Emerging coronavirus hotspots likely already exist undiscovered in the eight U.S. states planning to ease social distancing restrictions, a new analysis reveals.

The study discovers that in the eight states planning to ease social distancing measures in the coming weeks, almost two dozen small cities and rural counties have infection rates poised to skyrocket, despite their currently not having large numbers of confirmed cases.

The analysis, by technology firm Dataminr, uses artificial intelligence to scrutinize social media posts on platforms such as Twitter and Reddit. The company also looks at 10,000 public data sources in more than 100 languages. Its technology is used by both governments and companies, including many hedge funds, to predict emerging events. It is also used by the World Health Organization and the United Nations.

The company’s warning about previously unknown hotspots comes on the day that Georgia became the first U.S. state to ease measures imposed to stop the spread of outbreaks.

The state has allowed gyms, tattoo parlors, hair and nail salons, massage therapists, and other businesses to reopen after less than a month of closures. It is allowing in-person religious services to restart this weekend and will allow restaurants and theaters to reopen starting on Monday.

Yet Dataminr forecasts that two Georgia counties—Chatham and Clarke—are among those that, even with social distancing in place, are already one to two weeks away from experiencing an exponential surge in COVID-19 cases.

Ted Bailey, Dataminr’s chief executive, said that his company’s study indicated that infection rates varied widely within the states that were now reopening for business: “While it may be the case that the outbreak is ‘peaking’ in major cities, it is also likely the case that new virus waves are on the brink of emerging in other small metro and rural areas.” Though not an epidemiologist, Bailey warned that easing social distancing restrictions in the midst of this kind of “rolling” outbreak “could be detrimental to recovery.”

Georgia’s experience is being closely watched by other states as they consider how to emerge from strict shelter-in-place orders, business and school closures, and social distancing measures. States that have already announced plans to ease most restrictions in the coming week include Montana and Oklahoma, while South Carolina and Tennessee have announced that only some restrictions will be lifted during this time. Several other states have said they will reduce social distancing measures by early May.

Among the areas where Dataminr has identified new emerging hotspots are counties in Indiana and Michigan. Indiana’s lockdown is due to expire on May 1, and the state is currently debating whether to extend it. Michigan announced today that its restrictions will remain in place until May 15 but will be eased for some recreational activities, including boating and golf, as well as for some businesses like garden centers, plant nurseries, and landscaping companies.

Earlier in the course of the coronavirus pandemic, Dataminr’s software was able to accurately forecast infection hotspots in 14 states between seven and 15 days prior to the actual case totals in those locations beginning to grow exponentially, according to a study the company published in March.

The company was also among the earliest to warn the world of a possible emerging epidemic, issuing an alert based on a surge in social media posts related to a potential novel disease on Dec. 30. It is one of a handful of companies, including BlueDot and HealthMap, that have used A.I. to try to spot emerging epidemics and track their spread.

Bailey said that previous experience during this pandemic has shown his company’s forecasts to be “highly accurate” in identifying regions about to experience rapid growth in infection rates, starting with Wuhan, China, and continuing with accurate predictions for places like Italy, South Korea, Spain, and Iran.

Dataminr uses natural language processing—a kind of machine learning that can analyze language—to look for geographic clusters of social media posts related to the coronavirus. But rather than searching for key words or monitoring aggregate volumes of social media posts, Dataminr instead looks for what it calls “eyewitness” accounts.

In this case, Dataminr’s software searches for posts in which people indicate that they have tested positive for COVID-19; are experiencing symptoms consistent with the disease; think they’ve been exposed to the virus but have not yet been tested; or provide firsthand accounts stating that relatives, friends, and colleagues are sick. It also looks for mentions of COVID-19-related supply shortages and closures.

This data is then fed into another algorithm that tries to correlate the volume and acceleration of such posts with the likelihood that a particular location will experience an exponential increase in infection rates in the next two weeks.

In addition to the two counties in Georgia, Dataminr is forecasting likely surges in the following counties: