Can AI classify a GHS mixture, and where does it break down?
Summary
Only if a qualified reviewer checks it: AI can draft a GHS mixture classification, applying mixture test data, bridging principles and cut-off values in order. The draft holds when every ingredient has a known identity, concentration and classification; it breaks down at bridging judgments, unknown ingredients and concentration ranges.
US OSHA requires test data on the complete mixture first, then bridging principles, then ingredient-based methods such as cut-off values (OSHA, 29 CFR 1910.1200 Appendix A, A.0.4.1, 2024).
A Category 1 carcinogen ingredient at 0.1% or more classifies the whole mixture as a carcinogen (OSHA, 29 CFR 1910.1200 Appendix A, Table A.6.1, 2024).
An ingredient of unknown acute toxicity at 1% or more means the mixture cannot get a definitive acute toxicity estimate (OSHA, 29 CFR 1910.1200 Appendix A, A.1.3.6.2.3, 2024).
The ECHA enforcement guide on bridging allows only one bridging principle per hazard class, with aerosols the exception (AgencyIQ, 2025).
When can AI classify a GHS mixture, and when not?
AI can classify a GHS mixture reliably only for the arithmetic steps, such as summing classified ingredients against a cut-off value, and only when the ingredient data are complete; the judgment steps need a person. A judgment step is any place where the rule says data must be adequate, similar or conclusive. The table below decides, step by step, which parts of a GHS mixture classification an AI draft can be trusted with and what the reviewer checks.
Mixture classification condition | AI draft verdict | Why | What the reviewer checks |
|---|---|---|---|
Test data on the complete mixture | Only if the report is supplied | The draft can read a result, not judge test validity | Method, dose and whether the result is conclusive |
Bridging principle (dilution, batching, interpolation, substantially similar mixtures) | No, draft only | Each principle has conditions on ingredients and toxicity | Which principle, and that every condition is met |
Additivity sums and cut-off values for known ingredients | Yes | Fixed formulas and limits, such as a skin corrosive mixture at 5% or more Skin Category 1 ingredients (OSHA, 2024) | Inputs, units and the edition of the limit table |
Unknown or missing ingredient data | No | Unknown acute toxicity at 1% or more changes the method (OSHA, 2024) | That gaps are declared, not filled in |
Concentration given as a range on the SDS | Only if the range is handled | A trade secret SDS may give a prescribed range (OSHA, 2024) | Upper and lower bound both tested |
The key row is the third one: an AI draft can do the additivity sums and cut-off comparisons, provided a person has confirmed the inputs. The table covers health hazards under US OSHA; physical hazards rest mainly on test results, and EU CLP adds specific concentration limits for some substances, so the reviewer checks the rules of the market the product is sold into.
In what order do mixture test data, bridging principles and cut-off values apply under GHS?
Under GHS, mixture test data come first, bridging principles second and ingredient-based methods such as cut-off values third. US OSHA states the sequence plainly: where test data are available for the complete mixture, "the classification of the mixture will always be based on those data" (OSHA, 29 CFR 1910.1200 Appendix A, A.0.4.1, 2024). Only when test data are missing and bridging cannot be applied do the ingredient methods take over.
The order has an exception that an automated draft can miss. For carcinogenicity, germ cell mutagenicity and reproductive toxicity, OSHA classifies mixtures from ingredient information unless a case-by-case justification supports using the whole mixture (OSHA, 29 CFR 1910.1200 Appendix A, A.0.4.2, 2024).
The EU follows the same tiered approach. CLP Annex I, section 1.1.3 says sufficient data on similar tested mixtures and ingredients "shall be used" for bridging, as quoted in AgencyIQ's summary of the ECHA enforcement guide (AgencyIQ, 2025). That guide adds that only one bridging principle can be applied per hazard class, aerosols excepted, and that similarity must be shown at the level of ingredient identity, not classification alone (AgencyIQ, 2025).
Which GHS mixture classification steps does AI tend to skip or get wrong?
No published evaluation measures AI accuracy on GHS mixture classification specifically, so the failure points below are inference from how the method works. The authors of ChemSafetyBench, a chemistry safety benchmark, note that large language models often generate scientifically incorrect or unsafe responses (Zhao et al., 2024), which is reason enough to treat every step as unverified.
The likely failure points follow from where GHS mixture classification needs judgment rather than arithmetic. First, a draft may jump straight to cut-off values and skip the question of whether test data or a bridging principle should apply. Second, it may apply a bridging principle without checking its conditions, for example that a diluent has an equivalent or lower toxicity classification than the least toxic original ingredient (OSHA, 2024). Third, it may fill an unknown value with a guess. Fourth, it may take one number from an SDS concentration range.
The fix for the last two starts with what the model can read. A model answering from ingredient SDSs you supply can cite them and flag the gaps; Can ChatGPT answer from my company's technical documents? covers what that takes.
What should a reviewer check in an AI-drafted GHS mixture classification?
A reviewer of an AI-drafted GHS mixture classification should check the inputs first, then the method chosen for each hazard class, then the arithmetic. Responsibility does not move to the tool: under US OSHA, chemical manufacturers and importers "shall evaluate chemicals produced in their workplaces or imported by them" to classify them (OSHA, 29 CFR 1910.1200(d)(1), 2024).
Use this checklist for each hazard class:
Every ingredient has an identity, a concentration and a classification taken from a cited ingredient SDS.
Impurities and additives above a cut-off value are included.
The draft states why mixture test data or a bridging principle were or were not used.
Any bridging principle names its conditions and shows each one is met.
Ingredients with unknown data are listed, and the draft states their total percentage.
Strong acids and bases are judged by pH, not by the concentration table (OSHA, 2024).
Specific concentration limits, where the jurisdiction sets them, override generic ones.
Lubinpla's main agent answers from the ingredient SDSs you provide as cited sources, and asks you with a question card when a value is missing instead of filling it in.
Frequently asked questions
Which data does a GHS mixture classification need before AI can start?
A GHS mixture classification needs the full composition with the identity and concentration of each ingredient, each ingredient's classification from its SDS, and any test data on the mixture itself or on similar tested mixtures. Without these, an AI draft can only list what is missing.
Who is responsible for a GHS classification drafted by AI?
The supplier stays responsible. Under US OSHA, chemical manufacturers and importers must classify the chemicals they produce or import, and under EU CLP, suppliers must self-classify hazardous mixtures and document the basis of each decision (AgencyIQ, 2025).
Can AI classify the physical hazards of a mixture?
For physical hazards, classification of a mixture mainly relies on testing (ChemSafetyPro, 2016). An AI draft can read a supplied test report and compare it with the criteria, but it cannot replace the test.
Are cut-off values the same in every jurisdiction?
No. GHS treats cut-off value and concentration limit as interchangeable, while EU CLP defines them differently and sets specific concentration limits for some substances that take precedence over generic ones (ChemSafetyPro, 2016). Check the table of the jurisdiction you sell into.