BRENNEN

Industries / Industrial equipment and services / Industrial services / Industrial hygiene

Exposure records

Exposure records are the measurements of what workers breathe, touch and hear, kept by employers, inspectors and national databases, together with the statistics used to decide from a few samples whether a group of workers is safe.

7

Milestones achieved, from 1972 to 2019.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1972

    Germany pools its measurements

    Since 1972 the inspectors of Germany's statutory accident insurers, the Berufsgenossenschaften, have sent their workplace samples of chemicals to one institute, BIA, which analyses them and stores each result with up to 150 details about the workplace and method. By 2001 its MEGA database held about 1,000,000 measurements of more than 400 substances; in 1999 alone, 31,000 measurements were taken in 4,000 companies. France began a similar database, COLCHIC, in 1987 and had 400,000 results for 600 substances by 2001. Both sets of authors note limits: MEGA's sites are chosen where high exposure is suspected, not at random, and COLCHIC's use was held back by missing details.

    LessonA record is only as useful as the details stored with it.

    ebi.ac.ukebi.ac.uk

    Credit

    the Berufsgenossenschaften and BIA; France's regional health insurance laboratories and INRS

  2. 1980

    Workers may see their records

    OSHA's rule of 23 May 1980, in force from 21 August 1980, gave US workers and their representatives the right to see records of their exposure to toxic substances and their medical records. It replaced a 1978 interim rule, "Preservation of Records", which required employers to keep them. Under the rule as it stands, employers must keep exposure records for at least 30 years and medical records for the length of employment plus 30 years, and must give access within 15 working days.

    osha.govosha.gov

    Credit

    OSHA

  3. 2009

    Hygienists' guesses tested

    Logan, Ramachandran, Mulhausen and Hewett asked occupational hygienists to judge the exposure of tasks shown on video, with 0 to 8 air samples each, and checked the 4,386 answers against a statistical analysis of the data. With samples to look at, the hygienists were right 43% of the time, and 63% after a short training in simple rules for reading small, skewed sets of results. A 2022 review puts hygienists' accuracy without any samples at about 30%, where chance is 25%.

    LessonJudging exposure from a few samples needs statistics, not instinct.

    ebi.ac.ukpmc.ncbi.nlm.nih.gov

    Credit

    Logan, Ramachandran, Mulhausen and Hewett

  4. 2010

    OSHA posts its lab results

    OSHA inspectors have taken workplace samples since 1972, and from 1979 the results went into a computer system called IMIS. In 2010 OSHA put more than 1 million personal sample results from its Salt Lake City laboratory online as the Chemical Exposure Health Data. Comparing the two for lead, Lavoué, Friesen and Burstyn found only 39% of records in both. Results below the detection limit made up 71% of the records found only in the laboratory data, but 42% and 46% of the records in IMIS, which the authors read as inspectors leaving many results that found nothing out of IMIS.

    LessonA database that leaves out results that found nothing makes exposures look higher than they are.

    ebi.ac.uk

    Credit

    OSHA; analysis by Lavoué, Friesen and Burstyn

  5. 2012

    Nineteen countries pool data

    For SYNERGY, a pooled study of lung cancer, researchers gathered 356,551 measurements of five lung carcinogens (asbestos, chromium, nickel, PAHs and respirable crystalline silica) from 18 European countries and Canada, entered in one standard format and reaching back to 1951. The database, called ExpoSYN, was built to feed a job-exposure matrix, a table that gives the likely exposure for each job, country and period.

    ebi.ac.uk

    Credit

    Peters, Kromhout and 18 co-authors from European and Canadian institutes

  6. 2018

    A stricter test in Europe

    EN 689, Europe's standard method for testing whether a group of workers complies with an exposure limit, was revised in 2018, replacing the 1995 version. D'Errico and colleagues applied both versions to 1,383 respirable dust measurements from 210 groups of workers in industrial minerals companies. The new statistical test found almost twice as many groups non-compliant as the old one: 29% against 17%.

    pmc.ncbi.nlm.nih.gov

    Credit

    the European Committee for Standardization (CEN); study by D'Errico and colleagues

  7. 2019

    Free statistics for few samples

    Expostats, a free web toolkit from Lavoué and seven co-authors, answers the usual questions about a set of measurements: whether a group complies with a limit, whether some workers are likely to be overexposed, and whether a control made a difference. It uses Bayesian statistics, which give the probability of each answer and handle results below the detection limit. Its authors call it the first freely available toolkit of its kind. In 2026 other researchers built an open strategy simulator, SIHM, on its code, and reported that fewer than one-third of US practitioners use such robust assessment strategies.

