Researchers from Google and Harvard have teamed up to create an AI program that can help spot potentially dangerous restaurants. The AI program uses anonymized and aggregated data from users that opt to share their locations and other data with Google. Essentially, it identifies users who may be suffering from or have had foodborne illnesses, such as people searching for things like how to relieve stomach pain or people who may have wound up at their local emergency room, then traces their location history back to see what restaurants they’ve been to. The model aggregates the data across multiple users in order to establish a rough estimate of just how likely a random guest at each tracked restaurant would be to get sick after eating there. The model is called FINDER, short for Foodborne Illness DEtector in Real time. The model’s primary focus is to detect possible foodborne illness sources and outbreaks faster and more accurately than traditional methods, which include things like consumer complaints to relevant organizations and regularly scheduled health inspections for restaurants.
Background: The way FINDER works is pretty simple, in theory. As stated above, it looks at aggregated user data to figure out how likely restaurants are to cause foodborne illnesses. The machine learning part comes in with the inevitable flaw; what if a certain restaurant gets blamed for a foodborne illness spreading, when that’s not the case? This is why the machine is always learning from the results of its actions, including the results of inspections that it sparks. A pilot in Chicago and Las Vegas managed to get its success rate up above 50% when all was said and done, compared to traditional methods for the two cities averaging in the 20% range. It’s worth mentioning, on that note, that Chicago has one of the United States’ most stringent, advanced, and successful restaurant health inspection systems.
Impact: As this AI becomes better at its assigned function and rolls out to more areas, it will inevitably help combat the spread of foodborne illness in restaurant settings. That’s a given. The real impact, however, comes in the core setup of the AI and how it can be applied in other ways. Cross-referencing user data with other factors to obtain more cohesive data is a fine method, and it can easily be applied in things like business ratings, bargain hunting, rating the injury risk of activities, finding places off the beaten path and other consumer-centric use cases. On a grander scale, an AI set up similarly to this one could be used to get more accurate demographic information for regions, activities, and business, for example. It could also be used on a countrywide or even worldwide scale to hunt for indicative trends in all sorts of different use cases like product popularity, the success of and reaction to new laws, and more. Couple with further data sources, such as surveys, it could also be used to produce grander and more accurate demographic data that could help in decisions like what region to release a new movie in first, or what world population group should be the first to get to try a new type of medicine.