Human in Experimentation: Ethics, Privacy, Responsibility and End-to-End Workflow — Artificial Intelligence in Chemistry (Basic Science)

Gains:

Integrating the human principle into the end-to-end workflow in the experiment and being able to separate artificial intelligence/human roles at each stage
Ability to work with generic proxy examples without giving confidential structure and data to general artificial intelligence
Ability to place human approval gates at critical points and assume ultimate responsibility with transparency and traceability

In this final unit we combine the entire module into a single framework: human-in-the-loop. This principle says that no matter how capable the AI is, a human must understand, control, and approve at every critical point of a chemical decision. In this unit, we will first clarify the ethical and confidentiality aspects, then establish an end-to-end workflow that takes a molecule from idea to report, and see how AI and humans divide labor at every stage. The goal is to turn everything you've learned into a system that fits into daily work.

Privacy: what not to give to AI?

Chemistry often works with confidential information: unpatented molecules, company secret synthesis routes, unpublished data, patient samples. Text you enter into a public AI service may be stored or processed according to the service's policy. Therefore:

Check your institutional policy before releasing unpublished original structures (pre-patent molecules) to public AI.
Never enter personal/patient data (clinical samples) in an identifiable form.
Review confidentiality agreements before sharing trade secret synthesis pathways.
Opt for in-house/custom deployment AI tools when necessary.

Caution: Asking a general AI about a molecule may mean revealing the structure of that molecule. For a unique structure that has not been patented, this may jeopardize your priority (patent right). Ask for hidden structures with generic/surrogate examples.

Ethics: four principles

Transparency: Don't hide where you use AI; Comply with publication and institutional policies.
Responsibility: The human is responsible for the final decision; “AI suggested” does not excuse a security or integrity breach.
Non-maleficence: Do not use AI for dangerous/illegal synthesis, dual-use (weapons potential) applications.
Fairness and honesty: Don't distort data, don't embellish results, don't make up attributions.

End-to-end workflow: molecule → report

Now let's combine the entire module into a single flow. At each stage, pay attention to the distinction between what the AI does and what the human does.

Stage

Role of AI

The (critical) role of humans

1. Idea/literature

Concept summary, search term

Finding the real source, attribution confirmation

2. Structure description

generating SMILES

Authentication with RDKit, InChIKey

3. Synthesis plan

Retrosynthesis, alternatives

Literature confirmation, path selection

4. Security

SDS summary, preliminary mismatch

SDS reading, approval, PPE decision

5. Application

(None)

All laboratory work

6. Analysis

Spectrum comment support

Final structure assignment

7. Data processing

writing code

Run the code, verify the result

8. Documentation

Formatting, control

Observation, deviation, signature

9. Reporting

draft text

Number/citation verification, accountability

This table shows how the module's 11 units are arranged into a single job. AI accelerates; people verify and own.

Step by step: setting up the workflow

Clarify roles: For each task, “Does the AI or the human decide?” determine from the beginning.
Set up a privacy filter: Without giving something to the AI you can ask "is this private/private?" ask.
Put verification gates: A verification step at the end of each stage (table in unit 10).
Keep traceability: What output came from the AI, how was it verified, who approved it.
Set approval points: Human approval is mandatory at critical points such as security and final result.
Feedback: Record every error caught and improve workflow.

Four copyable templates

1) Privacy pre-filter:

Before asking a question, consider this:Content: [WHAT I AM ASKING]Task: Does this content contain unpublished original artifacts, personal data, or trade secrets? If it does, suggest how I can rewrite the question with a generic/surrogate example that will hide the structure.

2) End of phase verification gate:

I have completed the following phase: [PHASE, e.g. synthesis plan].Task: List all the elements that need to be verified coming out of this stage (number, citation, structure, security claim). For each, write the verification method and who should confirm it. Mark any open items that need to be closed before moving on to the next step.

3) Human-AI role distribution plan:

My project: [SHORT DESCRIPTION].Task: Divide this project into phases. For each stage, distinguish between: - What the AI can do (draft, code, summary) - What the human must decide (security, consequence, approval). Mark the critical points where the ultimate responsibility lies with the human.

