How Reuters collected and analyzed prison temperature data

Reuters

30-07-2025

During the process of reporting on a story highlighting the impacts of rising temperatures on the U.S. prison population, Reuters set out to answer: How many prisons provide cooling for inmates? How hot does it get inside prisons?
Reuters filed public records requests to every state and federal corrections department in the country asking for data on the presence of air conditioning systems in housing units and copies of indoor temperature logs.
The absence of air conditioning in prisons can create dangerously high temperatures, potentially causing severe health issues and, in extreme cases, death.
In February 2024, Reuters submitted public records requests, asking for a spreadsheet or electronic record indicating the presence of air conditioning in housing units in each adult prison. Thirty-five states responded, with 29 states providing records and six states either denying our request or responding they did not possess or maintain records.
The Bureau of Prisons, which oversees all 122 federal prisons, did not respond to Reuters' request seeking information on how many facilities have air conditioning.
In December 2024, Reuters submitted an additional round of public records requests, seeking spreadsheets or electronic records on the daily internal temperature readings for each housing area in every adult prison, covering the period from December 1, 2023, to December 1, 2024. Based on interviews with incarcerated individuals, the presence of air conditioning alone did not guarantee relief and adequately comfortable temperatures inside. Reuters identified 16 states across the country that were either in heat-prone areas or provided limited air conditioning. The federal Bureau of Prisons was not included in this request.
Responses to these requests came in the form of emails, spreadsheets and PDFs.
Reuters created a single standardized spreadsheet with the responses on the presence of air conditioning in housing units.
Each prison was assigned a unique id (facility_id in the data), which matches the Department of Homeland Security's (DHS) Prison Boundaries Database. The DHS dataset includes the geographic boundaries and attributes of prisons, detention centers, re-entry facilities, jails and juvenile detention centers.
If a facility_id is marked as NA in our spreadsheet, it may be because the facility is new and opened since the time of the last update by the DHS. In other instances, a facility may officially be categorized in another way, making it unavailable in the DHS database.
To categorize air conditioning status, Reuters created two additional fields: ac_status and ac_status_details.
For ac_status, a prison was marked as:
The field ac_status_details provides additional information:
Reuters found that nearly 50% of state prisons across 29 states have partial or no air conditioning in housing units. The Texas Department of Criminal Justice reported that 51 of its 58 facilities fall into this category—the highest number among all states that responded to Reuters. None of Maryland's 14 state prisons have air conditioning.
The presence of air conditioning in a facility does not necessarily mean the system is operational. Some states provided additional details indicating whether air conditioning units required maintenance. In cases where a state reported that air conditioning is available in housing units, Reuters assumed those systems are functioning and capable of providing cool air to incarcerated individuals.
In addition, it is possible that a prison has implemented air conditioning since we received the records.
Reuters received copies of indoor temperature logs from five states. Each state provided documents that were in their respective recording formats.
For instance, the Texas Department of Criminal Justice provided daily temperature data taken at 3 p.m. in housing areas without air conditioning, while data for areas with air conditioning was collected at 1 p.m. every few days.
The California Department of Corrections and Rehabilitation provided temperature records from May through September 2024 in the form of scanned PDFs containing handwritten entries. In total, 189 PDFs comprising over 28,000 pages were given to Reuters. Of those, Reuters identified 20,326 pages as directly related to internal temperatures.
We focused on processing information from California as some CDCR facilities were in the nation's top 10 most heat-exposed prisons, where average outdoor daily temperatures exceeded 85 degrees Fahrenheit (29 degrees Celsius) for multiple days in a row. California state prisons were also where most of the incarcerated individuals we interviewed were located.
To expedite processing this large volume of data, Reuters used Gemini 2.5 Pro, Google's most advanced AI thinking model, to read and extract the temperature log information, outputting the results in JSON format.
Each page of the PDF was uploaded to Gemini 2.5 Pro along with a set of detailed instructions for data extraction. The AI-extracted data was then manually cleaned and merged into a single dataset. We further filtered this dataset to include internally recorded temperatures from June through August 2024, which can be found here.
From June through August 2024, the average indoor temperature across all California state prisons was 76.2 F (24.6 C). The Sierra Conservation Center, a minimum and medium-security prison facility, recorded the highest average at 84 F (29 C). Several facilities experienced dangerously high indoor temperatures, with some reaching up to 104 F (40 C).
To validate the conversion of internal temperature logs from California, Reuters assessed the AI extraction process by calculating the exact match rate — the percentage of entries where the AI output exactly matched the original handwritten PDF — and the degree of difference between the values.
Reuters randomly sampled 384 temperature logs from California. Sample size was determined using Fisher's formula:
where:
E = margin of error or confidence interval to express degree of uncertainty
We chose a 95% confidence level (1.96 for 95% confidence), a proportion of 0.5 (50%) to be conservative and offer a maximum sample size and a margin of error of 5% (0.05 for ±5%). This yielded a sample size of 384. We then manually verified this sample of 384 AI outputs against the original PDFs from CDCR.
The model correctly read and transcribed the handwritten temperature data 95% of the time, with a ±5% margin of error at 95% confidence. However, most errors were small. The success rate rose to 97% for temperatures being within ±1 F (0.6 C).
In five cases, Reuters couldn't read the original handwriting during the manual checks, or AI left the output blank. In cases where there was both a legible recorded temperature and an AI-extracted output, the mean absolute error (MAE), or how far off the AI output was from the actual record, was 2.6 degrees F (1.4 degrees C). The root mean squared error (RMSE), which gives more weight to larger misses, was 3.5 degrees F (1.9 degrees C). This means that on average, the AI-extracted outputs were within about 3.5 degrees of the actual recorded temperatures.
These results indicate the model performed reliably, and often produced a result close to the original.
Mismatches mostly occurred when the PDF had:
All data and methodology are publicly available on GitHub at https://github.com/reuters-graphics/prison-heat-records