The pain of direct AI Chat when talking about chemistry
General-purpose AI models are great at image creation but struggle with chemical accuracy. We walk through a real-world example of structure errors and how ChemIllusion delivers precision.
At ChemIllusion, our central goal has always been clear: to add the power of generative AI to how chemists work, without compromising on the scientific rigor that defines the field. We want to enable researchers to create stunning, publication-quality visuals that are as accurate as they are beautiful. I'm not great at conveying this so I recently turned to AI for help. I was away from my desk so I used the ChatGPT app on my phone. My idea was to convey that chemists could benefit from AI but the AI needs to meet them where they work and chemists draw molecules. I thought an image of a chemist drawing on a hood might be a way to convey our central philosophy. And I like the result! In the process of making it I was reminded of another reason we created ChemIllusion. Getting chemically accurate images from direct chat with AI is hard. Here is the process for getting this image:
A vision of AI-augmented chemistry: GPT-Image-1.5 captures the goal, but can it handle the details?
The Illusion of Correctness
The final image above, generated with GPT-Image-1.5, looks good. It conveys a modern lab and a scientist interacting with an AI avatar. But when I tried to use direct AI tools to illustrate specific chemical concepts, the flaws became immediately apparent.
The problem isn't just artistic interpretation—it's that general-purpose AI models simply don't have enough chemical training data. They struggle with the precise terminology and spatial relationships that define organic chemistry. When I tried prompting with specific chemical instructions, the responses revealed this fundamental gap.
The Iterative Struggle
1. My first attempt: "A chemist at a hood drawing chemical structures on the glass of the hood and the hood glass is an AI pointing out features on her molecule"
The AI interpreted "AI" as a robot, not understanding the context of an AI assistant.
2. I clarified: "No robot. Just an avatar and smaller"
Better, but the chemical structure had some problems. An extra line below the NH made it a tertiary amide but either that methyl or the H would need to go. And the peptide bond and hydroxyl labels are incorrect. Maybe if I give it a simpler molecule...
3. More specific: "Smaller avatar in a box or circle. Anisole on hood and avatar pointing to the para position. Saying 'will add here'"
Not bad. Anisole doesn't have a para methyl group but I can work with that. However, the arrow points to a position that is ortho to the methyl but meta to the methoxy. That is not the most activated position for EAS. This shows the AI's lack of understanding of aromatic substitution patterns. I better clarify...
4. Correcting: "Point to the carbon above that instead and the text should say Activated"
Still not right - the AI couldn't grasp the specific carbon position. And we lost the methoxy group. Let's try again...
5. More precise: "Close. But keep the OCH3 group para to the methyl group and have the arrow point to the carbon on the ring next to the OCH3 attachment point"
The AI still struggled with the spatial relationships in the aromatic ring. Try again...
6. Final correction: "So close. But ortho is one more carbon away. Closer to the avatar. Two bonds from the O atom."
Finally got it right. Each iteration revealed more about the AI's limitations. Even with precise chemical terms the model struggled to grasp the spatial relationships fundamental to organic chemistry. General AI models simply don't have enough chemical training data to understand these concepts reliably.
Getting it Right with ChemIllusion
ChemIllusion provides drawing tools any chemist would recognize. And it uses your drawing to communicate with the AI, avoiding these types of errors. With our specialized pipeline—which separates the chemistry logic from the artistic rendering—we can create images that are both visually striking and scientifically accurate.
The finished product: precise geometry and publication-ready style.
The Real Cost of "Free" AI
When choosing between a general-purpose chat tool and a specialized platform like ChemIllusion, consider the actual resources required to get a usable result. The hidden cost of using general AI for science is the time spent correcting its fundamental misunderstandings of chemical concepts and terminology. ChemIllusion removes that burden by ensuring the chemistry is right from the start.
In my experience, what seemed like a quick 1-minute task with direct AI turned into lots of iteration because the model simply lacked the chemical training data to understand terms like "para position" or spatial relationships in aromatic rings.
Experience precision at ChemIllusion