4) Report integrity check:

Below is my report draft:[DRAFT]Task: Check and mark the following:1) Are there any unverified numbers/citations?2) Is the use of AI transparently stated?3) Does the data appear "beautified" (outlier hiding)?4) Are there any security/ethical flaws? [CHECK] where you are unsure.

Weak prompt / Strong prompt

Weak:

What is the best way to synthesize our company's new secret molecule?[actual original structure]

This exposes the original pre-patent structure to a general service; It puts priority and confidentiality at risk.

Strong:

I would like to learn the general retrosynthesis strategy for a similar structure of a class of aromatic ketones. Explain through a generic example (acetophenone derivative) without giving the concrete, original structure. I will apply the strategy to my own molecule in-house.

Difference: without revealing the latent structure, the same learning was achieved as with the generic surrogate example.

mini cases

Case 1 — Original structure disclosed. A team asked a general AI about a unique molecule before applying for a patent. The legal unit stated that this could bring priority into question; The process was delayed. Lesson: ask for unique structures with generic proxy, respect privacy policy.

Case 2 — Human consent prevented the accident. In an automated workflow, the AI produced a draft procedure and was to be implemented directly; But an experienced chemist caught an incompatible pair of reagents at the flow-through "safety clearance" gate. Lesson: at critical points, human approval gates save lives.

Case 3 — End-to-end discipline. A master's student implemented the entire module in one project: literature summary + actual citation verification with AI, structure verification with RDKit, data analysis with code, security with SDS, honest documentation, transparent reporting. The result was both faster and more robust; in the thesis defense "How did you verify AI?" He answered the question clearly. Lesson: AI + verification discipline work best together.

Common mistakes

Giving the latent structure to the general AI. It compromises patent priority and trade secret.
Entering identifiable personal/patient data. Privacy breach and legal risk.
Bypassing approval gates. Security and ultimately moving forward without human approval.
Transferring responsibility to AI. “AI suggested” does not excuse any mistakes.
Avoiding transparency. Concealing the use of AI undermines integrity.
Ignoring the risk of dual use. It is unethical to use AI for harmful applications.

Tip: Design your workflow with “every critical decision a human stops and approves.” Make AI a link that accelerates the chain, and the node holding the chain is always human.

In summary

Human principle in experimentation: in every critical chemical decision, a human understands, supervises and approves.
Privacy: do not provide unique artifacts, personal data and trade secrets to public AI; Use generic proxy.
Ethics: transparency, responsibility, non-maleficence, honesty are the four basic principles.
In an end-to-end workflow, AI accelerates at every stage; people verify and own.
Put human approval gates at critical points (security, end result) and maintain traceability.

Application task

Choose a project of your own (real or hypothetical) and design an AI-human workflow from start to finish. Adapt the 9-stage table above to your own project: what will the AI do at each stage, what will the human approve? Apply privacy pre-filter: what information do you not give to the AI, how do you genericize it? Define at least two human-approval gates and one traceability method. Write a one-page “workflow and verification plan.”

checklist

[ ] In the experiment, I grasped the human principle and why it was necessary.
[ ] I genericize the hidden structure/data without giving it to the general AI.
[ ] I comply with the principles of transparency, responsibility, non-maleficence and honesty.
[ ] I separated the AI/human roles in the end-to-end workflow.
[ ] I put human approval gates at critical points.
[ ] I track what output comes from the AI and how it is verified.

Module Exam

1. Which of the following is the most accurate positioning for artificial intelligence in chemistry?

A) Artificial intelligence is a code, draft and editing assistant; Responsibility for structure, number, resources, security and ethical decisions rests with humans ✔
B) Artificial intelligence is a chemistry engine and the molecular weights it gives are always exact.
C) Artificial intelligence only works in literature review, it has nothing to do with analysis and security
D) Since artificial intelligence is more impartial than humans, security decisions should be left to it

Description: Artificial intelligence; It is an assistant who writes code, produces drafts, builds a retrosynthesis skeleton, helps and organizes spectrum interpretation. However, the deterministic calculation of molecular weight and stoichiometry, verification of the structure with RDKit, confirmation of citations in the database, decision of safety with SDS/GHS and the final responsibility belong to the competent chemist. The big language model is not a calculator or a chemistry engine; It is a text generator that produces the 'next most likely word' and its unverified output may lead to an incorrect structure, number or hazard.

