BRENNEN

Industries / Mining and materials / Chemicals and petrochemicals / Specialty chemicals

Safety data and regulation

Safety data and regulation is the paperwork and law around a chemical's hazards: the labels and safety data sheets that travel with each product, and the rules on which substances may be sold at all.

7

Milestones achieved, from 1960 to 2024.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1960

    Flavor makers' expert panel

    The Flavor and Extract Manufacturers Association (FEMA) began a program in 1959 to assess the safety of flavor ingredients and set up its Expert Panel of scientists in 1960. The panel judges whether each ingredient is "generally recognized as safe" (GRAS), a status under the US Food Additives Amendment of 1958 that lets an ingredient be used without separate approval as a food additive. The joint FAO/WHO expert committee on food additives (JECFA) has published group summaries on about 2,200 FEMA GRAS flavoring substances.

    femaflavor.org

    Credit

    FEMA and its Expert Panel

  2. 1978

    Lead paint banned

    The US Consumer Product Safety Commission banned paint for consumer use containing more than 0.06% lead by weight of the dried paint, and toys and furniture bearing such paint, for products made after 27 February 1978. It found an unreasonable risk of lead poisoning in children. The ban covers paint wherever consumers can reach the dried surface, including homes, schools and playgrounds, but not paint for cars and boats. Congress cut the limit to 0.009% from 14 August 2009.

    law.cornell.edu

    Credit

    US Consumer Product Safety Commission

  3. 1983

    Data sheets become law

    The US Occupational Safety and Health Administration (OSHA) issued its Hazard Communication Standard on 25 November 1983 for manufacturing and extended it to all industries in 1987. It requires hazard information to travel with chemicals on labels and on material safety data sheets, documents that list a product's hazards, safe handling and first aid.

    LessonRequiring a sheet did not make it right. In a 1991 study OSHA commissioned, only 11% of the sheets examined were accurate in all four areas checked: health effects, first aid, protective equipment and exposure limits.

    osha.gov

    Credit

    OSHA

  4. 1991

    Asbestos ban struck down

    In 1989 the US Environmental Protection Agency (EPA) used the Toxic Substances Control Act of 1976 (TSCA) to ban asbestos in almost all products. On 18 October 1991 the US Court of Appeals for the Fifth Circuit sent the rule back. It found that EPA had not shown a ban was the least burdensome way to reach an acceptable risk, as TSCA required, and had not weighed the hazards of likely substitutes. EPA says the decision significantly weakened its authority under TSCA. Congress rewrote TSCA in 2016, and in March 2024 EPA banned the ongoing uses of chrysotile asbestos in the first rule finalized under the 2016 amendments.

    LessonA law that makes a ban the hardest rule to justify can block bans for decades; this one took a new law.

    law.resource.orgepa.gov

    Credit

    US Court of Appeals for the Fifth Circuit (Corrosion Proof Fittings v. EPA); US EPA

  5. 2003

    One hazard system worldwide

    A UN expert committee adopted the Globally Harmonized System of Classification and Labelling of Chemicals (GHS) in December 2002, and the UN Economic and Social Council endorsed it in 2003. The GHS sets shared rules for sorting chemicals into hazard classes and a fixed 16-section layout for safety data sheets, so sheets from every country that adopts it follow the same order. The US adopted it in 2012 and moved to a later GHS revision in 2024.

    osha.govosha.gov

    Credit

    UN Committee of Experts on the Transport of Dangerous Goods and on the Globally Harmonized System

  6. 2007

    No data, no market

    The EU's REACH regulation entered into force on 1 June 2007, and its registration rules applied from 1 June 2008. Its Article 5, headed "No data, no market", says a substance may not be made or sold unless it has been registered where the regulation requires it. Registering means the makers and importers file data on the substance's properties and uses.

    LessonA rule binds only what inspectors check. In ECHA's REF-12 enforcement project, 2,603 controls in 29 countries found that one in three substances in imported mixtures lacked the required registration, a June 2026 Council of the EU paper reports.

    legislation.gov.uklegislation.gov.ukdata.consilium.europa.eu

    Credit

    European Parliament and Council of the EU (Regulation (EC) No 1907/2006); European Chemicals Agency (ECHA)

  7. 2024

    A third fail inspection

    In an EU-wide project of ECHA's Enforcement Forum, called REF-11, inspectors in 28 EU and EEA countries checked more than 2,500 safety data sheets in 2023 for completeness and correctness. They found that 35% did not comply with the law. The final report came out in December 2024.

    kft.de

    Credit

    ECHA's Enforcement Forum and national inspectors

Watch for

  1. PFAS restricted as a class

    PFAS (per- and polyfluoroalkyl substances) are a family estimated at more than 10,000 substances that persist in the environment. In January 2023 authorities in Denmark, Germany, the Netherlands, Norway and Sweden proposed restricting them as a class under REACH. In March 2026 ECHA's Committee for Risk Assessment backed an EU-wide restriction on making, using and selling PFAS and estimated it could cut PFAS emissions by about 96% over 30 years. The draft opinion of ECHA's Committee for Socio-economic Analysis found a full, immediate ban likely not proportionate and favors exemptions for named uses; its final opinion is expected by the end of 2026.

