Ethics, Biosecurity, Confidentiality and Scientific Integrity — Artificial Intelligence in Biology

Gains:

Ability to protect sensitive human/genetic data in accordance with KVKK, GDPR and ethics committee rules and avoid giving them to the open model
Ability to question the validity of results by assessing risks of biosecurity/dual use and population bias
Understanding that ethical responsibility cannot be delegated to the vehicle and that meaningful human-in-the-loop is necessary

Using artificial intelligence strongly in biology is complemented by using it responsibly. Biology is a sensitive field that touches human health, genetic privacy, the environment and even biosecurity. In this unit, we will discuss the ethical, legal and security responsibilities you will face when using artificial intelligence in biological studies. This unit is a framework built on the principles of verification and honesty in all previous units.

Basic principle: AI is a tool; Ethical responsibility always lies with humans. "The model suggested" is not an excuse.

Four areas of responsibility

1. Data privacy and human data

Genetic, clinical or health data from human participants is extremely sensitive. A person's DNA sequence can reveal their and their relatives' disease risk and identity; Even genomic data thought to be anonymized can be re-identified. Pasting this data into a publicly available AI model may violate data protection laws (KVKK in Türkiye, GDPR in Europe) and ethics board (IRB/ethics committee) approvals.

Do not provide personal/clinical data to the open model.
Avoid use beyond the scope of consent; What did the participant consent to?
Opt for enterprise/local (on-premise) or data processing contract tools.

2. Biosecurity and dual use

Some biological information is "dual use": it can be both useful and harmful; for example, pathogen (disease-causing) sequences, toxin production. Asking AI for information on enhancing dangerous pathogens or designing harmful biological agents is unethical and illegal in most places.

Don't make dangerous dual-use requests.
Follow your institution's biosafety board (IBC) guidelines.

3. Scientific integrity

Everything we've seen in previous units comes together here: no spurious attribution, no data beautification, no p-hacking, no selective reporting. Artificial intelligence can make these easier or harder; The difference is in your honesty.

Report all results (including negative ones).
Declare the use of artificial intelligence.
Support reproducibility by sharing raw data and code (if possible).

4. Bias and fairness

AI models carry biases of the data on which they are trained. Genomic databases predominantly come from populations of European descent; this leads to poorer predictions in other populations. An image model may perform poorly in a condition that is underrepresented in the training data.

Question what population/data the model was trained on.
Evaluate whether the results hold across all groups.

Tip: Ask yourself: "If I give this data to the model, who does it belong to and did that person give permission?" If the answer is vague, don't give it. The breach of confidentiality is irreversible.

Step by step: responsible AI control

Classify the data source: Personal/sensitive or public?
Check consent and permissions: Is this use covered?
Choose the right tool: Secure/native environment for sensitive data.
Consider the risk of dual use.
Question bias: Is the result valid in all groups?
Document and declare use.

Copiable prompt templates

Review my analysis plan below from an ethical perspective: list cautions regarding data confidentiality, participant consent, potential bias, and risk of dual use. Plan: [text] (Note: I am NOT sharing actual patient data, just the plan.)

What privacy and consent questions should I answer before using this genomic dataset? A checklist is prepared in terms of KVKK/GDPR and the ethics committee.

How should I transparently declare the artificial intelligence tool I use in the method section of my article? Write an example declarative sentence; what information (tool, version, scope of use) should it contain?

How do I evaluate whether the results of this model have population bias? List the questions I should ask about the training data and the check steps.

Weak prompt / Strong prompt

Weak: "Analyze the following patient's genetic data: [real data]."

Güçlü: "I am setting up an analysis plan that works with clinical genomic data, but I do not share real patient data. Only from a methodological perspective: what confidentiality measures should I take, in what environment should I process the data, what approval and ethics committee conditions should I meet? You can show the flow with dummy/sample data."

Difference: No actual sensitive data is shared in the powerful prompt; Only the method is asked. Privacy is maintained, the model still helps.

three mini cases

Case 1 — Privacy violation: A researcher pasted what he thought was anonymous patient exome data into an open model to have it analyzed. When the ethics committee learned of this, the study was suspended; There was no data processing agreement. Lesson: sensitive data never enters the open model.

