Hestur AIHestur
    All Articles
    Chemical & Process Industries

    We tried ChatGPT for our safety data sheets. Here is why it did not work

    A general model writes fluent, confident, wrong answers about products it has never seen, with no record of where anything came from. Here is exactly where it fell down, and what to use instead.

    Hestur AI TeamΒ·6 min readΒ·UK focus

    It is a reasonable thing to try. The work is repetitive, the tool is right there, and the first answer it gives looks great. Then you check it against a product you actually know, and the problem becomes obvious. The issue is not that the model is not clever. It is that it is answering the wrong way. This article is practical guidance, not legal advice.

    1

    Why we tried it in the first place

    Safety data sheet work is high-volume and repetitive, which is exactly the kind of task you hope a language model can absorb. Paste in some context, ask for a section, get fluent text back in seconds. For generic phrasing it even helps. The trouble starts the moment the content has to be specifically, verifiably true about your product.

    2

    Where it fell down

    Four failures, and they are structural rather than fixable with a better prompt:

    • No knowledge of your products. It has never seen your formulations, so anything specific to your product is a guess dressed up as an answer.
    • No source of truth. There is nothing for it to be right against. It cannot check itself, because it has nothing to check against.
    • Confident wrong answers. The failure mode is not a blank. It is a fluent, plausible, wrong paragraph that looks exactly like a right one, which is the most dangerous kind of error in a regulated document.
    • No audit trail. A chat has no version history, no citation, and no record of what was produced or approved. If a customer or an auditor asks where a statement came from, you have nothing to show.

    The last one is the quiet killer. Even where the answer happened to be right, there was no way to prove it, no citation, no version, nothing to put in front of a customer or an auditor. That alone rules it out for a controlled document.

    Seen the confident-wrong-answer problem yourself?

    We build the grounded version: answers from your own verified documents, a citation on each one, and a person who approves. A short technical call, no pitch.

    Book a technical call
    3

    What actually works

    The fix is not a different chatbot, it is a different architecture. A grounded system retrieves the relevant passages from your own verified documents and writes the answer from those, with a citation back to each source. If the fact is not in your documents, it says so rather than inventing one. A named person reviews and approves before anything is used. That is the approach we describe in AI for regulatory affairs, and it is a genuinely different tool from the one that failed.

    4

    How to test any tool before you trust it

    Whatever a vendor shows you, run this test: pick a product you know cold, ask the tool a specific question about it, and ask it to show you where the answer came from. A good tool points at a passage in your own document. A bad one produces a confident paragraph and cannot tell you why. The test takes five minutes and it separates the two kinds of tool completely.

    5

    Frequently asked questions

    Can I use ChatGPT to write a safety data sheet?

    Not for anything you will rely on. A general model does not know your product, cannot check itself against a source of truth, and produces confident wrong answers with no audit trail. It can help with generic phrasing, but the substance must come from your own verified data and be approved by a competent person. This is practical guidance, not legal advice.

    Why does it give wrong answers so confidently?

    Because it generates fluent text based on patterns, not on your data. It has no mechanism to know that it does not know, so it fills the gap with something plausible. In casual use that is harmless; in a regulated document it is the core risk.

    So is AI useless for safety data sheets?

    No. The difference is grounding. A system that retrieves from your own verified documents and cites each answer, with a human approving, is a different tool from a general chatbot. The problem is not "AI", it is answering from general knowledge instead of from your source of truth.

    What is the one question to ask a vendor?

    Ask where the answer comes from, and to show you the citation. If the answer is "the model", walk away. If it is "your documents, and here is the reference", you are looking at the right kind of tool.

    6

    Sources

    Safety data sheet requirements are set in regulation. Confirm what applies to your products against the source. This article is practical guidance, not legal advice.

    Want the version that actually works?

    A short technical call on what grounded retrieval from your own documents looks like, with citations and a human approval gate.

    Book a technical call

    ChatGPT for SDS? Read this first

    Book a call