Industries / Consumer goods / Cosmetics and personal care
Mixing and filling
Mixing and filling is the factory work that turns a recipe into thousands of identical units: boiling soap, blending creams, molding lipsticks and filling bottles and cans, while keeping germs and other contaminants out.
4
Milestones achieved, from 1880s to 2021.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1880s
Soap moves into big works
After a fire at its first plant, Procter & Gamble built Ivorydale in Cincinnati in 1886 as a purpose-built factory, designed by the industrial architect Solon Beman, and later added its first research lab there. In England, William Hesketh Lever cut the first sod of his Port Sunlight soap works in March 1888, and the first boil of soap there took place in June 1889, according to the Port Sunlight Village Trust, an independent charity. Lever Brothers later became Unilever.
Credit
Procter & Gamble; William Hesketh Lever, Lever Brothers
2007
One rulebook for plants
ISO 22716:2007 sets out good manufacturing practice (GMP) for cosmetics: the rules for premises, equipment, staff, hygiene and records that keep each batch the same and free of contamination. In April 2011 the European Commission published it as the harmonized standard under the EU's cosmetics regulation, which gives makers who follow it a presumption that they meet the law's GMP requirement. The FDA's 2013 draft guidance on cosmetic GMP says the FDA reviewed the same standard and adopted, changed or left out parts of it.
Credit
International Organization for Standardization; European Committee for Standardization (CEN); European Commission; FDA
2019
Germs in recalled cosmetics counted
Scientists at Germany's Federal Institute for Risk Assessment (BfR) went through 13 years of alerts in the EU's rapid alert system for dangerous non-food products, then called RAPEX and now Safety Gate. Of 243 alerts for germs from 2005 to 2017, 111 concerned cosmetics. Pseudomonas bacteria, especially Pseudomonas aeruginosa, were the most common germs in cosmetics and toys, and two groups of bacteria, Pluralibacter gergoviae and Klebsiella, turned up only in cosmetics. The median time from a lab finding to an EU alert was eight weeks.
LessonGerm alerts reach the EU system weeks after the lab result, so contamination has to be caught at the plant rather than on the shelf.
Credit
Szilvia Vincze, Sascha Al Dahouk and Ralf Dieckmann, BfR
2021
Benzene traced to the propellant
In 2021 Valisure, an independent testing lab in New Haven, Connecticut, reported benzene, a chemical that causes cancer, first in aerosol sunscreens and then in body sprays, where it found benzene in 54% of 108 batches from 30 brands, and petitioned the FDA to act. In December 2021 Procter & Gamble recalled aerosol dry shampoos and dry conditioners, saying that "unexpected levels of benzene came from the propellant that sprays the product out of the can". In October 2022 Unilever recalled aerosol dry shampoos sold as Dove, Nexxus, Suave, TIGI and TRESemmé after "an internal investigation identified the propellant as the source". The FDA names isobutane, a propellant gas, among the possible sources of benzene.
LessonA harmful chemical can come in with an ordinary raw material, here the gas that pushes the product out of the can, so every supplier's material needs testing, not just the recipe.
Credit
Valisure (chief executive David Light); Procter & Gamble; Unilever; FDA
Watch for
US factory rules, still pending
The Modernization of Cosmetics Regulation Act of 2022 requires the FDA to propose good manufacturing practice rules for cosmetics plants by 29 December 2024 and to publish final rules by 29 December 2025, consistent "to the extent practicable" with national and international standards. Neither has appeared. A June 2026 update from EcoMundo, a regulatory consultancy, reports that the rule has moved to the FDA's list of long-term actions and that no proposal is expected within 12 months.
Achieved whenthe FDA publishes a final GMP rule for cosmetics facilities.
Claimed by
US Congress, December 2022; status reported by EcoMundo, June 2026
AI help at the line
Estée Lauder says Ella, its "Estée Lauder Line Assistant", is running in two of its factories. It helps operators deal with equipment problems and speeds up changeovers, the switch of a production line from one product to the next. The company has published no figures.
Achieved whenEstée Lauder or an outside party publishes changeover times and downtime for the same lines before and after Ella.
Claimed by
Florent Guillet-Caillot, vice president of supply chain at Estée Lauder, in BeautyMatter, May 2026
Open
01
Lipstick flaws in photos
A public set of photographs of lipstick bullets, the molded stick, with labeled flaws: air holes, cracks, dents, broken tips, uneven color, beads of oil on the surface (called sweating) and specks of dirt. Make the sticks by remelting cheap shop-bought lipsticks in a hobby lipstick mold and planting each flaw on purpose, photograph them under fixed light from several angles, and publish a baseline flaw detector scored on shades and shapes it has not seen. The sticks are for photographs only and are not worn.
