BRENNEN

Industries / Food, farming and forestry / Flavors and fragrances

Sensory testing

Sensory testing measures how people perceive a flavor or a scent, with trained panels, consumer panels and instruments that stand in for the nose.

7

Milestones achieved, from 1949 to 2015.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1949

    Flavor described by a panel

    Chemists at the Cambridge, Massachusetts consulting firm Arthur D. Little introduced the Flavor Profile method in 1949. A highly trained panel of usually four to six people scores a product's flavor notes, their strength, the order they appear in and the aftertaste, then agrees on one profile by discussion; the Society of Sensory Professionals calls it the "mother" of many later descriptive methods.

    thebhc.orgsensorysociety.org

    Credit

    Arthur D. Little, Inc

  2. 1952

    Liking put on nine points

    Work began in 1947 at the US Quartermaster Food and Container Institute for the Armed Forces on a rating scale to replace slow paired comparisons of foods. The nine-point hedonic scale, four steps of liking and four of disliking around a neutral middle, was introduced in 1952 and quickly became the method of choice for industry, government and universities. A 1955 study found the steps between its words were not equal, but the scale was never revised; its developers wrote that it "was, perhaps, too immediately successful".

    LessonA scale adopted too quickly can lock in flaws that later studies find.

    docs.ufpr.br

    Credit

    Peryam, Girardot, Pilgrim and colleagues at the Quartermaster Food and Container Institute

  3. 1964

    A perfumer sniffs the column

    G. H. Fuller, R. Steltenkamp and G. A. Tisserand described "the gas chromatograph with human sensor: perfumer model" in 1964. In gas chromatography–olfactometry (GC-O), the gas leaving a gas chromatograph, an instrument that separates a mixture into single compounds, is split between a detector and a nose-shaped port, where a trained assessor sniffs and presses a button at each odor, so the compounds that matter to smell can be picked out of a complex mixture.

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

    Credit

    G. H. Fuller, R. Steltenkamp and G. A. Tisserand

  4. 1974

    Panelists score alone

    H. Stone, J. L. Sidel and colleagues published Quantitative Descriptive Analysis (QDA) in 1974, developed by Tragon Corporation with the Department of Food Science at the University of California, Davis, to fix the poor statistics of the Flavor Profile and similar methods. A panel of 10 to 12 agrees on the words to use, then each member scores the products alone in a booth on a line scale, and the scores are analyzed with statistics.

    LessonConsensus scoring had been criticized in the 1960s and 1970s for giving too much room to a panel leader or a headstrong panelist, which scoring alone removes.

    sensorysociety.orgsensorysociety.org

    Credit

    H. Stone, J. L. Sidel and colleagues; Tragon Corporation; UC Davis

  5. 1982

    An electronic nose

    K. Persaud and G. Dodd built an electronic nose from semiconductor gas sensors, each responding broadly to many odors, and reported in Nature in 1982 that it could reproducibly tell apart a wide variety of odors. They built it to test their idea that the nose discriminates by combining signals from many broadly tuned receptor cells, without needing a receptor for each smell.

    ebi.ac.uk

    Credit

    K. Persaud and G. Dodd

  6. 1991

    Smell receptors found

    Richard Axel and Linda Buck, working in mice, published in 1991 a family of about 1,000 genes for odorant receptors. Each smell-sensing cell carries one receptor type, and the brain combines signals from many receptors into a pattern; they shared the 2004 Nobel Prize in Physiology or Medicine for this work.

    nobelprize.org

    Credit

    Richard Axel; Linda Buck

  7. 2015

    A trillion smells unproven

    In 2014 C. Bushdid, M. O. Magnasco, L. B. Vosshall and A. Keller estimated from mixture tests that people can tell apart at least 1 trillion smells, against the usual figure of 10,000. In 2015 R. C. Gerkin and J. B. Castro showed that the same data could give almost any number, that the formula gives an upper bound rather than a lower one, and that there was no evidence for the claim.

