Industries / Retail, hospitality and personal services / Hospitality and tourism / Restaurants and food service
Food safety
This work keeps food made and served in restaurants and canteens safe to eat: cooking and cooling it to the right temperatures, keeping sick workers away from food, telling customers what is in it, and the inspections that check all of this.
7
Milestones achieved, from 1993 to 2021.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1993
Burgers cooked through
E. coli O157:H7 in undercooked hamburgers from one restaurant chain made more than 500 people ill in Washington, Idaho, California and Nevada between November 1992 and February 1993, and four died, by the CDC's count. Nation's Restaurant News later reported that the chain, Jack in the Box, cooked its patties to an internal 140°F (60°C) under an older standard, though a Washington State rule a few months old required 155°F (68°C). After the outbreak the FDA's model food code raised the minimum for ground beef from 140°F to 155°F. The chain hired David Theno, who put HACCP, a system of checks at each step where a hazard can be controlled, into its restaurants and held managers to account for food safety.
LessonA cooking rule protects diners only when every kitchen knows it and checks it with a thermometer.
Credit
Washington State; the US Food and Drug Administration; Jack in the Box and David Theno
1998
Grades in the window
From January 1998 Los Angeles County made restaurants post their inspection grade as a card customers could see. A 2005 study by Paul Simon and colleagues, among them the economists Ginger Zhe Jin and Phillip Leslie, compared the county with the rest of California and found foodborne-disease hospitalizations 13.1% lower in 1998, and still lower in 1999 and 2000. New York City began letter grades in 2010. Daniel Ho of Stanford Law School studied more than 700,000 inspections in ten places and found that nearly every restaurant in San Diego got an A, that New York's scores were inconsistent and did not predict later cleanliness, and that New York's grades had not discernibly reduced foodborne illness.
LessonA posted grade helps only if inspectors score alike, and a scale where nearly everyone gets an A tells diners nothing.
Credit
Los Angeles County; the New York City Department of Health and Mental Hygiene; studies by Simon and colleagues (2005) and Ho (2012)
2012
Reviews flag food poisoning
From July 2012 New York City's health department, with Columbia University and Yelp, ran software that screens Yelp restaurant reviews for signs of food poisoning. In the first nine months it scanned about 294,000 reviews and passed 893 to an epidemiologist; interviews with reviewers turned up three outbreaks nobody had reported, with 16 people ill (CDC's Morbidity and Mortality Weekly Report, 2014). By 2018 the department's team reported that the system had helped find 10 outbreaks and 8,523 complaints of foodborne illness linked to city restaurants.
Credit
New York City Department of Health and Mental Hygiene; Columbia University; Yelp
2013
Ratings posted by law
The Food Hygiene Rating (Wales) Act 2013 requires food businesses in Wales to display the hygiene rating sticker that inspectors give them. Northern Ireland made display compulsory in 2016, while in England it stayed voluntary. In the 2024 audit for the UK Food Standards Agency, carried out by IFF Research through 1,348 unannounced visits and published in June 2025, the sticker could be seen at 94% of premises in Wales, 90% in Northern Ireland and 72% in England.
Credit
The National Assembly for Wales; the Food Standards Agency
2015
A model picks inspections
In 2015 the Chicago Department of Public Health began using a machine learning model, a program trained on past results, to decide which of its more than 15,000 food establishments to inspect first, and the city published the code and data. In the city's own simulation the model found restaurants with critical violations seven days earlier on average. In the first independent analysis, in 2019, Vinesh Kannan, Matthew Shapiro and Mustafa Bilgic of the Illinois Institute of Technology found that the model's strongest input was which group of inspectors had last visited, that inspectors differed in strictness, and that the rise in violations found after 2015 might not be due to the model. They recommended dropping the inspector input.
LessonA model trained on inspection results learns the inspectors' habits along with the kitchens' risks, and publishing the code is what let outsiders find this.
