Why a Compliance Verdict Should Never Come From an LLM

A wave of AI-powered SDS and hazard-classification tools has arrived, most of them built on large language models, most of them promising to "automate" classification by letting a model read your formulation and tell you the answer. The pitch is seductive, and the underlying technology is genuinely impressive. It is also the wrong tool for the one job that matters most in this domain: producing a defensible classification verdict.

This isn't an anti-AI argument. It's an argument about where AI belongs in the pipeline — and where it categorically does not.

What a classification verdict actually is

When you classify a mixture as, say, Acute Tox. Category 4 (Oral) or Skin Irrit. Category 2, you are not offering an opinion. You are asserting that a specific set of inputs — component identities, concentrations, each component's classification and relevant cut-off values, M-factors, additivity rules — when run through the criteria written in the regulation, produces exactly that result. The verdict is a deterministic function of the inputs and the rules. Given the same data and the same regulation, there is one correct answer, and it is reproducible by anyone who applies the rules correctly.

That property — same inputs, same rules, same output, every time — is the whole game in compliance. It's what makes a classification auditable. It's what lets you stand behind an SDS when a customer's regulatory team, or an inspector, asks you to show your work.

What an LLM does instead

A large language model does not apply rules. It predicts likely text. When you ask an LLM to classify a mixture, it generates the most statistically plausible answer-shaped string given its training and your prompt. Often that string is correct, because the model has seen a lot of correct classifications. Sometimes it is confidently, fluently wrong — and the failure mode is the dangerous one, because the wrong answer is dressed in the exact vocabulary and structure of a right one.

Three properties make this disqualifying for the verdict step:

It is non-deterministic. Ask the same question twice and you can get two different answers. For a compliance artifact, "it depends on the sampling temperature" is not an acceptable property.

It is not auditable. When a deterministic engine classifies something Skin Corr. Category 1B, it can name the cut-off it crossed and the component that drove it. When an LLM does it, the honest answer to "why?" is "because that text was probable." You can ask it to generate a rationale, but a post-hoc rationale generated by the same probabilistic process is not a citation — it's a plausible-sounding story, which may or may not reflect anything real.

It fails silently. A rules engine that hits a case it doesn't cover throws an error or returns "unclassified — insufficient data." An LLM almost never says "I don't know." It fills the gap. In compliance, a confident guess is worse than a blank, because a blank gets escalated to a human and a confident guess gets shipped.

The cautionary tale here isn't hypothetical. The most ambitious science-trained model ever publicly demoed was pulled within days precisely because it produced authoritative-sounding scientific output that was frequently fabricated. Training on the right corpus buys you the register of correctness, not correctness itself. That gap is exactly where compliance liability lives.

Where AI absolutely does belong

None of this means there's no role for AI in an SDS platform. There's a large and legitimate one — it's just upstream of the verdict, never inside it.

• Data extraction. Pulling component names, CAS numbers, and concentrations out of a messy supplier SDS PDF is a language task. An LLM is excellent at it. Crucially, its output here is proposed input, which a human or a deterministic check can verify before it touches a classification.
• Translation. Rendering an authored SDS into another jurisdiction's language is a language task. Again: the hazard verdict is fixed by the rules engine first; translation only changes the words it's communicated in.
• Drafting and summarization. Narrative sections, first-draft precautionary phrasing, plain-language explanations — all fair game, all reviewable.

The line is simple and bright: AI can read, extract, translate, and draft. It never decides the classification. The classification comes from a deterministic engine applying the published criteria, and that engine can show you the rule path for every result.

The architecture in one sentence

Use the probabilistic tool for the probabilistic problems — reading unstructured documents, moving between languages — and use a deterministic rules engine for the one thing that must be reproducible and auditable: the verdict.

This is the design principle behind GHSLogic. The classification logic is deterministic and rule-driven; AI is used only for extraction and translation, never to render a hazard determination. When the engine says a mixture is Category 3, it can tell you which cut-off, which component, and which rule — every time, the same way.

If you're evaluating an AI-based SDS tool, there's one question worth asking the vendor directly: when your tool assigns a hazard category, is that decision made by a rules engine or by the model? The answer tells you whether you're buying a compliance tool or a very articulate guess.

Try the deterministic engine yourself: run a mixture through GHSLogic's free calculators and watch them show the math behind every category it assigns.