Lubinpla

Writing a safety data sheet with ChatGPT: what it gets wrong

Summary

No, not on its own: ChatGPT drafts the 16-section layout, but hazard classification, exposure limits and composition need source data and qualified human review. The legal duty stays with the manufacturer, importer or supplier who issues the safety data sheet, under US OSHA and EU REACH rules alike, whatever tool wrote the first draft.

  • Under US OSHA rules, the manufacturer, importer or employer preparing a safety data sheet must ensure it accurately reflects the scientific evidence used in the hazard classification, and must add significant new hazard information within three months (OSHA, 2024).

  • Under EU REACH Annex II, a safety data sheet shall be prepared by a competent person who has received appropriate training, including refresher training (European Commission, 2020).

  • NIST AI 600-1 (NIST, 2024) names confabulation, confidently stated but erroneous or false content, as a risk of generative AI.

  • As of 2026-10-23, manufacturers, importers and distributors evaluating mixtures must comply with the 2024 OSHA Hazard Communication amendments no later than November 19, 2027 (OSHA, 2026).

When can ChatGPT write a safety data sheet, and when not?

ChatGPT can write the frame of a safety data sheet, but not the content that carries legal weight: hazard classification, composition, exposure limits and regulatory status. The 16 headings are fixed in law, under 29 CFR 1910.1200(g)(2) in the US (OSHA, 2024) and REACH Article 31 in the EU (European Union, 2006), so a language model reproduces them well.

The table below decides, section by section, where a ChatGPT draft can be used and where it cannot.

SDS section or task

Verdict

Why

16-section layout and headings

Yes, as a draft frame

Headings are fixed by OSHA (g)(2) and REACH Article 31

Section 1 identifier and supplier

Only if you supply the product identifier, recommended use and restrictions, supplier address and emergency number

The model cannot know your label, intended uses or contacts

Section 2 classification

No

Must follow from the evidence, under each jurisdiction's criteria

Section 3 composition and CAS numbers

No

The formulation is in your records, not in public text

Section 8 exposure limits

No

Limits are national and change by jurisdiction

Section 15 regulatory status

No

Status depends on jurisdiction and date

Answering from the manufacturer SDS with a citation

Yes, if a person checks the cited section

The source is the issued document, not the model

The Section 2 row matters most: a wrong classification flows into the label, Section 8 and the training a customer gives its workers.

The Section 1 row is the easiest to fix: give the model your product identifier, recommended use and restrictions, address and emergency number, and it can lay out the section for a person to check. Appendix D lists all four for Section 1 (OSHA, 2024). These limits apply to a safety data sheet you issue. They do not cover reading or summarizing one a supplier issued.

Which safety data sheet sections does ChatGPT get wrong most often?

The safety data sheet sections ChatGPT is most likely to get wrong are those that depend on data it does not have: Section 2 classification, Section 3 composition, Section 8 exposure limits and Section 15 regulatory status. No published study ranks error rates by section, so this ranking is an inference from what each section needs.

NIST describes the underlying risk. Confabulation is "the production of confidently stated but erroneous or false content" (NIST, 2024), and it is "particularly relevant when it comes to open-ended prompts for long-form responses and in domains which require highly contextual and/or domain expertise" (NIST, 2024). A 16-section safety data sheet is exactly that kind of long-form, expert output.

Section 8 shows why. In the EU, Section 8.1 lists the national limit values applicable in the Member State where the safety data sheet is provided (European Commission, 2020), while a US safety data sheet lists the OSHA PEL and ACGIH TLV where available (OSHA, 2024). A limit that is correct in one country can be wrong on a safety data sheet for another, and the model gives no sign of which it used.

Who is legally responsible for the accuracy of a safety data sheet?

The company that prepares and supplies the safety data sheet is legally responsible for its accuracy, not the tool that drafted it. In the US, 29 CFR 1910.1200(g)(5) places that duty on "the chemical manufacturer, importer or employer preparing the safety data sheet" (OSHA, 2024), who must ensure the information reflects the scientific evidence used in the hazard classification.

In the EU, REACH Article 31 places the duty to provide a safety data sheet on the supplier of the substance or mixture (European Union, 2006). Annex II adds that the sheet "shall be prepared by a competent person who shall take into account the specific needs and knowledge of the user audience" (European Commission, 2020), and suppliers must ensure that person is trained.

The US rules are also in transition. As of 2026-10-23, manufacturers, importers and distributors evaluating substances had to comply with the 2024 amendments by May 19, 2026, and those evaluating mixtures must comply by November 19, 2027 (OSHA, 2026). A model trained before an amendment can draft to the older version without saying so.

What should you use instead of a ChatGPT-written safety data sheet?

Instead of a ChatGPT-written safety data sheet, use the manufacturer's issued safety data sheet for the right jurisdiction, and use AI only to find and quote from it. When you supply a product, request the current safety data sheet for the customer's country and language from the manufacturer. When you prepare one yourself, have a competent person or specialist authoring service classify from test data and the composition on file.

AI is useful on the reading side. A question such as "what PPE does the SDS for this product require?" can be answered by quoting Section 8 of the issued document, with the section named, so a person can check it in seconds. The difference between that and a ChatGPT answer from general knowledge is covered in whether ChatGPT can answer from your company's technical documents.

Lubinpla answers a team's questions from its own records and stored documents, such as the manufacturer's safety data sheet, and marks the source behind each sourced sentence so a person can open it before replying.

Frequently asked questions

Can I use ChatGPT to translate a safety data sheet?

Treat a machine translation as a draft for a qualified translator or specialist to check. In the EU, the safety data sheet must be supplied in an official language of the Member State where the product is placed on the market, unless that Member State provides otherwise (European Union, 2006), so the translation carries the same legal weight as the original.

Is a ChatGPT-written SDS illegal?

No rule names the drafting tool. The issuer is responsible for accuracy under 29 CFR 1910.1200(g)(5) in the US (OSHA, 2024), and in the EU the sheet must be prepared by a trained competent person (European Commission, 2020). In the EU an AI draft that no trained competent person prepared and checked does not meet Annex II, and in the US the issuer answers for any error it contains.

Can ChatGPT tell me whether a supplier's SDS is out of date?

Not reliably from general knowledge. Ask the supplier for the current revision and compare the revision date in Section 16 with your copy (OSHA, 2024). Under US rules, significant new hazard information must be added within three months (OSHA, 2024).

Related posts