    ebi.ac.ukebi.ac.uk

    Credit

    Lavoué, Joseph, Knott, Davies, Labrèche, Clerc, Mater and Kirkham

Watch for

  1. AI fills chemical inventories

    TNO and IDEWE, a Belgian prevention service, report that their pipeline of large language models copied fields such as company name, product name and hazard statements from safety data sheets into an online chemical inventory with an average accuracy of 0.83 against copying by hand. A Seoul National University group reported that Gemini 2.5 Pro listed the substances and their amounts in the data sheets of 59 cleaning products with an F1 score, a combined measure of missed and wrong items, of 1.00, against 0.94 for ChatGPT-4o.

    Achieved whensomeone other than the builders tests such a pipeline on a large public set of data sheets it was not tuned on and publishes error rates for each field.

    publications.tno.nlebi.ac.uk

    Claimed by

    TNO and IDEWE, November 2025; Lee, Lee and Lee of Seoul National University, May 2026

Open

  1. 01

    Rebuild a database by AI

    A pipeline that reads published exposure studies and copies every set of measurement results (agent, job, industry, year, number of samples, average and spread, limit, sampling method) into one standard table with the page and sentence each value came from, tested first against a database that people built by hand.

    The brief
    Credit · unclaimed
  2. 02

    Score AI exposure judgments

    A public test of 200 written exposure scenarios, each taken from a published study: the task, chemical, amounts, controls and duration, with the measured results held back. AI models, and for comparison volunteer hygienists, place each scenario in one of four exposure categories, each a band of its exposure limit. The answer key comes from the study's own measurements.

    The brief
    Credit · unclaimed

Open 01 · the brief

Rebuild a database by AI

Credit · unclaimed

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

What to build

A pipeline that reads published exposure studies and copies every set of measurement results (agent, job, industry, year, number of samples, average and spread, limit, sampling method) into one standard table with the page and sentence each value came from, tested first against a database that people built by hand.

Why it is still open

Koh and colleagues built a database of US lead exposures in 2015 by reviewing 175 papers holding 1,111 sets of summary results published from 1940 to 2010, and ExpoSYN needed partner institutes to re-enter 356,551 measurements in one format. The AI extraction studies in this field found by the search read safety data sheets, not measurement studies, and a search found no published test of an AI model scored against a hand-built exposure database.

ebi.ac.uk

Built with what exists

Europe PMC's open-access full texts and its REST API; a general AI model that fills fixed fields and quotes the sentence it used (Notes: "The jobs agents repeat"); a second model or a person reviewing flagged rows (Notes: "Seven more flagged"); Python.

  1. First step

    Ask Koh and colleagues for the lead database, and download the open-access papers among the 175.

  2. Done looks like

    On the papers behind the lead database, the pipeline recovers at least 95% of the 1,111 summary sets, with values matching in at least 98% of those recovered, and a public list of every miss and error; then the same pipeline run on a second agent, such as benzene, with 200 rows checked by hand.

Open 02 · the brief

Score AI exposure judgments

Credit · unclaimed

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

What to build

A public test of 200 written exposure scenarios, each taken from a published study: the task, chemical, amounts, controls and duration, with the measured results held back. AI models, and for comparison volunteer hygienists, place each scenario in one of four exposure categories, each a band of its exposure limit. The answer key comes from the study's own measurements.

Why it is still open

Hygienists' judgments have been scored against measurements since at least 2009: about 43% correct with a few samples and about 30% with none, where chance is 25%. A search found no study that scores AI models' exposure judgments against measured data; the AI tests published in this field extract text from data sheets instead.

ebi.ac.ukpmc.ncbi.nlm.nih.gov

Built with what exists

Expostats, free at expostats.ca, to set each scenario's true category from its data; general AI models answering fixed-choice questions (Notes: "A thousand questions, one pass"); open-access exposure studies found through Europe PMC.

  1. First step

    Pick 20 open-access studies that describe the task and give individual results, write one scenario from each, and run Expostats on the results to set the key.

  2. Done looks like

    200 scenarios covering at least 20 agents, each with its source and answer key; the accuracy of at least three AI models, and whether each tends to guess high or low, set beside the published 30% and 43% figures for hygienists; all prompts and answers public.