Page 285 - GHES-3-3

Global Health Economics and
Sustainability
Carbon footprint of smartphones in healthcare

make decisions that align with responsible management of LLMs are trained on public data, reinforcement learning,
the environment. As healthcare systems around the world and the internet, they should have answers to our queries.
strive for sustainability, it is essential to comprehend all We then compare the actual data from the CSR reports to
the elements that affect the environment, including the the LLM answers.
digital tools that physicians and patients use on a daily The LLM query is: “What is the average carbon footprint
basis. This study is intended for physicians and healthcare of the entire lifecycle of (smartphone model)?” These
professionals, aiming to contribute to the growing body of queries were conducted for flagship devices, including the
knowledge about smartphones and their carbon footprint.
Apple iPhone 14, 15, and 16 series, as well as the Samsung
2. Methods Galaxy S24 series, Z Flip6, and Z Fold6.

We chose Apple and Samsung smartphones since they In addition, we asked the same queries to smart home
are the most popular smartphones in terms of sales and devices to determine if they produced any emissions as
shipment market share in the United States (Counterpoint individual household assistants. We conducted the same test
nd
Research, 2025). Carbon emissions data for flagship described above using the 2 generation of Apple HomePod
smartphones from Apple and Samsung were obtained and the latest versions of Google Home and Echo.
from publicly available sources, including CSR reports, The LLM responses were compared directly to the
official corporate websites, and product environmental emissions data reported by Apple and Samsung. If the LLM
reports (Apple Inc., 2024; Samsung Electronics, 2024). model failed to provide an answer or explained that the
The designation of “actual emissions data” was based data were publicly available, it was noted as a missing data
on the corporate materials of each company. For Apple point. To quantify the accuracy of each model, the percent
devices, we obtained the carbon footprint values directly error was calculated using Equation II.
from Apple’s official product environmental reports (Apple
Inc., 2025), which provided detailed lifecycle emissions Percenterror= AIvalue-actualvalue ×100 (II)
in kg CO emitted for each model that we used to test Actualvalue
2
our LLMs. Similarly, for Samsung devices, we referred
to the official Galaxy Environmental Reports (Samsung Error rates were analyzed for each LLM across all
Electronics, 2025), which documented carbon emissions queried devices, with a particular focus on patterns of
across the whole product lifecycle. overestimation or underestimation. This approach allowed
us to assess the reliability and consistency of LLMs in
To evaluate the accuracy of AI-generated data from
the chatbots, we posed the same question to four LLMs: estimating smartphone carbon footprints.
ChatGPT-4.0, Gemini 1.5 Flash, Claude.ai, and Meta AI. 3. Results
ChatGPT-4.0 was chosen because it has the most daily
queries (570,000), while Gemini was chosen because it is The analysis revealed variabilities in the accuracy of LLMs’
linked to Google with 8.5 billion daily searches. Claude. carbon emissions data compared to manufacturer-reported
ai was chosen because it uses “constitutional AI,” meaning values. Responses from ChatGPT-4.0, Gemini, and
that once humans set the algorithm, the AI trains itself. Claude.ai frequently align closely with actual emissions for
Meta AI uses LLM Meta AI, which is a different AI Apple models (Table 1), such as the the iPhone 14 (61 kg
algorithm from ChatGPT or Gemini. These models were CO ) reported by the three AI-LLMs vs. 61 kg CO₂ actual.
2
built using distinct development pipelines, including Answers given by ChatGPT and Gemini for the Samsung
different architectures, data, and training methodologies, models were mostly accurate, with only minor errors.
which resulted in each LLM being unique on its own. We However, Claude.ai’s response was less reliable and showed
omitted Co-pilot since OpenAI designed both ChatGPT higher percent errors.
and Co-Pilot, and they use the same AI architecture. For queries to Meta AI, the LLM often failed to provide
We examined Apple HomePod, Google Home, and definitive answers for iPhone and Samsung devices
Echo, but only Apple provided CSR literature on its website (Tables 1-3), citing that the information was “not publicly
or 10-K statements. The latest versions of Google Home disclosed” or recommending referral to official product
and Echo did not have any current CSR data available to environmental reports. For Table 1, there is very little
the public. We used the same query for the smart home information available on the company’s 10-K statements
devices. or CSR report for any more information.
We focused on smartphones to highlight the companies’ Overall, the average percent errors varied significantly
transparency in revealing their products’ emissions. Since among models. ChatGPT-4.0 demonstrated lower average

Volume 3 Issue 3 (2025) 277 https://doi.org/10.36922/ghes.8359