Does AI “Understand” Chemistry? A Sober Look at Expectations and Reality

Introduction: Can AI Become a “Chemist”?

In recent years, we’ve seen more and more headlines like “AI discovers new drugs” and “AI automatically designs chemical reactions.” However, the more deeply one has studied chemistry, the more likely one is to feel a sense of unease, wondering, “Can we really say it ‘understands’ that?”
If AI doesn’t truly “understand,” what should we expect from it, and where should we exercise caution?

This guide breaks down how AI handles chemistry from the ground up, distinguishing between “why it appears to understand” and “where it actually still has weaknesses.” We’ll proceed without rushing to a conclusion, so that we can calmly determine appropriate limits for its use.

How Does AI Handle Chemistry?

Chemistry Is Input as a “Form of Expression” Rather Than a “Concept”

Modern AI deals not so much with chemistry itself as with the formats used to represent it. Typical examples include the following:

Molecular structures: SMILES, InChI, molecular graphs (atoms = nodes, bonds = edges)
Reactions: Input and output of reaction equations (reactants → products) and conditions (solvent, temperature, catalyst, etc.)
Physical properties: Numerical labels such as melting point, boiling point, solubility, toxicity, and reactivity

AI learns “patterns found in large datasets” from these inputs and returns predictions for unknown molecules or reactions. What is happening here is behavior closer to pattern extraction and analogy than to deduction from first principles.

It Does Not “Understand” Physical Laws from Within

AI does not inherently embody the laws of thermodynamics or quantum chemistry.
The fact that its output sometimes appears to align with these laws is simply because much of the training data consists of “real-world results that follow physical laws”; this does not constitute direct proof that the model autonomously possesses these laws.

Misunderstanding this point leads to overconfidence—the belief that “since AI understands, we can trust it.”

Why Does It “Seem to Understand”?

It Can Generate Explanations That Sound Like They Were Written by a Chemist

Generative AI can reproduce the writing patterns found in research papers and textbooks with high accuracy. As a result,

The choice of technical terms seems authentic
It follows a typical order of explanation
It provides plausible justifications

This results in output that makes it appear as though the AI “understands what it’s talking about.”
However, the fluency of the text is no guarantee of understanding. In fact, the more fluent the text, the greater the risk that verification will be overlooked.

It Appears Strong in Fields Where “Hits” Are Common

In certain areas of drug discovery and materials exploration, results have been reported showing that AI can accelerate the search for candidates. This is because statistical predictions work effectively when evaluation metrics are relatively clear and a large body of historical data is available.
However, in many cases, the AI cannot fully explain the mechanism behind “why” a candidate is good, so the final judgment remains with humans and experimentation.

Areas Where AI Struggles

Vulnerable to Data Bias and Missing Data

Chemical data suffers from structural bias. This is because successful examples are easily shared, while failed examples (negative cases) are rarely shared.
This bias causes models to be trained in a “convenient world,” making them prone to overlooking real-world “failure conditions.”

Inadequate coverage of conditions under which reactions do not occur
Omission of on-site factors, such as scale-up and the effects of impurities
Vulnerability to special conditions (high pressure, extremely low temperatures, trace moisture, etc.)

AI is particularly prone to confidently making incorrect predictions in areas where “data is scarce” or “failures are not documented.”

Poor at Explaining Causality and Mechanisms

While AI is good at identifying correlations, it struggles to explain causality (why something happens) in terms of mechanisms.
Human theoretical understanding and expert verification remain essential for evaluating the validity of reaction mechanisms, transition states, and electron flow.

A Practical Approach to AI Utilization

Design It as an “Aid,” Not a “Replacement”

It is more realistic to use AI as an aid for exploration and organization rather than as a substitute for chemists.

Literature review and summarization: Overview of related research, clarification of terminology, and extraction of comparison criteria
Expanding hypotheses: Enumerating candidates, proposing combinations of conditions (subject to verification)
Preparing for predictions: Prioritization, narrowing the search space

The key here is to be mindful of treating the AI’s output as “candidates” rather than “definitive answers.”

Determine the Verification Workflow First

When implementing AI, operational design is more important than accuracy.

Which outputs should be treated as “suggestions,” and which should be considered for “adoption”?
How to conduct refutation (negation) tests
Who bears ultimate responsibility (including safety, regulations, and ethics)

If these boundaries are left ambiguous, the AI will be praised only when results are achieved, while frontline staff will become exhausted when failures occur.

Internal Links

This article provides a detailed explanation of the “plausibility” of generative AI and verification procedures
https://example.com/ai-verification-playbook
This article provides a detailed explanation of design principles for realistically advancing the use of AI in research and development
https://example.com/rnd-ai-practical-design

Summary: Does AI “Understand” Chemistry?

As a summary of the current state of knowledge, the following conclusions are reasonable:

Rather than “understanding” chemistry, AI is learning patterns in chemical data
Fluent explanations are not evidence of understanding; rather, they tend to lead to insufficient verification
Its strengths lie in “prediction” and “exploration support,” while its weaknesses are “explaining causality and mechanisms” and “vulnerability to missing data”
In practice, it is safest to use AI as a supplementary tool and to design verification workflows and clear lines of responsibility in advance

AI is not a magical chemist, but when positioned correctly, it can serve as a tool to accelerate research and development. The most solid first step is to start small by focusing on three areas—“generating candidates,” “organizing,” and “refuting”—and then systematically bridge the gap between expectations and reality through quantitative analysis.

Chemistry × AI is a field where results are determined not by flashy headlines but by the nuances of operational design. Therefore, simply adopting the perspective of “verification, not understanding” at this stage will surprisingly increase your chances of success in practice.