As a patent attorney who has been working with chemical inventions over the past two decades, I was intrigued by AI and jumped onto the wave when it first came out a few years ago. At first, most of the tools were not up to par. Further down the line, it seems to have occurred that more and more AI tools started appearing in the field, but most were focused on mechanical inventions. Chemical inventions seemed to be the bottleneck, or the holy grail. It seemed to be what was the toughest to crack, to get AI to understand chemistry.

Over the past year, I once again spent some time experimenting with some of the newer generation of AI tools that are starting to appear in our field. As they get more and more advanced, what has surprised me is not just how much improvements these tools have been getting at understanding the language of chemistry.

For a long time, chemistry felt like a natural boundary for general AI systems. Our work is not purely textual. It is structural, spatial, and deeply contextual. You cannot infer a molecule’s behaviour from words alone. And yet, something has shifted recently, largely because of how AI models are starting to interact with structured chemical notations, or what is now known as SMILES.

SMILES has existed for decades. Many of us learned about it early in our training, even if we did not use it daily once graphical drawing tools became standard. At its core, SMILES is simply a way of encoding molecular structures into strings of characters. Take aspirin, for example, a molecule that I synthesized as a freshman student. In SMILES form, it’s written as: C(=O)OC1=CC=CC=C1C(=O)O.

Whereas most chemists would nowadays use chemical drawing programs to represent chemical structures, SMILES has had always been an easy format to store molecular information. Historically, SMILES was mainly useful for storage and data transfer. But now, AI tools have given SMILES a new purpose: the translation of the molecular structure into characters. By turning molecular structures into machine-readable character sequences, SMILES has effectively created a bridge between chemistry and language-based AI systems.

Whereas the generic AI tools often struggle with molecular structure input, specialized AI tools can “read” a SMILES string as a real molecular structure. It can spot viable structure variants, highlight inconsistencies, suggest sensible alternatives, and compare with known prior art molecules.

That is not trivial. A large part of chemical drafting involves navigating large, structured spaces of possibilities. If these specialised AI-tools can help us map that space more systematically, this can change how we approach chemical drafting.

That said, I remain sceptical of any narrative that suggests AI is “understanding” chemistry in the human sense.

Chemistry is ultimately about mechanisms, experimental nuance, and accumulated judgement. No model understands why a reaction failed in the lab at 3am, or why a seemingly minor substitution changes everything. But that does not mean these tools are not useful. In fact, used properly, they can be extremely powerful: particularly for tasks that are structurally repetitive or data-heavy.

If machines can help with structure consistency checks, variant exploration, or early prior art mapping, that frees us to focus more on the parts of the job that are genuinely intellectual: invention framing, claim architecture, and anticipating how an examiners (or competitors) might think about the chemistry.

Stepping back, I think what we are seeing is part of a broader shift. AI is moving away from being a general productivity layer and toward becoming domain-native. In chemistry, that means systems that can operate on structures, not just describe them.

AI will never be able to replace chemical patent attorneys. But it is starting to meet us closer to where we work.