The briefCredit · unclaimed02
Where the germs came from
A public table that gives, for each cosmetics recall for germs, the cause that the company or a regulator stated: a water system, a contaminated raw material, a preservative too weak for the formula, poor cleaning of equipment, or none stated. Start from the recalls already compiled, and look up each firm's recall notice, any FDA warning letter and, where released, the FDA inspection report (Form 483, the list of problems an inspector hands over at the end of a visit), plus the national notices behind EU alerts.
The briefCredit · unclaimed
Open 01 · the brief
Lipstick flaws in photos
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of photographs of lipstick bullets, the molded stick, with labeled flaws: air holes, cracks, dents, broken tips, uneven color, beads of oil on the surface (called sweating) and specks of dirt. Make the sticks by remelting cheap shop-bought lipsticks in a hobby lipstick mold and planting each flaw on purpose, photograph them under fixed light from several angles, and publish a baseline flaw detector scored on shades and shapes it has not seen. The sticks are for photographs only and are not worn.
Why it is still open
Inspection makers sell systems for this job without publishing their images or error rates. SEA Vision Group and ARGO Vision say their lipstick system learns from "proprietary datasets", a mix of real and synthetic images, and a July 2026 description of EyeVision's lipstick inspection lists cracks, dents, contamination and color variation with no accuracy figures. A 2022 conference paper by Zhu and colleagues reports a lipstick surface-defect system tested in a case study, and its abstract mentions no public image set. A public lipstick set on Roboflow Universe has 92 images labeled for where the lipstick is, not for flaws, and searches of the web and of GitHub for a lipstick defect dataset found none.
process-worldwide.commvpromedia.comapi.crossref.orguniverse.roboflow.com
Built with what exists
A silicone or aluminum hobby lipstick mold and a hot plate; a phone or USB camera on a stand inside a light box; Label Studio, an open-source labeling tool; Anomalib, Intel's open-source library for anomaly detection, which learns from good parts only; Ultralytics YOLO to draw boxes around flaws.
First step
Remelt ten lipsticks of different shades, mold 50 sticks with and without planted air holes and cracks, and photograph them to settle the lighting and camera angles.
Done looks like
At least 2,000 labeled images of at least 300 sticks in 20 shades, with at least 100 examples of each flaw type, released under an open license, plus a baseline whose share of flawed sticks caught and false-alarm rate on held-out shades are published with the code.
Open 02 · the brief
Where the germs came from
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table that gives, for each cosmetics recall for germs, the cause that the company or a regulator stated: a water system, a contaminated raw material, a preservative too weak for the formula, poor cleaning of equipment, or none stated. Start from the recalls already compiled, and look up each firm's recall notice, any FDA warning letter and, where released, the FDA inspection report (Form 483, the list of problems an inspector hands over at the end of a visit), plus the national notices behind EU alerts.
Why it is still open
The recalls have been counted, but not traced. Venkatesh and colleagues (Journal of the American Academy of Dermatology, 2025) counted 334 US cosmetics recalls from 2011 to 2023, 76.8% of them for germs, mostly Pseudomonas and Burkholderia bacteria, and listed incomplete information on contaminants among their limits. NSPOMdb (Scientific Data, December 2025) compiles 371 cosmetics recalls for germs from 2006 to 2024, 207 from the EU's Safety Gate and 164 from the FDA, with the germ named; in its recall table, which this session downloaded, only two of the 371 recall reasons mention a water system or a preservative. A search for a study that sorts cosmetics germ recalls by cause found none.
Built with what exists
NSPOMdb's recall table (CC BY 4.0, on figshare); the Safety Gate weekly-report XML download; FDA recall notices, warning letters and the FDA Data Dashboard's inspection citations; a small AI model to pull stated causes out of long documents, such as Meta's Muse Ember (see the Note "The jobs agents repeat"); a spreadsheet in which two people code causes separately.
First step
Download NSPOMdb's recall table, take the 164 FDA cosmetics rows, and search the FDA's warning-letter page for each recalling firm to see how often a cause can be found at all.
Done looks like
At least 150 cosmetics germ recalls, each coded with a stated cause or "none stated" and linked to the document it came from; two coders agree on at least 85% of codes before settling differences; the share of recalls with each cause published with the table under an open license.