Exploring the Limitations and Strengths of Large Language Models for Chemical Structure Generation

1. Introduction

The advent of Large Language Models (LLMs) has marked a transformative era in artificial intelligence, particularly in their capacity to reshape digital interactions across healthcare, education, and economic sectors.[1] Within educational contexts specifically, these AI systems have introduced paradigm-shifting capabilities in personalized instruction, adaptive tutoring, and dynamic content creation. Contemporary LLM implementations - including ChatGPT (OpenAI), Gemini (Google), Copilot (Microsoft), - have achieved remarkable linguistic fluency, enabling natural language interactions accessible through common digital platforms. Their integration into learning ecosystems offers significant pedagogical advantages: facilitating immediate formative feedback, enhancing learner motivation, and providing intuitive access to complex problem-solving across STEM and humanities disciplines - all through conversational interfaces requiring no technical specialization.[2]

In chemistry education, LLMs offer a compelling yet underexplored opportunity. Chemistry is inherently multimodal: learning often depends on interpreting symbolic, visual, and spatial information. Students are expected to construct and decode molecular structures, Lewis diagrams, orbital representations, and reaction mechanisms—skills that require both conceptual understanding and visual literacy.[3,4,5,6,7] While LLMs excel in linguistic tasks and factual recall, their ability to process or generate chemical imagery accurately remains a significant limitation.[8,9]

Recent studies suggest that although LLMs can describe chemical principles correctly, they frequently struggle with visual conventions and structural accuracy. For example, when prompted to generate Lewis structures or stereoisomeric diagrams, models often produce distorted or chemically invalid results.[10,11] This discrepancy raises concerns about the pedagogical reliability of these tools, particularly in introductory courses where students rely heavily on accurate visual aids to develop foundational understanding.[12]

At the same time, the integration of image input and interpretation capabilities in modern LLMs introduces promising pathways for chemistry education. Multimodal models—capable of processing both text and images—can now identify functional groups in chemical diagrams, recognize reaction types from schemes, and classify molecular properties from image prompts .[13,14] These features may be particularly useful in visual tasks such as distinguishing between polar and nonpolar molecules, understanding the significance of conjugated systems in natural dyes, or interpreting organic transformations like reduction and hydrolysis. Such affordances point to a future where AI could serve as a conceptual scaffold, enabling students to explore chemical ideas interactively, even if the model lacks perfect visual precision.[15,16]

Despite their educational potential, large language models (LLMs) continue to exhibit limitations in solving general chemistry problems that require multistep reasoning, including stoichiometry, significant figures, electron configurations, and Lewis structures.[17] Nevertheless, LLMs have already been incorporated into undergraduate chemistry instruction. For example, students have used ChatGPT as an information source to review and revise their responses during laboratory experiments involving the oxidation of aldehydes and ketones with Tollens’ and Fehling’s reagents. [18] More recently, studies have demonstrated that LLMs such as ChatGPT, Gemini, and Copilot can support a wide range of chemistry and chemometric tasks, highlighting their growing potential as educational and scientific tools. [19,20] However, despite these advances, little is known about how accurately current multimodal LLMs generate chemical representations that are essential for chemistry learning, such as Lewis structures, molecular geometries, and molecular polarity. This gap motivated the present study, which evaluates the image-generation capabilities of ChatGPT, Gemini, and Copilot using representative chemistry tasks.

Despite these advances, there remains limited evidence regarding how effectively current LLMs generate chemical visualizations that are essential for chemistry teaching. Accurate representations of Lewis structures, molecular geometry, and molecular polarity are central to developing students’ representational competence, yet the reliability of AI-generated chemical images has not been systematically evaluated. Furthermore, the rapid evolution of LLMs suggests that their visual capabilities may be improving at a pace that has not been documented. Therefore, this study investigates the ability of three widely available LLMs—ChatGPT, Gemini, and Copilot—to generate and explain chemical representations commonly encountered in introductory chemistry. By comparing their performance over a one-year period (2025–2026), this work provides insight into the progress of multimodal LLMs and identifies their current strengths and limitations for chemistry education.

2. Materials and Methods

The LLMs used in this study were described in Table 1. The same prompts (Q1-Q6) were submitted to each chatbot in June 2025 and again in July

3. Results and Discussion

3.1. Building Thalidomide Structure

Q1: generate thalidomide Lewis structure and show the enantiomers

This section aimed to evaluate whether LLMs can generate chemical structures. As a model compound, thalidomide was selected due to its well-known role in stereochemistry education [21]. Thalidomide exists as two enantiomers: the (R)-thalidomide, which exhibits sedative properties, and the (S)-thalidomide, which is teratogenic and responsible for severe birth defects. Its historical and pharmacological significance makes it a common example for illustrating the importance of chirality in drug development.