Credit
Chicago Department of Public Health; Kannan, Shapiro and Bilgic
2020
Sick workers, record fine
Outbreaks at Chipotle Mexican Grill restaurants from 2015 to 2018 made more than 1,100 people ill. In Boston in December 2015 norovirus, a stomach virus that spreads easily from infected workers to food, sickened 141 people, probably through an apprentice manager who kept working after vomiting in the restaurant. In Powell, Ohio, in July 2018 about 647 people fell ill from Clostridium perfringens, a germ that grows when cooked food is held at the wrong temperature. In April 2020 Chipotle agreed to a $25 million criminal fine, which the US Department of Justice called the largest ever in a food safety case, and a deferred prosecution agreement requiring a better food safety program. Court filings said some employees reported pressure to work while sick and too little staffing and training.
LessonA written rule to send sick workers home fails when staff feel they cannot afford to stay away. CDC's study of 800 restaurant and other retail food outbreaks from 2017 to 2019 found ill or infectious workers among the causes of 41% of those whose contributing factors were identified, and only 43.6% of managers said any workers had paid sick leave.
Credit
US Department of Justice; Chipotle
2021
Labels for shop-packed food
Natasha Ednan-Laperouse died on 17 July 2016 of anaphylaxis, a severe allergic reaction, after eating a Pret a Manger baguette that contained sesame. The coroner, Sean Cummings, found that allergens "were not labelled adequately or clearly" on the packaging. From 1 October 2021 England's Food Information (Amendment) Regulations 2019, known as Natasha's Law, require food prepacked for direct sale, such as sandwiches a shop packs itself before customers choose them, to carry its name and a full list of ingredients with allergens labelled. The Food Standards Agency's guidance on the rule covers England, Wales and Northern Ireland.
LessonThe EU rules England inherited left food prepacked for direct sale out of the labelling required for other packed food, and the gap closed only after a death.
Credit
The Ednan-Laperouse family; the UK government
Watch for
AI checks delivery kitchens
The delivery platform Taobao Shangou says its Bright Kitchen 2.0 program, started in July 2025 with China Telecom and DingTalk, runs more than 10 billion AI checks a year of restaurant back kitchens seen through cameras, and that 90% of merchants warned about a problem did not repeat it within three months. The figures come from Huanqiu's report on proposals that the platform's president, Fang Yongxin, made to the Shanghai People's Congress.
Achieved whenA regulator or an independent researcher publishes how accurately the cameras flag real violations, or compares inspection results in camera-watched kitchens with others.
Claimed by
Taobao Shangou, February 2026
England to require ratings
The Food Standards Agency's report to its September 2026 board lists "mandatory Food Hygiene Rating Display online across the UK and in-store in England" among planned changes. Detailed design runs to early 2027, with a consultation planned for June 2027 and advice to ministers in January 2028.
Achieved whenA law requiring food businesses in England to display their rating is in force.
Claimed by
Food Standards Agency, September 2026
Open
01
Cooling curves for big batches
A public set of full cooling curves for foods that kitchens cook in bulk and chill for later, such as rice, chili, stock and braised meat, and a calculator that predicts from the food, batch size, container depth and cooling method whether a batch will cool from 57°C to 21°C (135°F to 70°F) within 2 hours and to 5°C (41°F) within 6 hours in all, as the FDA Food Code requires. Cook each recipe in a test kitchen, cool it by the methods restaurants use (walk-in refrigerator, ice bath, ice wand, shallow pans, blast chiller), log the temperature at its centre every minute, and then throw the food away.
The briefCredit · unclaimed02
Check a sick-worker policy
A checker that reads a restaurant's written policy on sick employees and lists what it leaves out of the FDA Food Code's employee health rules: the symptoms workers must report, such as vomiting and diarrhea, the illnesses they must report, when a worker must be kept away from food or sent home, and when they may come back. Test it on a public set of at least 200 policies, made up of templates published by health departments and policies written for the test with planned gaps, each with an answer key.