Case 2 — Population bias: A polygenic risk score tool was trained on data of European origin; In another population, risk estimates were significantly inaccurate. When the model suggested questioning the composition of the training data, the team decided not to use the tool in that population. Lesson: bias saves or harms lives.

Case 3 — Disclosure and transparency: A team wrote data cleaning code with AI and stated this clearly in the method section; He also shared the code. Reviewers welcomed this because repeatability was strong. Lesson: transparency breeds trust.

comparison chart

area

safe behavior

risky behavior

patient data

Local/secure environment

Paste to open model

consent

Use within scope

Do not exceed scope

Biosecurity

Avoiding dual use

dangerous demand

integrity

Report all results

selective reporting

bias

Query the population

Apply blindly

transparency

Declare usage

hiding

Common mistakes

Giving sensitive data to the open model: Irreversible privacy violation.
Exceeding the scope of consent: Use not authorized by the participant.
Ignoring bias: Generalizing results to all groups.
Concealing use: Not declaring AI contribution.
"Model suggested" defense: Attributing responsibility to the vehicle.

Caution: In high-consequence biological work (clinical decision, biosafety, human data) AI output is not a substitute for competent expert verification and ethical oversight; it only speeds it up. The ultimate responsibility always lies with the human being, along with the ethical and scientific consequences of the decision.

Environment, ecology and traditional knowledge

Ethical responsibility is not limited to human data. Caution is also needed when using artificial intelligence in ecology and conservation biology. For example, the precise location data (camera trap coordinates) of an endangered species is sensitive due to the risk of poaching; Releasing this data to an open model or public could harm the species. Similarly, ethical and legal (such as the Nagoya Protocol) issues arise when the traditional biological knowledge of local communities (e.g. the use of a plant) is used without permission. Before using the data, "who does this information belong to and who would be harmed by its disclosure?" Always ask the question.

Human oversight and final decision

The essence of this entire module boils down to one principle: meaningful human oversight. AI produces a suggestion, writes a draft, shows a pattern; But a biological, clinical or ethical decision is never based directly on model output. Especially in high-risk areas (diagnosis, treatment, species conservation decision, release), the role of the model is not to make decisions, but to accelerate the expert's decision. The "human-in-the-loop" approach is not a slogan, but a necessity.

For this surveillance to work, the supervisor must be competent. Blind consent (“model said, acceptance”) is not oversight. True oversight requires expertise that can question the output, catch its error, and reject it when necessary. Therefore, artificial intelligence does not replace the expert; It makes the expert even more indispensable. The one who uses the vehicle best is the one who can control it best.

In summary

Responsible use of AI in biology covers four areas: data privacy (protecting sensitive human data), biosecurity (avoiding double use), scientific integrity (honest and complete reporting) and bias awareness (validity of results across all groups). Do not provide sensitive data to the open model, declare usage transparently, question population bias. Ethical responsibility can never be delegated to the agent; It is always in humans.

Application task

Consider a scenario from your own workspace (e.g. using a human dataset or an image model). Have the AI come up with the ethical checklist of this scenario (confidentiality, consent, bias, dual use, transparency) without sharing real data. For each item, mark it as “appropriate/risky/uncertain” in your situation and write an action plan for those that are risky/uncertain. Also have an AI use statement drafted for your article.

checklist

[ ] I did not provide sensitive/personal data to the open model.
[ ] I checked the scope of consent and ethics committee.
[ ] I have assessed the dual use/biosecurity risk.
[ ] I reported all results honestly.
[ ] I questioned population/data bias.
[ ] I have transparently declared the use of artificial intelligence and taken responsibility.

Module Exam

1. A student asked the language model 'how many strings are in this FASTA file?' he asks and the model answers '48'. Which is the most correct approach?