Q1 evaluated the LLMs ability in generating the Lewis structure of an important pharmaceutical

In June 2025, ChatGPT (Figure 1) and Copilot (Figure 2) generated incorrect structures. In July 2026, the LLM still generated incorrect structures, Copilot (Figure 3), Gemini (Figure 4), ChatGPT (Figure 5) generated incorrect structures.

This outcome suggests that LLMs are not yet capable of accurately generating molecular structures, particularly for stereoisomeric compounds. However, LLMs correctly identified the pharmacological differences between the (R)- and (S)-enantiomers of thalidomide, consistently recognizing the teratogenic effect of the (S)-enantiomer and the sedative property of the (R)-enantiomer.

3.2. Building the Simple Molecules Lewis Structure

Constructing Lewis structures and analyzing molecular geometry and hybridization are fundamental topics in general chemistry. In this section, Q2 evaluated the ability of LLMs to generate Lewis structures for a selection of molecules commonly taught at the introductory level,

Q2: generate Lewis structures for BF₃, CH₄, NH₃, PCl₃, PCl₅, SF₄, SF₆, XeF₂, and XeF₄.

In July 2025, the LLMs were prompted multiple times—up to seven attempts per molecule in some cases—to assess their consistency and accuracy. However, none of the tools reliably produced correct or pedagogically useful Lewis structures. Many responses contained structural errors, misplaced electron pairs, or nonstandard representations that deviated from accepted conventions in chemistry education. Additionally, the visual formatting of the diagrams was often confusing, making them unsuitable for instructional purposes.

As illustrated in Figure 6, ChatGPT’s generated Lewis structures displayed several inaccuracies, including incorrect electron distributions and ambiguous bond placements. These flaws could potentially mislead students or obstruct their conceptual understanding of molecular structure and bonding.

In July 2026. The LLMs were able to produce the Lewis structures of these simple molecules. ChatGPT (Figure 7) committed a mistake showing the Lewis structure of SF4, but Gemini (Figure 8) and Copilot (Figure 9) generated all structures correctly

After the LLMs generated the correct Lewis structures, they were asked in Question 3 to determine and illustrate the corresponding molecular geometries.

Q3. Illustrate the molecular geometry of each molecule.

Among the three LLMs evaluated, ChatGPT (Figure 10) correctly generated the molecular geometries for all molecules except SF₄, which was incorrectly represented rather than adopting the expected seesaw geometry. Furthermore, it failed to depict the lone pair on the central sulfur atom. Conversely, Gemini (Figure 11) and Copilot (Figure 12) accurately represented all molecular geometries, including the appropriate bond angles and lone pairs, in agreement with VSEPR theory.

3.3. Comparing NH₃ and NF₃ Polarities

Comparing the polarities of NH₃ and NF₃ is a classic exercise in general chemistry that illustrates how molecular geometry and bond dipoles together determine the overall molecular dipole moment. To evaluate the ability of LLMs to explain this concept and generate appropriate molecular representations, the following prompt was used:

Q4. Show the molecular geometry and dipole moments of NH₃ and NF₃. Illustrate their Lewis structures and indicate the direction and magnitude of the molecular dipole moment.

All three LLMs correctly identified NH₃ as the more polar molecule. Their explanations consistently attributed this result to the higher electronegativity of fluorine relative to hydrogen and the vectorial addition of the bond dipoles, which leads to reinforcement of the molecular dipole in NH₃ and partial cancellation in NF₃. The reported dipole moments (July 2025) were also consistent with literature values, approximately 1.47 D for NH₃ and 0.23 D for NF₃.

Despite providing conceptually correct explanations, none of the LLMs generated chemically accurate Lewis structures for these relatively simple molecules. As illustrated in Figure 13, the structure produced by ChatGPT contained formatting artifacts and incorrect bonding patterns that could mislead students and compromise their understanding of fundamental concepts such as Lewis structures, molecular geometry, and molecular polarity. Similar structural inaccuracies were observed in the outputs generated using Gemini and Copilot, indicating that, although current LLMs can successfully explain chemical concepts and retrieve quantitative information, they remain unreliable for generating chemically accurate structural diagrams without expert verification.

In July 2026, ChatGPT (Figure 14) and Copilot (Figure 15) correctly generated the Lewis structures and molecular geometries of NH₃ and NF₃, and both identified that NF₃ is less polar than NH₃. However, neither AI system adequately illustrated the vector addition responsible for the molecular dipole moments. Specifically, they did not clearly demonstrate that the N–H bond dipoles reinforce the lone-pair contribution in NH₃, producing a relatively large net dipole moment, whereas in NF₃ the N–F bond dipoles oppose the lone-pair contribution, resulting in substantial cancellation and a much smaller molecular dipole moment. In contrast, Gemini (Figure 16) accurately illustrated the vector addition and cancellation of the bond dipoles relative to the nitrogen lone pair, providing a clear visual explanation of why NH₃ is more polar than NF₃.

