Protecting families from contaminated food and water
All fields are required
Protecting families from contaminated food and water
Posted in Our Blog on May 23, 2025
Last year, Make Food Safe considered the possibility of using AI to detect foodborne illness using restaurant reviews. But new technological discoveries often age like milk. In just a year, so much has changed.
A new study from the UK Health Security Agency revisited that concept. Even finding that the limitations observed in the aforementioned post have been resolved using new parameters.
Could we see a future where health and safety investigators mine data with AI to detect foodborne illness that would be potentially missed with traditional reporting methods?
Possibly!
Let’s discuss their findings.
A recent study, titled “Restaurant Review Gastrointestinal Illness Detection and Extraction with Large Language Models” assessed different types of Artificial Intelligence (AI) for their ability to detect and classify text using online restaurant reviews. A task that could potentially identify and target investigations of foodborne illness outbreaks.
“We are constantly looking for new and effective ways to enhance our disease surveillance,” said Professor Steven Riley, Chief Data Officer at UKHSA. “Using AI in this way could soon help us identify the likely source of more foodborne illness outbreaks, in combination with traditional epidemiological methods, to prevent more people from becoming sick.”
Technology to detect foodborne illness is advancing
He notes that further work is needed before using this technology in a routine investigative approach.
But the technology and learning how to harness AI to detect foodborne illness is advancing.
Foodborne gastrointestinal illness is a significant cause of sickness in the United Kingdom (UK), The United States, and the rest of the world. According to the UKHSA, there is an estimated 2.4 million cases of foodborne illness in the UK each year. For the United States, that number is even higher – 48 million!
The UKHSA expects that the true number of people afflicted by foodborne illness is significantly higher.
As much as 90% higher!
It is estimated that only 10% of people experiencing gastrointestinal illness seek medical attention
Only those who seek medical attention are tested for pathogens responsible for foodborne illness. Most cases of foodborne illnesses are mild enough to resolve on their own.
So, the sick individual never reaches out for medical care. And therefore, their sample is never tested. Leaving a significant number of missing diagnoses unreported!
Restaurant review sites are all over the internet.
You can leave a Google Review. There is Yelp, Trip Advisor, Open Table, and even restaurant social media accounts. All free to access, free to review, and free to data mine.
Some people are serial reviewers. Those kind people praise great service and great experiences when they go out.
But most people only use that service when things go wrong.
Exceptionally bad experiences. Or… When they get sick.
This is where scientists believe that they can use AI to detect foodborne illness outbreaks.
If several people leave reviews indicating they became sick after eating at the restaurant around about the same time, it could indicate a foodborne illness outbreak.
But how do these scientists plan to use AI to detect foodborne illness outbreaks?
UKHSA scientists build on some of their previous language models research used to identify key words. Chat GPT-4, GPT-4o, and Llama-3.3 were used in these investigations.
A comprehensive list of keywords were included in the search parameters.
These included:
Certain symptoms were intentionally omitted. Mostly because they are not “sufficiently specific to GI illness.” Symptoms like headache, fever, and respiratory symptoms.
AI language models were trained with annotated data from around ten thousand restaurant reviews. For this data input, epidemiologists first reviewed the data and provided the correct answer. A process commonly used to train AI.
Then they put the technology to work.
Over three thousand reviews were both manually annotated and then mined through AI to detect foodborne illness in the dataset.
Again. Results were promising.
Results for GI illness detection, notation of symptoms, and food information were identified.
Llama-3.3 was the best performing of the AI models.
Some of the prior concerns for limitations have been overcome in the past year.
Concerns such as differences in spelling or misspelling, gender biased terms, and even marital status details seem to be a non-issue for current AI technology.
There are, however, persistent limitations investigators worry about. Aspects of the process that still need improvement.
Only “catered food” would appear in a restaurant review. A gathering, such as a wedding, family barbeque, or home cooked food would not be available for data analysis and therefore otherwise go undetected.
This limitation also potentially leads to income bias.
The study indicates the concern that “people with higher incomes are more likely to consume catered-food, though it is unclear whether certain types of diners leave reviews more regularly.”
With various incubation periods, the afflicted may not appropriately attribute illness to the appropriate source.
Depending on the germ involved, you may begin feeling sick within a few hours or up to several days after eating contaminated food. The true source of the illness may not be accurately reported.
The study also indicates that most of the time people attribute foodborne illness to catered food. Though evidence shows that non-catered food can also lead to illness. Or even some other source entirely.
According to the study, the weakest part of the process is the “approach to food disambiguation.” AI is taught to focus on named ingredients, rather than likely ingredients found in certain dishes.
In many cases, scientists have found that even with a “long look up table” used in an attempt to overcome this limitation, AI are still sometimes unable to match an ingredient to one of the labels it is looking for.
While the results of current research indicate that AI can be a useful tool to help detect potentially unreported foodborne illness outbreaks, it is far from a perfect system.
AI will not detect every foodborne illness outbreak through publicly available restaurant reviews. Additionally, results should be approached with caution when AI does indicate potential outbreaks.
If you’d like to know more about food safety topics in the news, like “UK Scientists Improve AI to Detect Foodborne Illness using Restaurant Reviews,” check out the Make Food Safe Blog. We regularly update trending topics, foodborne infections in the news, recalls, and more! Stay tuned for quality information to help keep your family safe, while The Lange Law Firm, PLLC strives to Make Food Safe!
By: Heather Van Tassell (contributing writer, non-lawyer)