# Building a Chemistry AI Solver That Shows Its Work > Source: > Published: 2026-08-12 19:55:25+00:00 Chemistry problems look deceptively simple to automate. A user types an equation, uploads a worksheet, or asks for a stoichiometry calculation, and the application returns an answer. But the final answer is the least interesting part of the system. When I started working on [Chemistry AI](https://chemistryai.chat/), the harder product question was: **how can a student tell whether an AI-generated chemistry solution is trustworthy?** A chemistry solver can produce a plausible number while still making an error in one of several places: That means a good interface needs to expose the path from the question to the result. The goal is not just more text. It is a sequence of steps a learner can inspect. The workflow I found most useful can be thought of as four stages. Typed questions, photos, and worksheets all need to become a consistent representation. Before solving anything, the system should identify formulas, quantities, units, reaction arrows, charges, and the actual question being asked. This stage should preserve the original input too. If OCR confuses a chemical symbol, users need enough context to spot it. Chemistry questions often omit details that a human silently infers. Is the gas ideal? Is the reaction at standard conditions? Are activities approximated by concentrations? A solver should state those assumptions before applying a formula. This makes the result easier to audit and also helps students learn when a method is valid. For a stoichiometry problem, the conceptual path is usually: Showing this structure is more useful than presenting one long calculation. It also makes debugging easier: if the answer is wrong, we can identify whether the mistake came from chemical reasoning or arithmetic. Different chemistry topics support different checks: These checks do not guarantee perfection, but they turn an opaque response into something testable. Some users want a quick answer so they can check homework. Others want guidance without having the entire solution revealed immediately. That suggests two interface modes: Both modes can share the same underlying structured solution. The difference is how much of that structure the interface reveals at once. The most valuable next step is better uncertainty handling. Instead of confidently solving a blurry photo, the system should highlight the uncertain symbol and ask for confirmation. The same principle applies when a problem has multiple interpretations. I also want verification to become more visible. A small atoms-conserved or units-verified indicator can communicate more than another paragraph of explanation. The main lesson is simple: an educational AI tool should not be judged only by whether it reaches the right answer. It should help users understand, inspect, and challenge the path that produced it. I would be interested to hear how other developers design verification into AI learning tools, especially when the input comes from images or handwritten work.