    LessonA headline number from a sensory test is only as good as the model used to extrapolate from it.

    ebi.ac.ukebi.ac.uk

    Credit

    C. Bushdid, M. O. Magnasco, L. B. Vosshall and A. Keller (Science); R. C. Gerkin and J. B. Castro (eLife)

Watch for

  1. Human-level smell, claimed

    Osmo, a start-up that Alex Wiltschko spun out of his work at Google Brain in 2023, says its smell model has reached "human-level olfactory intelligence": about as good as an average person at detecting and telling smells apart, and as good as a junior perfume designer at creating them. It says that in October 2026 the model outperformed in-house perfumers at one of the world's largest consumer goods companies. The smell scientist Avery Gilbert has argued that some of its marketing claims are undocumented.

    Achieved whena group independent of Osmo tests the model against a human panel on smells it has not seen and publishes the method and data.

    technologyreview.com

    Claimed by

    Osmo founder Alex Wiltschko, reported by MIT Technology Review, November 2026

Open

  1. 01

    Independent test of AI noses

    A preregistered test in which a panel of about 20 trained volunteers rates 100 aroma chemicals on a fixed list of odor words, using molecules that appear in no public odor dataset. Odor models, such as OpenPOM, any company model offered for testing and general AI models, must lodge their predictions before the panel's ratings are released.

    The brief
    Credit · unclaimed
  2. 02

    A robot that dips strips

    A bench robot arm that dips paper smelling strips into raw materials or finished perfumes to a set depth, prints a blind code on each strip, and presents the strips to evaluators at set times after dipping, so that the same scent can be judged fresh, after an hour and after a day.

    The brief
    Credit · unclaimed

Open 01 · the brief

Independent test of AI noses

Credit · unclaimed

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

What to build

A preregistered test in which a panel of about 20 trained volunteers rates 100 aroma chemicals on a fixed list of odor words, using molecules that appear in no public odor dataset. Odor models, such as OpenPOM, any company model offered for testing and general AI models, must lodge their predictions before the panel's ratings are released.

Why it is still open

Osmo says its model has reached "human-level olfactory intelligence", and the smell scientist Avery Gilbert says some of its marketing claims are undocumented. Searches of Europe PMC in November 2026 found new odor models compared on existing public datasets, and no test on a fresh panel run by a group independent of a model's makers.

technologyreview.com

Built with what exists

OpenPOM, an open, MIT-licensed reproduction of the 2023 Principal Odor Map model (github.com/ARY2260/openpom); Pyrfume's open odor datasets, to check that each molecule is new (github.com/pyrfume/pyrfume-data); RDKit for chemical structures; the Open Science Framework for the preregistration; paper smelling strips; aroma chemicals bought from perfumery suppliers, limited to materials allowed in fine fragrance under the IFRA Standards and diluted to 1% in dipropylene glycol.

  1. First step

    Draft the odor word list and the protocol, pick 150 candidate molecules sold by perfumery suppliers, drop any that appear in Pyrfume, and post the preregistration.

  2. Done looks like

    Every model's predictions time-stamped before the panel data opens; each panelist rates 20 of the molecules twice, which gives the panel's own repeat agreement as the ceiling; each model's correlation with the panel average reported against that ceiling; ratings, code and the molecule list published.

Open 02 · the brief

A robot that dips strips

Credit · unclaimed

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

What to build

A bench robot arm that dips paper smelling strips into raw materials or finished perfumes to a set depth, prints a blind code on each strip, and presents the strips to evaluators at set times after dipping, so that the same scent can be judged fresh, after an hour and after a day.

Why it is still open

The strip machines found in patents are made for shops. One, filed in 2016 by the Gaudier Kuppel Group, stores strips, sprays perfume onto each and marks it so it can be identified later; its background describes a person spraying perfume from a small bottle onto a strip by hand. Searches of patents, GitHub and Europe PMC in November 2026 found no robot that dips strips for blind evaluation in a laboratory.

patents.google.com

Built with what exists

An SO-101 robot arm run with Hugging Face's open LeRobot software (github.com/huggingface/lerobot; see the Note "No download"); a label printer and barcode reader; an analytical balance reading to 0.1 mg to weigh the liquid each strip picks up; a fume extractor.

  1. First step

    Weigh 50 strips dipped by each of two people into the same perfume, to measure the spread the robot has to beat.

  2. Done looks like

    200 robot dips whose liquid load per strip, measured by weighing, varies less than the hand dips of two perfumers in a side-by-side test; no code mix-ups across 200 strips, checked by barcode; every strip presented within 30 seconds of its scheduled time.