to the language model.
- [ ] I did not provide confidential/personal data to the open model.
- [ ] I accepted that the final decision and responsibility lies with me.