Gains:
- Being able to distinguish where artificial intelligence saves time in the chemistry workflow (idea, design, analysis, reporting) and where structure, number, resource and safety decisions are left to humans, depending on the level of risk.
- Ability to apply a discipline that independently checks each output with the four steps of Structure-Number-Resource-Security
- Understand why made-up compounds and numbers, false references, and security misconceptions are risks to this profession that must be considered from the very beginning.
Chemistry is an experimental science that studies the structure, transformation and properties of matter. A chemist weighs, dissolves, heats and measures in the laboratory; At his desk, he draws molecules, plans reactions, interprets spectra, scans the literature and writes reports. Artificial intelligence (AI for short, here specifically software that generates text by predicting the “next most likely word,” trained on a large language model, i.e., massive text data) can dramatically speed up many of these desk jobs. But it doesn't build the lab itself, it doesn't read the scales, and most importantly: it doesn't verify any numbers, formulas, or citations it produces on your behalf. This module teaches how to use AI in chemistry end-to-end but responsibly.
In this first unit, I want to plant one idea in your mind: AI is an assistant in chemistry, not a chemist. Writes code, produces outlines, builds retrosynthesis framework, helps interpret spectrum peaks, summarizes literature. However, it is the competent chemist who decides whether the molecule can actually be synthesized, whether a reaction is safe or not, whether a reference is genuine or not, and the ultimate responsibility.
Where does AI fit into the chemistry workflow?
Let's roughly divide a chemistry project into five phases: (1) idea and literature, (2) molecule and reaction design, (3) laboratory application, (4) analysis and characterization, (5) reporting. AI carries very different values at these stages.
- Ideas and literature: AI quickly summarizes a topic, explains key concepts, generates search terms. But he can make up attribution; We will see this in a moment.
- Design: Generates SMILES (a form of writing the molecule as a text string; more on that later), suggests retrosynthesis steps, lists reactive alternatives. Suggestion is good, not a guarantee.
- Laboratory: AI does nothing here. Weighing, heating and security measures are carried out by humans. AI only helps write the procedure; Before performing that procedure a human has to supervise it for safety.
- Analysis: In interpreting NMR, IR, mass spectrum (these are measurement techniques that show the structure of the molecule), YZ helps to group the peaks and list possible structures. The final assignment is yours.
- Reporting: Produces experimental notebook, article draft, safety form draft. It is very powerful; but the numbers should match yours.
Tip: The safest place to use AI is in "if something goes wrong it'll be noticed and fixed immediately" areas: draft text, code skeleton, term description. The riskiest places are the "if false, it spreads silently" areas: constants, references, security assertions, numerical results.
Why do we have to be careful? Three structural risks
The big language model is not a calculator or a chemistry engine. Produces text that statistically "looks likely". This leads to three concrete risks in chemistry:
- Made up (hallucination): Can confidently write down a compound that is not a model, a reaction that does not exist, or an incorrect molecular weight. For example, it says "The molecular weight of compound X is 154.2 g/mol", whereas the correct value is 168.2.
- Fake citation: May give a source that looks realistic but does not exist at all, such as "Smith et al., 2019, Journal of Organic Chemistry." He can even make up a DOI number.
- Safety fallacy: Can present a dangerous combination of reagents (e.g. an oxidizer and an organic solvent in an inappropriate ratio) as "no problem".
These three risks will appear again and again in each unit of this module. The solution is not to escape from AI, but to establish the discipline of verifying each output with an independent source.
mini cases
Case 1 — Fitted molecular weight. A graduate student asked AI the molecular weight for para-nitrophenol. “139.11 g/mol,” the AI said. The student was confident and planned to weigh 6.96 g for 0.05 mol. However, the molecular weight of para-nitrophenol is 139.11 g/mol, and this time it was correct — but the student had never checked. On the next compound (para-aminophenol, actual value 109.13 g/mol) YZ said “123.15”; This time the student caught it by calculating it by hand from the formula (C6H7NO2): 6×12.011 + 7×1.008 + 14.007 + 2×15.999 = 109.13. Lesson: calculate each molecular weight without formula.
Case 2 — Fake DOI. A researcher asked AI for 5 sources in his literature review. The DOI of two of the five returned "not found" when clicked; The headlines were factual, but the articles were not. Waste of time 40 minutes. Lesson: not every citation should be used without verification in the database (DOI, PubMed, Reaxys).