    Achieved whenthe European Commission adopts the restriction and it is published in the EU's Official Journal as an entry in Annex XVII of REACH.

    arnoldporter.com

    Claimed by

    ECHA's Committee for Risk Assessment, March 2026

  2. AI checks data sheets

    SDS Manager, which sells software for managing safety data sheets, says its AI validation, released in October 2026, reads all 16 sections of every sheet in a customer's library and flags problems such as a hazard class listed without matching precautionary statements, or classifications that disagree. It says an internal analysis of 100,000 commonly used sheets found 4.3% missing section 15, on regulations, and 44% with incomplete data in sections 1 and 2. It publishes no test of how often its flags are right.

    Achieved whenan outside party has experts review a random sample of the tool's passes and flags and publishes how many flags were right and how many problems it missed.

    sdsmanager.com

    Claimed by

    SDS Manager, October 2026

Open

  1. 01

    Audit data sheets at scale

    An open checker that reads the public safety data sheets of mixtures such as paints, glues and sealants, takes each ingredient and the low end of its concentration range from section 3, works out the hazard classes the EU's classification rules (the CLP regulation) require at those amounts, and flags any sheet whose hazard section, section 2, leaves one out. Then a published audit of 1,000 sheets.

    The brief
    Credit · unclaimed
  2. 02

    Data sheets workers understand

    A tool that turns a safety data sheet into a one-page summary for the person using the product (what can hurt you, what to wear, first aid, spills), where every sentence cites the section of the sheet it came from and sentences without a matching passage are rejected. Then a reading test that compares the summary with the full sheet.

    The brief
    Credit · unclaimed

Open 01 · the brief

Audit data sheets at scale

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

An open checker that reads the public safety data sheets of mixtures such as paints, glues and sealants, takes each ingredient and the low end of its concentration range from section 3, works out the hazard classes the EU's classification rules (the CLP regulation) require at those amounts, and flags any sheet whose hazard section, section 2, leaves one out. Then a published audit of 1,000 sheets.

Why it is still open

Inspectors still check sheets by hand: REF-11 covered a little over 2,500 sheets across 28 countries and found 35% non-compliant. OSHA wrote in 2012 that the size of the accuracy problem was unclear because only very limited numbers of sheets had been evaluated. Published AI work so far measures how well models copy data out of sheets, not whether the sheets are right; in a 2026 benchmark the best model extracted 84% of fields correctly. A vendor tool that flags inconsistent classifications in customers' libraries appeared in October 2026, with no accuracy test and no published audit (see Watch for).

kft.deosha.govarxiv.org

Built with what exists

Safety data sheets that makers publish free on their websites; the EU's list of harmonized classifications in Annex VI of the CLP regulation; go-clp, an open-source (Apache-2.0) library that applies CLP's mixture rules and carries Annex VI's concentration limits by date; an AI model for extraction from long documents, such as Muse Ember (see the Note "The jobs agents repeat"); a coding agent that tests its own work, such as Inkling 2 Small (see the Note "Checked without being asked").

github.com

  1. First step

    Download 50 sheets for wall paints and wood glues from five makers and check section 2 against section 3 by hand with go-clp, to settle which rules the checker applies.

  2. Done looks like

    1,000 sheets from at least 50 suppliers checked. A person with chemical regulatory training reviews every flag, and at least 90% of flags hold up. A random 100 unflagged sheets are also reviewed to estimate misses. The share of sheets with a missing hazard is published with a 95% confidence interval, along with the code.

Open 02 · the brief

Data sheets workers understand

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A tool that turns a safety data sheet into a one-page summary for the person using the product (what can hurt you, what to wear, first aid, spills), where every sentence cites the section of the sheet it came from and sentences without a matching passage are rejected. Then a reading test that compares the summary with the full sheet.

Why it is still open

OSHA's 2012 rule cites a 1991 study in which 91 workers understood about two-thirds of the health and safety information on the sheets they were given, and a survey in which 46% of workers said the sheets are too long. Published AI studies on these sheets extract data for chemical inventories and exposure assessment, and searches for this file found none that tests whether workers understand an AI summary better.

osha.govpmc.ncbi.nlm.nih.gov

Built with what exists

Published safety data sheets for common products; a general AI model to write the summaries; a citation checker built with an agent toolkit, such as Cursor's Agent SDK, whose example reviewer cites the note behind each finding (see the Note "Findings that cite notes"); an online survey tool for the reading test.

  1. First step

    Write the question set for five sheets for household paints and glues and try it with ten volunteers reading the full sheets.

  2. Done looks like

    20 sheets summarized. Two safety professionals confirm that no summary drops or contradicts a hazard in section 2. At least 60 adult volunteers are randomly given either the full sheet or the summary and answer the same questions on route of exposure, health hazard, protection and where to get help. Scores for both groups are published against a target, set in advance, of at least 15 points higher for the summary with no hazard missing.