2. What is the most reliable way to describe a molecule to artificial intelligence?

A) Writing only the common Turkish name
B) Giving a structure representation like SMILES and confirming the identity with InChIKey ✔
C) Just telling the molecular weight
D) It is sufficient to write the trademark

Explanation: Names (especially trade and synonym names) are ambiguous and can lead to the wrong molecule. Giving the molecule a structure notation such as SMILES provides precision; His identity is confirmed in the database with InChIKey. In addition, the SMILES given by artificial intelligence must be parsed with RDKit and verified by canonicalization.

3. How should you use a molecular weight given by artificial intelligence?

A) Using it directly, because AI is reliable in numbers
B) Calculate and verify independently from the formula manually or with RDKit ✔
C) It is enough just to ask another artificial intelligence and compare
D) Molecular weight is unimportant, no need for verification

Explanation: Molecular weight is a quantity calculated deterministically from the molecular formula. So asking the language model is the weakest way; It should be independently verified with a tool like RDKit or by calculating it manually from a formula. The language model can fit such numbers with a 'probable looking' value.

4. What should be done before using a literature citation (article/DOI) returned by artificial intelligence?

A) Nothing; If artificial intelligence gives a reference, it is real
B) Just looking at the title to make it look convincing is enough.
C) Confirming each citation and DOI by decoding them in a database (doi.org, Crossref) ✔
D) The long DOI number proves that it is real.

Description: The language model can generate articles, authors, and DOIs that look realistic but do not exist (fake citation). Just because it gives a DOI does not mean it is valid. Each citation must be verified by resolution in a database such as doi.org, Crossref, or the publisher's site; The safest flow is for a human to find the real article and have the text summarized by artificial intelligence.

5. In which area is artificial intelligence weakest and requires confirmation in reaction planning?

A) In recognizing which general class the reaction belongs to
B) Exact condition numbers, yield values and reagent/catalyst names ✔
C) Explaining a mechanism in plain language
D) Brainstorming for alternative routes

Explanation: AI generally predicts the type of reaction and the main product of well-documented reactions well. However, exact condition numbers (temperature, time, equivalent), yield values, rare reactions, and reagent/catalyst names are the weakest and most prone to fabrication; these should be confirmed with literature and catalogues.

6. What is the most reliable way to use artificial intelligence in spectrum interpretation?

A) Just give a few peaks and ask 'what spectrum is this?' to ask
B) Giving the expected structure and full raw data and having a consistency check (hypothesis testing) ✔
C) Writing only the shift values without performing integration and division
D) Having definitive structure assignment done with a single technique (IR only)

Description: Ask AI 'which compound is this spectrum?' Asking from scratch increases the error rate. The most reliable way is to give the expected structure (SMILES) and the raw data (integration, splitting, solver, peak list) together and ask 'is this data consistent with this structure?' is to ask; that is, using artificial intelligence as a hypothesis tester. The final assignment is for the chemist examining the experimental spectrum.

7. What is the most reliable way to obtain the result of a chemical calculation (e.g. percent yield)?

A) Requesting the result directly from artificial intelligence verbally
B) Print and run the account's Python code and test it with a known example ✔
C) Ask the same question several times and get the number that comes up most frequently
D) Estimating quickly without specifying the unit

Explanation: The language model does arithmetic by 'predicting the likely outcome' and can be wrong even with very simple multiplications. The most reliable way is to have the AI write its CODE (e.g. Python), not the account itself, and run that code; the code is deterministic and auditable. The code should also be tested with a small sample whose answer is known.

8. What does the concept of 'domain of applicability' mean in molecular property prediction?

A) Processing power of the computer on which the model runs
B) Molecular region similar to the training data, where the model gives reliable predictions ✔
C) Assurance area where the number given by the model is always exact
D) Temperature range in which the molecule can be synthesized in the laboratory

Explanation: A QSAR/prediction model works well on molecules that are similar to the molecules it was trained on. Given a molecule that is very different from the training data, the model still produces a number, but this is unreliable; this is called 'being outside the scope of applicability'. The model always gives an answer; It is up to the human to evaluate whether the answer is reliable or not.