A) Directly using the number 48 given by the model; The model is reliable because it sees the file
B) Ask the number once again and accept it as correct if the two answers are the same
C) Reading the file with Biopython, counting the number of sequences with code and using the output ✔
D) Opening the file manually and counting by eye decision

Explanation: Deterministic tasks such as counting and measuring should be left to the code you run, not to the language model. The model does not 'remember' the number, it produces a probabilistic value and may be mistaken; Counting with a tool like Biopython gives accurate and reproducible results.

2. You want to count cells in a microscope image and measure the area of each. What is the most appropriate approach for this task?

A) Using an expert segmentation tool like Cellpose/StarDist and verifying the result manually ✔
B) Paste the image into the general language model and ask 'how many cells are there'
C) Describe the image to the model and ask for an estimated number
D) Accepting the number of pixels as the number of cells without any calibration

Explanation: Cell segmentation and measurement is not the domain of the general language model, but of specialized image models (Cellpose, StarDist) specially trained for this task. The language model writes the code that calls these tools; The expert model makes the measurement, and the biologist verifies the result.

3. A graduate student adds 8 sources produced by the language model to his thesis introduction. The consultant finds that 3 of these do not exist at all. How could this situation be prevented?

A) By asking the model to produce resources once again
B) By selecting only the citations with DOI (if there is DOI, it is real)
C) Requesting the list by instructing the model to 'fit'
D) Independently verifying each citation on PubMed/doi.org and citing only the articles he has read ✔

Explanation: General language models are not a literature database; Instead of searching for real records, it 'generates' realistic-sounding attributions (fabricated attribution). Each byline and DOI should not be used without independent verification in sources such as PubMed/doi.org and only citing articles that have been actually read.

4. What is the most powerful way to ensure the accuracy of a data cleaning code written by artificial intelligence?

A) If the code works without errors, accept it as correct.
B) Testing the result with a small known sample and verifying the expectation with assert ✔
C) Ask the model 'is the code correct?' asking and trusting the answer 'yes'
D) Run the code several times and get the same output each time

Remark: Just because the code works without errors does not mean it is correct; 'wrong code working without errors' is the most dangerous situation in biology. Testing with a small sample whose result you know in advance and embedding the expectation in the code with assert is the strongest shield against silent wrong results.

5. In an RNA-seq analysis, 20,000 genes are tested and 3,800 genes are found to be 'significant' with p<0.05 without correction. What is the main problem with this conclusion?

A) The number of samples is too high, it should be reduced
B) p threshold is too high, 0.10 should have been used
C) Multiple comparison correction (FDR) was not performed; There are many false positives ✔
D) Samples should have been tested instead of genes

Explanation: When you test thousands of genes simultaneously, many genes appear 'significant' by chance, even if there is no real difference; This is called the multiple comparison problem. The solution is to apply FDR (false discovery rate) correction with a method such as Benjamini-Hochberg; With correction, the list usually reduces to tens to a few hundred genes.

6. The best match in a BLAST result has an E-value of 2.0. What does this mean?

A) The match is most likely random; ✔ not a reliable match
B) The coupling is very strong; The greater the e-value, the better
C) The sequence matched at a rate of 2%
D) There is a 2 base difference between the two sequences

Explanation: The E-value indicates the expectation of the match being random; It's better to be small. An e-value greater than 1 indicates that the match is most likely random. For reliable matches, very small thresholds such as e < 1e-5 are generally sought.

7. In an AlphaFold model, the terminal region of the protein shows a low pLDDT score (<50). How should this region be interpreted?

A) AlphaFold made an error in this region, the region should be removed from the analysis
B) This region is definitely an alpha helix
C) The model is completely unreliable, the structure should not be used
D) The region is likely irregular/elastic; should be interpreted as 'unclear' rather than 'false' ✔

Description: pLDDT is the local confidence score for each amino acid; Values below 50 often reflect truly irregular (flexible, non-fixed) regions. It is wrong to automatically delete these regions as 'wrong guesses'; A low score should be read as 'uncertain/flexible' and its biological significance should be evaluated.

8. A researcher reports cell areas only in pixels and the results are inconsistent with the literature. What's the missing step?