Case 3 — Hazard that seems safe. A user asked the AI about a cleaning procedure; AI indirectly produced a suggestion of mixing a hypochlorite-containing solution with an acidic solution—a combination that could release chlorine gas. Luckily the user looked at the safety data sheet (SDS) and stopped. Lesson: every procedure is human-inspected for SDS and hazard before being implemented.
Weak prompt / Strong prompt
Weak prompt:
Synthesize this molecule.
This claim is vague: what molecule, what starting materials, what scale, what constraints? AI makes up the gaps.
Powerful prompt:
Role: You are an experienced assistant in organic synthesis. Target molecule: 4-methoxyacetophenone (SMILES: COc1ccc(cc1)C(C)=O). We have anisole as the starting material. Task: Give a 2-step retrosynthesis proposal. For each step: reagents, conditions (temperature, solvent), and POSSIBLE side reactions. Constraint: Mark "VERIFY" where you are not sure, fitting.Output: numbered steps + list of points to be verified.
Difference: role, exact target (with SMILES), context, number of steps, output format and "fitting" constraint. These are the five components of good prompting in chemistry: role, precise structure representation, context, task, verification constraint.
Types of AI tools: language model or chemistry tool?
In chemistry you encounter two different "AI" and it is critical not to confuse them:
Genre
What does
reliability
Example usage
General language model
Generates text/code, explains, writes drafts
Low on numbers, high on language
Procedure outline, terminology explanation
Chemistry library (RDKit etc.)
Processes the molecule deterministically, calculates the molecular weight
High (rule-based)
SMILES verification, MA account
Custom prediction model (retrosynthesis, feature)
Gives data-trained prediction
Medium, depending on area
Retrosynthesis, solubility estimation
Literature/search tool
Real database queries
High depending on source
Attribution confirmation, reaction search
The most robust workflow combines these: the language model generates the idea, the chemistry library calculates the number deterministically, the literature tool verifies the attribution, the human validates. We will establish this chain in the next units.
Verification discipline: FOUR steps
Take each AI output through these four steps before entering it into the lab or report:
- Structure: Is the molecule/reaction notation (SMILES/InChI) valid? Can it be parsed with a tool (RDKit)?
- Number: Have molecular weight, stoichiometry, yield calculation been verified independently (by hand or library)?
- Source: Has every claim and attribution been verified in a real database?
- Safety: Have the SDS been read for each reagent in the procedure, are there any dangerous combinations?
Note: These four steps apply "always", not "sometimes". 20 situations in which AI is correct do not excuse 1 situation in which it is incorrect; because in chemistry that 1 situation could be an explosion, poisoning or a retracted article.
Common mistakes
- Mistaking the AI output as a “resource”. AI is not a resource, but a generator that tries to remember resources (sometimes incorrectly). It is the source database.
- Trusting the number. Using numbers such as molecular weight, pKa, boiling point without verifying them. These can be calculated/searched deterministically; Asking the language model is the weakest way.
- Outsourcing security to AI. "AI didn't say it's dangerous" doesn't mean it's safe. Hazard assessment is the job of man and SDS.
- Vague prompt. Not giving the molecule by its name but by its structure (SMILES/InChI); This leads to incorrect molecules due to name confusion.
- Trusting in just one try. Asking the same question again in a different way and not checking consistency.
In summary
- AI is a powerful desk assistant in chemistry: it produces code, outlines, descriptions, scan skeletons; but the lab does not do it and does not verify its output.
- There are three inherent risks: fake compound/number, fake attribution, security fallacy.
- Separate and combine general language model and deterministic chemistry tools (RDKit, databases).
- Take each output through the four steps of Structure–Number–Source–Security.
- Good prompt: includes role, explicit structure representation, context, task, validation constraint.
Application task
Choose a molecule that you have studied yourself (e.g. aspirin, SMILES: CC(=O)Oc1ccccc1C(=O)O). Ask AI three things: (1) molecular weight, (2) two reference papers, (3) a synthesis step. Then verify each independently: manually calculate the molecular weight from the formula, confirm citations with DOI, examine the synthesis step with a security eye. Which output came out correct and which required correction? Keep a half-page “verification log.”
checklist
- [ ] I realized that AI is an assistant, not a chemist.
- [ ] I recognize the risks of fabrication, false attribution, and mistrust with examples.
- [ ] I can distinguish between language model and deterministic chemistry tools.
- [ ] I will apply the four steps of Structure–Number–Source–Security to each output.
- [ ] I describe molecules by structure notation (SMILES/InChI), not by name.
- [ ] I kept at least one verification log.