9. What is the correct approach to dealing with observations in experiment documentation?

A) Writing observations to artificial intelligence to save time
B) The person doing the experiment writes down the observations; AI only formats and checks ✔
C) It is sufficient not to record the observations, just write the final yield
D) Hiding deviations from the plan to make the report look nice

Explanation: Artificial intelligence does not know what it is; Making observations (color change, gas release, etc.) fit into it produces records that seem plausible but did not happen, and this is scientific fraud. AI formats the recording, recalls missing fields, and checks for repeatability; but the human experimenter must provide the observations and facts. Deviations from the plan must also be recorded.

10. What is the role of artificial intelligence in laboratory safety and the source of truth?

A) If artificial intelligence says 'safe', there is no need for another source
B) The source of truth is SDS and GHS; AI is helpful but can't have the final say ✔
C) Security decisions can be completely delegated to artificial intelligence
D) SDS is unnecessary; Internet forums provide sufficient information.

Description: Safety is the highest priority in chemistry, and artificial intelligence can never have the last word here. The source of actual safety information is the SDS (Safety Data Sheet) and GHS classification provided by the manufacturer. AI can summarize SDS text, generate form drafts and perform preliminary non-compliance screening; However, all claims must be verified with SDS and the procedure must be approved by an experienced person/security officer.

11. Which of the following combinations of reagents is a known hazardous (incompatible) match?

A) Mixing water and table salt
B) Mixing an acidic solution with a solution containing hypochlorite (risk of chlorine gas) ✔
C) Mixing ethanol and water
D) Dissolving sugar and water

Explanation: While some chemicals are relatively safe alone, they are dangerous together. Mixing acid and hypochlorite releases chlorine gas; Oxidizer + organic poses risk of fire/explosion, water + reactive metal heat/hydrogen. Although AI knows most of these combinations, it does not guarantee all of them and can sometimes present a dangerous suggestion in innocent language; each procedure should be screened for non-compliance with SDS.

12. What is the most accurate expression for 'hallucination' (artificial intelligence producing fabricated information)?

A) It is a rare software error and is completely eliminated with an update.
B) It is a behavior inherent in the language model; The solution is to independently verify each output ✔
C) It only appears in incorrectly written prompts, never in correct prompts.
D) If artificial intelligence speaks in a safe language, it is not a hallucination

Explanation: The hallucination is not a malfunction but the result of the nature of the language model that produces the 'next most likely word'; The model aims at the possible, not the true. The most insidious part is that it presents false information in the same confident language as the truth. The solution is not to try to fix the model, but to build a type-appropriate validation shell around each output; fluency is never proof of accuracy.

13. What does 'triangulation' mean to verify a critical value (e.g. molecular weight)?

A) Asking the same question to the same artificial intelligence three times
B) Comparing and reaching the same conclusion in more than one independent way ✔
C) Calculating the weight of three different molecules at the same time
D) Rounding the result to three decimal places

Explanation: Triangulation is reaching the same conclusion through multiple independent paths; For example, comparing the molecular weight according to artificial intelligence, RDKit calculation and manual calculation. If two independent paths overlap, trust is high; If someone deviates, it is stopped and investigated. Instead of relying on a single source, at least two independent methods should be sought.

14. What is the correct way to ask a general AI service about a unique molecule that has not been patented?

A) Giving the original structure directly, because speed is important
B) Learning a strategy through a generic/surrogate example without revealing the original structure ✔
C) Giving structure is not a problem, AI never hides information
D) Not giving the name of the molecule but only sharing its entire SMILES is sufficient protection.

Description: Content entered into a general artificial intelligence service may be stored or processed according to the service policy; Giving an original structure risks confidentiality and patent priority. The correct approach is to learn the strategy from a generic/surrogate example (a similar class of structures) and apply it to its own molecule in-house, without revealing the original structure. Personal/patient data should also not be entered in an identifiable form.