A) More cells should have been counted
B) Scale calibration (pixel-to-micrometer conversion) was not performed, the results were not converted to physical units ✔
C) Segmentation tool was chosen incorrectly
D) Image resolution is too high

Explanation: Scale calibration (how many micrometers per pixel) is essential to extract physical size from the image. If this step is skipped the values remain incomparable, depending on the microscope setting. Areas must be converted to physical units such as square micrometers.

9. A student takes 10 tissue samples from a single mouse and uses 'n=10' in statistics. Why is this incorrect?

A) 10 samples are too few for statistics, at least 30 are needed
B) Tissue samples had to be taken from different organs
C) These 10 examples are technical repetitions; biological n is still 1, this is the so-called repetition ✔
D) Samples are invalid because they were taken on the same day

Explanation: Repeated measurements taken from the same individual are technical repeats; where statistical n is the number of biological units (here mice). Considering technical repetition as biological (pseudoreplication) produces false significance. Proper design means working with multiple individuals.

10. One analysis found p=0.03. What is the correct interpretation of this value?

A) There is a 3% chance of seeing this or more extreme result when in reality there is no difference ✔
B) The probability of the effect being real is 97%
C) The probability of the null hypothesis being true is 3%
D) The result is 97% repeatable

Explanation: p-value is not 'probability that the hypothesis is true' or 'probability that the effect is true'. The p-value is the probability of seeing a difference as or more extreme than that observed when there is actually no difference. Therefore p=0.03 does not mean 'the effect is 97% real'; results should also be evaluated by effect size and confidence interval.

11. In a presentation, a 3% difference between two groups is shown as huge by compressing the y-axis to the 95-100 range. What is this an example of?

A) A good visualization technique; highlights the difference
B) Colorblind friendly palette problem
C) An acceptable style choice
D) A misleading and dishonest chart that exaggerates the difference with the truncated axis ✔

Explanation: Not initializing the axis from zero (truncated axis) misleads the viewer by visually exaggerating a small difference. This is a violation of the principle of honesty; Especially in bar charts, the axis should start from zero and the difference should be shown with its actual size.

12. A researcher pastes what he thinks is anonymous patient exome data into a public language model to have it analyzed. Why is this behavior problematic?

A) There is no problem; data can be shared if anonymous
B) Providing sensitive patient data to an open model is a violation of privacy and ethical/legal (KVKK/GDPR) and cannot be reversed ✔
C) It is problematic only because the analysis will be slow
D) The model is problematic because it cannot understand the data

Description: Patient genetic/clinical data is extremely sensitive; Even genomic data thought to be anonymous can be re-identified. Providing this data to a public model violates KVKK/GDPR and ethics committee (IRB) approvals and is an irreversible breach of privacy. Sensitive data should be processed in local/secure, contracted environments.

13. In one study, the difference they thought was 'patient vs control' was actually due to the samples being processed on two different days. What is this situation called and how is it handled?

A) It is the outlier effect; dots should be deleted
B) It is a lack of normalization; only scale correction is sufficient
C) It is the batch effect; should be balanced in the design and added to the model as a batch variable ✔
D) p-hacking; the test should be changed

Explanation: Technical difference due to sampling batch/processing day is called batch effect and can be confused with biological difference. The correct approach is to balance the batches in the design and add the batch variable to the statistical model (e.g. '~batch + condition'), thus separating the technical signal from the biological signal.

14. You want to calculate the GC ratio of a DNA sequence. Which is the correct and repeatable approach?

A) Print the code that calculates the GC rate with Biopython, run it and use the output ✔
B) Give the model the sequence and ask it to verbally tell the GC rate
C) Confirming the ratio given by the model by another model
D) Looking at the first 100 letters of the series and guessing by eye

Explanation: GC rate is a deterministic calculation; should be based on the code that processes the array, not on a value that the language model 'remembers'. Instead of having the model calculate it, printing the calculating code with a tool like Biopython and running it yourself will provide an accurate output that gives the same result every time you run it.