The briefCredit · unclaimed
Open 01 · the brief
Cooling curves for big batches
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of full cooling curves for foods that kitchens cook in bulk and chill for later, such as rice, chili, stock and braised meat, and a calculator that predicts from the food, batch size, container depth and cooling method whether a batch will cool from 57°C to 21°C (135°F to 70°F) within 2 hours and to 5°C (41°F) within 6 hours in all, as the FDA Food Code requires. Cook each recipe in a test kitchen, cool it by the methods restaurants use (walk-in refrigerator, ice bath, ice wand, shallow pans, blast chiller), log the temperature at its centre every minute, and then throw the food away.
Why it is still open
CDC's Environmental Health Specialists Network (EHS-Net), a group of health department staff, measured cooling food in about 50 restaurants in 2009 and 2010 only at the start and end of each visit, and estimated cooling rates from those two points. By that estimate 65% of cooling observations met the 6-hour guideline, and the authors wrote that future work could validate the method and study "the full cooling process". In an earlier EHS-Net survey of 420 restaurants, 41% of managers said they did not monitor time or temperature during cooling. Kevin Roberts and colleagues at Kansas State University logged full curves at one-minute intervals in 2013, but for two school recipes only, chili and tomato sauce, in pans 5.1 cm and 7.6 cm deep cooled in a walk-in refrigerator, a walk-in freezer or an ice bath, and in a stockpot with a chill stick; only the 5.1 cm pans in the freezer met both Food Code limits, and the curves appear as charts in the paper. This research found no open dataset of full cooling curves across many foods.
pubmed.ncbi.nlm.nih.govpubmed.ncbi.nlm.nih.govfoodprotection.org
Built with what exists
Waterproof DS18B20 or type K thermocouple probes read by an ESP32 or Raspberry Pi Pico microcontroller, or commercial data loggers; a reference thermometer checked in ice water; Python with SciPy to fit the curves; Zenodo to publish the data.
First step
Build a four-probe logger, cook a 6-litre batch of rice, and log it cooling in a deep pot and in two shallow pans side by side in the same refrigerator.
Done looks like
At least 300 logged cooling runs covering 8 foods, 3 batch sizes and 5 methods, published with recipes, container sizes and room and refrigerator temperatures, and a calculator whose predicted time to 5°C is within 30 minutes of the measured time on a held-back 20% of runs.
Open 02 · the brief
Check a sick-worker policy
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A checker that reads a restaurant's written policy on sick employees and lists what it leaves out of the FDA Food Code's employee health rules: the symptoms workers must report, such as vomiting and diarrhea, the illnesses they must report, when a worker must be kept away from food or sent home, and when they may come back. Test it on a public set of at least 200 policies, made up of templates published by health departments and policies written for the test with planned gaps, each with an answer key.
Why it is still open
In CDC's study of 800 outbreaks at restaurants and other retail food places from 2017 to 2019, only 16.1% of the establishments had policies covering all four Food Code components assessed, and 23.0% of policies on telling a manager listed all five symptoms. The FDA's Employee Health Policy Tool, based on Part 2-201 of the 2022 Food Code Supplement, walks a business through writing a new policy but does not review one it already has. This research found no tool that does.
Built with what exists
The text of Part 2-201 of the FDA Food Code; health department templates; AI models that read documents and photos, such as Claude Fable 5.2 (see the Note "Three working devices") and the open-weight Mistral Large 4 (see the Note "A trillion open parameters"), given the same instructions; Python for scoring.
First step
Collect 20 public templates from state and local health departments, write 30 policies with planned omissions, and build their answer key from section 2-201.
Done looks like
200 policies with answer keys agreed by two people who know the Food Code, and a published table of each model's accuracy for each component, with a separate count of unsafe errors, where a model calls a policy complete when it is not.