Can AI Predict Reaction Products? What It Gets Right, and Where It Hallucinates

You draw two reactants on the canvas, add a reagent above the arrow, and ask the AI: “What’s the product?” A second later you get a clean structure, a confident sentence, and a plausible-looking mechanism. The molecule looks right. But you have an experiment to run on Monday, and now you have to decide whether to trust it. So — can AI predict organic reaction products reliably, or are you looking at a well-dressed guess?

The honest answer is “sometimes, and you can usually tell which times.” AI reaction prediction is genuinely useful for a large slice of routine chemistry and unreliable in exactly the places where a chemist’s judgment earns its keep. This post breaks down where it tends to be right, where it tends to hallucinate, and a verification workflow that treats the output as a hypothesis rather than an answer.

How AI reaction prediction works, briefly

Two broad families of tools exist. The first is purpose-built reaction-prediction models — transformer and graph-neural-network architectures trained on large reaction corpora (much of it derived from patent reaction data). The Molecular Transformer, published in ACS Central Science, treats a reaction like a translation problem: reactant SMILES in, product SMILES out, with no hand-coded mechanism. It sits behind IBM RXN for Chemistry, a free interface many chemists have tried.

The second family is general-purpose large language models — the kind that power a chat panel beside your structure editor. These predict the next token in a sentence, which means they predict the most likely-sounding product given everything they have read, not the product a reaction mechanism actually produces. That distinction matters more than any benchmark number.

Neither family runs a quantum-mechanical simulation of your flask. They pattern-match against precedent. When your reaction resembles its training data closely, the match is strong. When it doesn’t, the model still answers — just with no warning that it has left familiar ground.

Where AI can predict reaction products reliably

For well-precedented transformations on common substrates, predictions are frequently correct on the major product. These are the cases that dominate both the literature and the training data:

• Common functional-group interconversions — ester hydrolysis, reduction of a ketone to an alcohol, oxidation of a primary alcohol, Boc protection and deprotection. The transformation is unambiguous and densely represented in precedent.
• Workhorse couplings on simple substrates — amide bond formation, a Suzuki coupling between an unhindered aryl halide and a boronic acid. The bond that forms is predictable when there is only one reasonable site.
• Brainstorming and recall — “what reagents convert a nitrile to a primary amine?” or “name a few ways to make this amide.” Here the AI is acting as a fast index into reaction precedent, and you verify the suggestion before committing.
• Naming and lookup — identifying a reaction class, recalling a named reaction, or surfacing the typical conditions. This is recall, not prediction, and it is comparatively safe.

In these situations the AI usually gets the connectivity of the major product right. That is a real time-saver when you treat it as a first draft.

Where it fails and hallucinates

The failures cluster in the parts of chemistry that depend on context the model cannot reliably infer from connectivity alone.

• Selectivity edge cases. Regio-, stereo-, and chemoselectivity are where prediction quality drops. When a substrate has two plausible reactive sites, or the question is which diastereomer or enantiomer forms, the model is choosing among close alternatives — and selectivity depends on subtle steric and electronic effects, catalyst, and conditions. Treat any predicted stereocenter or regiochemical outcome as unverified.
• Novel substrates outside the training distribution. A heavily functionalized intermediate, an unusual heterocycle, or a transformation with thin literature precedent pushes the model off its data. It will still answer confidently; the confidence is not a signal of correctness.
• Fabricated reagents and conditions. A general-purpose LLM can invent a plausible-sounding reagent, a temperature, or a catalyst loading that reads like real chemistry and isn’t. The format mimics a genuine procedure, which is exactly what makes it hard to catch.
• Overconfident wrong products. The most damaging failure: a clean structure and a fluent rationale for a product that the reaction does not give. There is no mechanistic guarantee underneath — the sentence sounds right because the model is good at producing sentences that sound right.

A practitioner’s verification workflow

The reliable way to use AI reaction prediction is to treat the output as a hypothesis and run it through a short checklist before it touches a flask or a report.

1. Treat it as a hypothesis, not a result. Read the predicted product as “the model suggests this” — a starting point for your own reasoning, not a conclusion.
2. Verify the structure is valid. Check valence and that the SMILES parses to the molecule you think it is. A quick way to do this is to render the proposed structure and eyeball it — the free SMILES-to-structure converter turns the AI’s SMILES into a drawing you can inspect for valence errors and unintended connectivity.
3. Check against precedent. Does this transformation appear in the literature on similar substrates? Search by reaction class or by analogous structures. If the AI cited a reagent or paper, confirm it exists and says what the answer claims — do not assume a named citation is real.
4. Confirm the selectivity reasoning yourself. If the substrate has more than one reactive site or a stereochemical outcome, work through the regio- and stereoselectivity using mechanism, not the model’s assertion. This is the step most worth your time, because it is where prediction is weakest.
5. Never promote “AI Suggested” to “Computed.” A property a deterministic tool calculated and a product an LLM guessed are different kinds of claim. Keep them labeled differently in your own notes the way a good tool keeps them labeled differently on screen.

How ChemStitch handles the honesty problem

ChemStitch’s reaction prediction and retrosynthesis suggestions are explicitly labeled AI Suggested — a yellow badge that marks them as generative output to verify, not deterministic fact. That label is deliberate. Properties computed by RDKit (molecular weight, logP, TPSA, Lipinski) carry a separate Computed badge, and the product never renders an AI guess as if a deterministic tool produced it.

The reasoning is the same one driving this whole post: chemists are professionally skeptical of AI output, and mislabeling a guess as a computed fact erodes trust permanently. A predicted product shown as “AI Suggested” invites the verification it needs; the same product dressed up as “Computed” invites a mistake. Honest provenance is the difference between a tool that helps you think and one that quietly leads you wrong.

If you want to go deeper on planning routes by hand rather than leaning on a model, see the manual step-by-step retrosynthesis workflow, the practice of choosing a disconnection yourself, and how synthons and synthetic equivalents turn a disconnection into real reagents. For tooling, we’ve compared the AI retrosynthesis software medicinal chemists use and written up reaction-mode synthesis planning.

So, can you trust it?

Use AI reaction prediction the way you’d use a sharp, fast colleague who has read more papers than anyone alive and occasionally makes things up with total confidence. For routine transformations on familiar substrates, it’s a genuine accelerator. For selectivity calls, novel substrates, and anything heading into a procedure or a filing, it’s a hypothesis you verify. The model can’t tell you which case you’re in — but you can, and the verification workflow above is how. Draw a structure in ChemStitch’s SMILES-to-structure converter and check the AI’s next answer before you trust it.