3.4. Comparing Methylene Chlorides Polarity

Q5 addressed the polarity of the chloromethane series, a topic that is widely taught in introductory general chemistry courses because molecular polarity is a fundamental concept underlying intermolecular forces, solubility, and molecular reactivity. All three LLMs correctly predicted the polarity trend of the chloromethane series, CH₃Cl > CH₂Cl₂ > CHCl₃ > CCl₄, and generated molecular dipole moments consistent with the expected values of 1.87 D, 1.60 D, 1.04 D, and 0 D, respectively. However, ChatGPT (Figure 17) and Copilot (Figure 18) had difficulty accurately illustrating the vector addition and cancellation of the bond dipole moments responsible for the resultant molecular dipole moment. In contrast, Gemini (Figure 18) clearly represented the individual bond dipole vectors and their vector summation, providing a more effective visual explanation of how molecular symmetry determines the overall polarity of the chloromethane series.

Q5: Compare the polarity of the methylene chlorides (CCl₄, CHCl₃, CH₂Cl₂, and CH₃Cl). Generate their Lewis structures, illustrate the molecular geometry and bond dipole moments, and indicate the direction and magnitude of the resulting molecular dipole moment for each compound

Figure 17. Comparison of the polarity of CCl₄, CHCl₃, CH₂Cl₂, and CH₃Cl,. Generated using ChatGPT (2026; GPT 5-5).

Figure 17. Comparison of the polarity of CCl₄, CHCl₃, CH₂Cl₂, and CH₃Cl,. Generated using ChatGPT (2026; GPT 5-5).

Figure 18. Comparison of the polarity of CCl₄, CHCl₃, CH₂Cl₂, and CH₃Cl Generated using Copilot (2026; GPT-5.6 Think).

Figure 18. Comparison of the polarity of CCl₄, CHCl₃, CH₂Cl₂, and CH₃Cl Generated using Copilot (2026; GPT-5.6 Think).

Figure 19. Comparison of the polarity of CCl₄, CHCl₃, CH₂Cl₂, and CH₃Cl,. Generated using Gemini (2026; 3.1 Pro).

Figure 19. Comparison of the polarity of CCl₄, CHCl₃, CH₂Cl₂, and CH₃Cl,. Generated using Gemini (2026; 3.1 Pro).

3.5. DDT

Q6: Generate the chemical structure of DDT and identify its stereogenic center

DDT was selected because of its historical importance in organic chemistry and environmental science. Once widely used as an insecticide, DDT was banned in many countries after its persistence in the environment and biomagnification were linked to severe ecological impacts, including the dramatic decline of bald eagle populations due to eggshell thinning.[22] Q6 was intentionally designed as a distractor by asking students to identify a stereogenic center, although DDT does not contain one. Therefore, the correct response required recognizing the absence of a stereogenic center rather than assuming one was present. ChatGPT (Figure 20) and Copilot (Figure 21) correctly generated the molecular structure of DDT and identified that the molecule does not contain a stereogenic center. Gemini (Figure 22) also generated the correct molecular structure and identified a minor component associated with DDT that has stereogenic center.

4. Conclusions

This study evaluated the ability of ChatGPT, Gemini, and Copilot to generate and interpret chemical representations that are fundamental to chemistry education, including Lewis structures, molecular geometry, stereochemistry, and molecular polarity. Comparing the responses produced in 2025 and 2026 revealed a remarkably rapid improvement in the visual capabilities of multimodal large language models. Within a single year, all three systems progressed from generating largely unreliable chemical images to accurately producing Lewis structures and VSEPR geometries for many simple molecules while maintaining strong conceptual explanations of molecular structure and polarity.

Despite these advances, important limitations remain. The evaluated models were still unable to consistently generate chemically correct structures for more complex molecules, such as thalidomide, and frequently failed to illustrate vector addition of bond dipoles accurately, despite providing correct textual explanations of molecular polarity. These findings indicate that current multimodal LLMs can explain chemical concepts more reliably than they can generate chemically rigorous visual representations, highlighting the need for expert verification before AI-generated figures are used in instructional materials.

The rapid improvements observed over only one year suggest that multimodal LLMs are evolving from text-based assistants into increasingly capable tools for generating chemical representations. Although they cannot yet replace dedicated molecular drawing software or textbook-quality illustrations, they have considerable potential to support interactive learning, visualization, and conceptual discussion in chemistry education. Because these systems continue to evolve rapidly, periodic benchmarking will be essential to monitor their capabilities, identify persistent weaknesses, and establish evidence-based recommendations for their effective and responsible integration into chemistry teaching.